Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning

Huajie Tan1,2,∗ , Yuheng Ji2,3,4,∗ , Xiaoshuai Hao2,∗,† , Minglan Lin2 ,

Pengwei Wang2,† , Zhongyuan Wang2 , Shanghang Zhang1,2,B
1
State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University
2
Beijing Academy of Artificial Intelligence 3 Institute of Automation, Chinese Academy of Sciences
4
School of Artificial Intelligence, University of Chinese Academy of Sciences
arXiv:2503.20752v2 [[Link]] 27 Mar 2025

+ ANS-SFT + COT-SFT + Reason-RFT-Zero + Reason-RFT

80
Only SFT-based Training + 6.3
ACC/%

+ 8.6
In-Domain
60
VLMs Gap
Large-Scale
Training Data Out-of-Domain
+ 4.2
40

Transient Adaptation Gap

Question: For the pair of similar

In-Domain

20 Out-of-Domain

figures, find the area of the green Visual-Counting Structure-Perception Spatial-Transformation Steps

figure in <image>. Choices:

(A) 12.5
(B) 20.0 Greedy Reward Transient Reasoning
8cm OpenAI GPT-4o Qwen2VL-2B（Zero-Shot）

(C) 51.2 5cm Qwen2VL-2B + ANS-SFT

Qwen2VL-2B + COT-SFT
Qwen2VL-2B + Reason-RFT-Zero
Qwen2VL-2B + Reason-RFT
Stratification Adaptation Gap Redundancy
(D) 81.9 S=? Answer from Reason-RFT Model
S = 32
<think>Let‘s start by understanding the COT Reasoning Data
Answer from COT-SFT Model problem. We have … and blue one is 8 cm.
Since the rectangles are similar, the ratio
<think>Let's start by understanding the of their corresponding sides is equal.
Question + Images
relationship between the dimensions of the That means the scale factor from blue to
blue and green figures. Since the figures …, green is 5/8. Now, when it comes to
for similar figure problems. areas, the ratio isn't just the same as the
Now, let's calculate the area of the green side ratio; it's the square of the scale
VLMs
figure using the assumed dimensions. factor. So, the area ratio would be (5/8)²,
which is 25/64. The blue area is 32 cm²,
Question + Images
The area of the green figure is calculated as
(8 cm)²* (5 cm) = 32 cm²* 5 cm = 160 cm². so I need to apply this ratio to find the
However, the original answer indicates that green area. This calculation is A1 A2 A3 … An
the side length is actually 5 cm, not 8 cm. straightforward: 32 ÷ 64 = 0.5, and 0.5 × VLMs …
25 = 12.5. Let me double-check. If the
Re-calculating with the actual side length:
(5 cm)²* (5 cm) = 25 cm²* 5 cm = 125 cm².
blue area is 32 and the scale factor is 5/8, …
then the area should be (5/8)² times 32.
Thus, the correct area is 125 cm², which Yep, 25/64 * 32 = 12.5. So, the green area COT Reasoning Process
corresponds to choice (D). </think> must be 12.5 cm², which corresponds to + GRPO
<answer> D </answer> choice (A). </think> Final Answer
<answer> A </answer>
STAGE 1: SFT-based Activation STAGE 2: RL-based Enhancement
Figure 1. Overview of Reason-RFT. Compared to traditional SFT-based methods, our proposed Reason-RFT framework demonstrates
superior generalization in visual reasoning tasks, excelling in reasoning improvement, out-of-domain performance, and data efficiency.

Abstract and limiting its real-world applicability. To address these

limitations, we propose Reason-RFT, a novel reinforce-
Visual reasoning abilities play a crucial role in understand- ment fine-tuning framework that significantly enhances gen-
ing complex multimodal data, advancing both domain- eralization capabilities in visual reasoning tasks. Reason-
specific applications and artificial general intelligence RFT introduces a two-phase training framework for vi-
(AGI). Existing methods improve VLM reasoning via Chain- sual reasoning: (1) Supervised Fine-Tuning (SFT) with cu-
of-Thought (CoT) supervised fine-tuning, using meticu- rated Chain-of-Thought (CoT) data activates the reason-
lously annotated training data to enhance visual reason- ing potential of Vision-Language Models (VLMs), followed
ing capabilities. However, this training paradigm may lead by (2) Group Relative Policy Optimization (GRPO)-based
to overfitting and cognitive rigidity, restricting the model’s reinforcement learning that generates multiple reasoning-
ability to transfer visual reasoning skills across domains response pairs, significantly enhancing generalization in
visual reasoning tasks. To evaluate Reason-RFT’s vi-
∗Equal contribution. sual reasoning capabilities, we reconstructed a compre-
†Project leaders.
B Corresponding author.
hensive dataset spanning visual counting, structure percep-

1
tion, and spatial transformation, serving as a benchmark heterogeneity of multimodal data. Motivated by these ob-
to systematically assess visual cognition, geometric under- servations, this paper investigates advanced RL techniques
standing, and spatial generalization. Experimental results to enhance the adaptability and transferability of visual rea-
demonstrate Reasoning-RFT’s three key advantages: (1) soning, aiming to overcome generalization challenges and
Performance Enhancement: achieving state-of-the-art re- improve performance in real-world applications.
sults across multiple tasks, outperforming most mainstream
In this paper, we propose Reason-RFT, a novel two-
open-source and proprietary models; (2) Generalization
phase reinforcement fine-tuning framework designed to
Superiority: consistently maintaining robust performance
enhance generalization in visual reasoning tasks. First,
across diverse tasks and domains, outperforming alterna-
we employ Supervised Fine-Tuning (SFT) with Chain-of-
tive training paradigms; (3) Data Efficiency: excelling in
Thought (CoT) reasoning to activate the model’s potential
few-shot learning scenarios while surpassing full-dataset
reasoning capabilities, using a high-quality domain-specific
SFT baselines; Reason-RFT introduces a novel paradigm
visual reasoning dataset tailored to stimulate related rea-
in visual reasoning, significantly advancing multimodal re-
soning abilities. Subsequently, we further enhance reason-
search. Project website: ReasonRFT.
ing potential through Group Relative Policy Optimization
(GRPO), demonstrating that Reason-RFT achieves supe-
rior generalization by pushing the model’s reasoning lim-
1. Introduction its. To evaluate its capabilities, we reconstructed a high-
Visual reasoning is pivotal for understanding complex mul- quality dataset spanning visual counting, structure percep-
timodal data and advancing artificial general intelligence tion, and spatial transformation, serving as a benchmark for
(AGI) [38, 45], making it a central focus in intelligent assessing visual cognition, geometric understanding, and
systems research. Recent advancements in image recog- cross-task generalization. Extensive experiments highlight
nition [27, 48] and scene understanding [9, 64] have en- three key advantages of Reason-RFT: (1) Performance Im-
abled transformative applications in healthcare [47, 66], provement: It significantly outperforms mainstream VLMs
robotics [28, 36], and autonomous driving [19–22]. Con- in tasks like visual counting, structure perception, and spa-
sequently, enhancing visual reasoning capabilities has gar- tial transformation; (2) Enhanced Generalization: It con-
nered significant attention from both industry and academia sistently surpasses SFT-only and RL-only paradigms across
for its potential to drive transformative advancements. diverse tasks, validated through multi-dimensional evalua-
Researchers have explored two primary categories of tions; (3) Data Efficiency: It achieves over 95% of the per-
methods to enhance visual reasoning capabilities: (1) formance of SFT-only approaches using less than 20% of
neural-symbolic methods [4, 7, 13, 16, 69], which integrate the data. These results underscore the effectiveness and ef-
symbolic reasoning with neural networks to improve inter- ficiency of Reason-RFT, offering a robust solution for ad-
pretability and modularity, and (2) Supervised Fine-Tuning vancing visual reasoning tasks.
(SFT) based on vision-language models (VLMs) [57, 63], Our main contributions are summarized as follows.
which utilize end-to-end training to strengthen reasoning
abilities. However, both approaches face significant lim- • We propose Reason-RFT, a novel two-phase reinforce-
itations. Neural-symbolic methods are hindered by high ment fine-tuning framework that significantly enhances
complexity and a strong reliance on program generation, the visual reasoning capabilities of VLMs by effectively
while SFT is constrained by its dependence on high-quality combining the complementary strengths of SFT-based
Chain-of-Thought (CoT) annotated data and meticulously and RL-based methods.
designed data mixing strategies, leading to issues such as • We systematically analyze SFT-based and RL-based
overfitting, cognitive rigidity, and limited transferability methods in visual reasoning tasks, and highlight the limi-
from in-domain to out-of-domain tasks. These challenges tations of SFT-based approaches and the superior general-
reduce their effectiveness in real-world applications. ization of RL-based methods in reasoning improvement,
Recent advances in models such as GPT-o1 [45], out-of-domain performance, and data efficiency.
DeepSeek-R1 [15], and Kimi-1.5 [56] have demonstrated • We reconstructed a comprehensive dataset that covers
that large-scale reinforcement learning (RL) during post- three core domains: visual counting, structure perception,
training significantly enhances reasoning abilities in math- and spatial transformation, which serves as a benchmark
ematics and coding. By providing a dynamic and adaptive to systematically assess visual cognition, geometric un-
alternative to static datasets, RL effectively addresses tradi- derstanding, and cross-task generalization.
tional limitations and improves reasoning, even with min- • Extensive experiments validate the effectiveness of
imal data. However, purely RL-based methods face chal- Reason-RFT, offering valuable insights to advance visual
lenges in generalizing across visual reasoning tasks, often reasoning and introducing a novel paradigm that signifi-
limited by the model’s inherent cognitive constraints and the cantly pushes forward multi-modal research.

2
2. Related Work visual input. This process can be represented as a mapping:
Visual Reasoning Visual reasoning is a key challenge in R : (I, T ) → A,
advancing artificial general intelligence (AGI), requiring
models to perform complex cognitive tasks based on vi- where I ∈ RH×W ×C is the visual input with H, W , and
sual perception [18, 38, 42]. It has broad applications, C representing height, width, and channels, T is the tex-
including visual counting [37, 38], geometric problem- tual description or question typically in natural language,
solving [12, 30, 42, 51, 70], visual transformation reason- and A is the derived conclusion or answer, often in natural
ing [23], scientific research [31, 41], and robotic task plan- language or structured data. Through this mapping, visual
ning [17, 24, 28]. Traditional methods rely on program gen- reasoning models integrate visual and textual information
eration [16, 29, 53] or neural-symbolic approaches [4, 7, to achieve effective reasoning.
13, 69], while recent vision-language models (VLMs) lever-
age large language models (LLMs) for enhanced reason- 3.2. Reason-RFT
ing. For example, LLaVA-CoT [63] uses multi-stage super- In this section, we propose Reason-RFT, a novel two-phase
vised fine-tuning (SFT) with chain-of-thought (CoT) [61], hybrid training strategy to enhance the reasoning capabili-
and Insight-V [10] combines SFT with reinforcement learn- ties of VLMs in complex visual reasoning tasks. As shown
ing (RL). DeepSeek-R1-Zero [14] introduces a rule-based in Fig. 2, the framework comprises two stages: (1) SFT-
RL method, significantly improving reasoning. Building on based Visual Reasoning Activation, which uses SFT with
DeepSeek-R1 [14], our work compares SFT-based and RL- high-quality CoT reasoning data to activate the model’s
based paradigms, highlighting the superior performance of domain-specific reasoning capabilities, and (2) RL-based
R1-based methods in visual reasoning. Visual Reasoning Reinforcement, which employs the GRPO
Post-Training Post-Training is a crucial phase for en- algorithm with rule-based rewards to further push the upper
hancing the performance of LLMs and VLMs, bridging pre- limits of the model’s reasoning potential.
trained models and their real-world applications [8, 32, 54,
68]. It primarily involves two methodologies: SFT [58, 62] 3.2.1. STAGE 1: SFT-based Reasoning Activation
and RL [43, 46, 52, 65, 71]. SFT adapts pre-trained mod- In the initial phase, we employ Supervised Fine-Tuning
els to specific tasks using task-oriented datasets, often for- (SFT) on a structured visual reasoning dataset containing
matted as instructions. Research like FLAN [60] high- step-by-step reasoning processes. This phase trains the
lights the importance of diverse instruction-tuning datasets model to decompose complex tasks into logical steps. Each
for improving zero-shot performance, while iterative pro- sample is represented as (x, q, r, a), where x denotes the in-
cesses, such as Llama 3.1’s six-round strategy [11], inte- put images, q is the question, r is the reasoning steps, and
grate rejection sampling, synthetic data, and human an- a is the final answer. The training objective maximizes the
notations. RL aligns models with human preferences or likelihood of generating both r and a given (x, q):
task-specific goals through feedback mechanisms. Rein-
T
forcement Learning from Human Feedback (RLHF) [46] X
refines models using human preference data, as seen in LSFT = −E(x,q,r,a)∼D log πθ (yt | x, q, y<t ), (1)
t=1
Llama 3.1 [11] and Nemotron-4 [2], which use reward mod-
eling techniques like DPO [49] and RPO [2]. For exam- where D is the dataset, y the concatenated sequence of r and
ple, TÜLU 3 [34] employs length-normalized DPO, while a, and πθ the model’s token distribution. The output model
DeepSeek-V3 [39] combines rule-based and model-based πCoT serves as the initialization for the next stage, ensuring
reward systems. Recently, DeepSeek-R1 [14] achieved sig- a robust foundation for reinforcement learning.
nificant text reasoning improvements through pure RL [50].
Our work adapts R1 methodologies to VLMs, enhancing vi- 3.2.2. STAGE 2: RL-based Reasoning Enhancement
sual reasoning, and systematically compares SFT-based and In the second phase, we refine πCoT using Group Rela-
R1-based paradigms, demonstrating the superiority of R1- tive Policy Optimization (GRPO), leveraging reinforcement
based methods in visual reasoning tasks. learning for its efficiency and scalability. Unlike Proximal
Policy Optimization (PPO), which requires a computation-
3. Methodology ally expensive value network, GRPO calculates relative ad-
vantages by comparing rewards within a group of sampled
3.1. Problem Definition actions, reducing computational overhead and simplifying
Visual reasoning can be formally defined as follows: given optimization. This makes GRPO particularly suitable for
a visual input I (e.g., images or videos) and a correspond- visual reasoning tasks.
ing textual description or question T , the goal is to derive a Sampling Action Groups For each input state s =
conclusion or answer A by analyzing the information in the (x, q), where x is the visual encoding of the input image

3
 S1: SFT-based Activation S2: RL-based Enhancement

Problem Reward Evaluation

Visual Reasoning COT Data KL Divergence Reference
Model
SFT for Limited Steps Policy Format Reward
Pretrained VLM
Text- Prompt Model Sampled Candidate Responses

GRPO
Accuracy Reward
Initial Policy Model Images
Policy Gradient Update

Format Reward Mathematical Accuracy Reward Function-based Accuracy Reward

Format: <think> {Reasoning Process} </think> 0 0

<answer> {Final Answer Here} </answer> 1 2

Area ? Angle ? 2
S1
1, 𝑖𝑓 𝐴𝑛𝑠𝑤𝑒𝑟 adheres to 𝑓𝑜𝑟𝑚𝑎𝑡 S1
3
1
3

𝑹𝐅𝐨𝐫𝐦𝐚𝐭 = ቊ
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 4
5
4
5
6 6

Discrete-valued Accuracy Reward

Prompt: You are a visual geometry expert. Prompt: You are a spatial visual reasoning
Prompt: You are a visual counting Given the <image>, you should solve the expert. Given the center-view <image> of initial
expert. Given the <image> in front problem as shown below: {problem}, and state, the center-view <image> (or the left-view
of you, if I add two large buses, and provide the answer in LaTeX format or as a <image> or the right-view <image>) of final
then remove all the bikes, now how floating-point number? state, and the {Function List}, you should
many buses are there in the table? GT: <answer> 3.14 or π (as example) <answer> determine a sequence of transformations that
GT: <answer> 4 (as example) <answer> can achieve the final state.
1, 𝐴𝑛𝑠𝑤𝑒𝑟 − 𝐺𝑇 < 𝜖1 ∗ |𝐺𝑇| GT: <answer> {sequence of function-type steps} <answer>
1, 𝐴𝑛𝑠𝑤𝑒𝑟 = 𝐺𝑇 𝑹𝑨𝒄𝒄 = ቐ𝐶𝑂𝑆𝐼𝑁𝐸 𝑇𝑦𝑝𝑒, 𝑏𝑒𝑡𝑤𝑒𝑒𝑛 𝜖1 ∗ 𝐺𝑇 𝑎𝑛𝑑 𝜖2 ∗ |𝐺𝑇|
𝑹𝑨𝒄𝒄 = ቊ 0, 𝐴𝑛𝑠𝑤𝑒𝑟 − 𝐺𝑇 > 𝜖2 ∗ |𝐺𝑇| 𝑹𝑨𝒄𝒄 = 𝐅𝐮𝐥𝐥 matching reward + 𝐏𝐚𝐫𝐭𝐢𝐚𝐥 matching reward
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒

Figure 2. Framework of Reason-RFT. Reason-RFT introduces a two-phase training framework for visual reasoning. First, Supervised
Fine-Tuning (SFT) with Chain-of-Thought (CoT) reasoning activates the model’s reasoning potential using a high-quality domain-specific
visual reasoning dataset in stage 1. Subsequently, in stage 2, Group Relative Policy Optimization (GRPO) enhances reasoning capabilities
further, enabling Reason-RFT to achieve superior generalization by pushing the model’s reasoning limits. Specifically, reward evaluation
consists of format reward and three different types of accuracy reward.

and q the textual encoding of the question, GRPO samples Based on these advantages, the policy is updated to rein-
a group of actions {a1 , a2 , . . . , aG } from the current policy force actions with positive advantages and reduce the prob-
πθ , initialized from πCoT . The sampling process is: ability of less effective ones. Policy updates are further con-
strained by minimizing the KL divergence between the up-
ai ∼ πθ (a | x, q), for i = 1, 2, . . . , G. (2) dated and reference models, ensuring stable RL learning.
This strategy ensures diverse responses, promoting explo-
3.3. Reward Design for Visual Reasoning Tasks
ration and preventing premature convergence.
Reward Evaluation Each sampled action ai is assigned For the diverse requirements of visual reasoning tasks, in-
a reward R(ai ) based on verifiable criteria, resulting in a cluding Visual Counting, Structure Perception, and Spatial
reward set {r1 , r2 , . . . , rG }. For visual reasoning tasks, Transformation, our reward design integrates two essential
the reward function R(ai ) combines two components: For- components: Format Reward and Accuracy Reward. The
mat Reward Rformat (ai ) and Accuracy Reward Racc (ai ). Format Reward is uniformly applied across all tasks, en-
The Format Reward ensures structured response formats, suring that the model strictly adheres to a structured re-
while the Accuracy Reward evaluates correctness, balanc- sponse format, which significantly enhances interpretability
ing structured reasoning and factual accuracy. The reward and consistency. For the Accuracy Reward, we carefully tai-
function is defined as: lor the design to the specific characteristics of each task, as
shown in Fig. 2, creating task-specific reward mechanisms
R(ai ) = Rformat (ai ) + Racc (ai ). (3) to evaluate the correctness of the model’s responses.
Policy Update with Relative Advantage Rewards are Format Reward This component ensures structured and
normalized within the sampled group to compute relative interpretable responses by requiring the model to adhere
advantages {A1 , A2 , . . . , AG }, defined as: to a predefined template: reasoning within <think> and
</think> and the final answer within <answer> and
ri − mean{r1 , r2 , . . . , rG } </answer>. A reward of 1 is given for strict adherence,
Ai = . (4) while deviations result in a reward of 0.
std{r1 , r2 , . . . , rG }

4
 Accuracy Reward This component evaluates the cor- 4. Experiment
rectness of the model’s responses, ensuring alignment with
ground truth across diverse visual reasoning tasks. To ad- 4.1. Experimental Details
dress task diversity, we design tailored reward mechanisms Datasets In this paper, we comprehensively evaluate the vi-
for discrete-valued, mathematical, and function-based prob- sual reasoning capabilities of our method by leveraging six
lems. Each mechanism is crafted to handle the unique char- existing datasets, enhanced through subtask categorization,
acteristics of its problem category, enabling precise and fair error-prone data filtering, and dataset restructuring. De-
evaluation. Below, we introduce the three reward forms. tailed protocols for data filtering and restructuring are pro-
• Discrete-valued Type This reward type applies to Visual vided in the supplementary materials. Specifically, we de-
Counting and partial Structure Perception tasks, where fine three task categories as follows.
answers are discrete values (e.g., multiple-choice or
• Visual Counting is a multimodal reasoning task evaluat-
integer-based responses). The accuracy reward Racc (ai )
is defined as: ing the integration of linguistic, visual, and mathematical
( skills by solving arithmetic problems in 3D block-based
1, if apred = agt scenes, which includes four subtask types, with datasets
Racc (ai ) = (5)
0, otherwise, from CLEVR-Math [38] and Super-CLEVR [37]. Specif-
where apred is the model’s predicted answer and agt is ically, CLEVR-Math provides 35k training and 1k test
the ground truth. This binary reward penalizes deviations samples, while Super-CLEVR serves as an out-of-domain
from the exact ground truth, ensuring precision in tasks test set with 1k samples to assess generalization.
requiring unambiguous answers. • Structure Perception is a visual reasoning task requir-
• Mathematical Type This reward type is designed for ing models to perceive and interpret structural infor-
Structure Perception tasks involving numerical answers, mation in mathematical geometries, medical imaging,
such as floating-point values or LaTeX-formatted expres- chart layouts, and architectural designs. It includes both
sions. It uses a tolerance-based evaluation to account choice and non-choice formats, with datasets sourced
for minor numerical deviations. The accuracy reward from Geo170K [12], Math360K [51] (collectively Geo-
Racc (ai ) is defined as: Math), and Geometry3K [40]. GeoMath provides 4.5k
1
 
|apred − agt | − ϵ1 × |agt |
  training and 820 test samples, while Geometry3K serves
Racc (ai ) = cos π × +1 , as an out-of-domain test set with 800 samples to evaluate
2 (ϵ2 − ϵ1 ) × |agt |
(6) cross-domain adaptability.
where apred is the predicted answer, agt is the ground • Spatial Transformation is a spatial-visual reasoning task
truth, ϵ1 is the tolerance threshold for an exact match (e.g., requiring models to infer single-step or multi-step trans-
0.05), and ϵ2 is the upper bound for partial rewards (e.g., formation actions by analyzing initial and final visual
0.20). If |apred − agt | < ϵ1 × |agt |, the reward is 1 (ex- states for 3D scenes from multiple perspectives (e.g., cen-
act match); if |apred − agt | > ϵ2 × |agt |, the reward is 0 ter, left, right). The dataset TRANCE [23] features four
(incorrect). This formulation ensures smooth transitions distinct difficulty levels, with 60k training and 6k test
between full and partial rewards, enabling fair evaluation samples. Out-of-domain test samples are generated by
of numerical accuracy. rendering identical data from left and right viewpoints to
• Function-based Type This reward type is designed for rigorously assess robustness to perspective variations.
Spatial Transformation tasks requiring a sequence of Evaluation Metrics We use accuracy-rate (Acc) as the
transformation functions. The accuracy reward Racc (ai ) primary metric [67]. For numerical answers, correctness is
evaluates the alignment between the predicted sequence
verified by mathematical equivalence to the ground truth.
Tpred and the ground truth Tgt , computed as:
For multiple-choice questions, we perform a string match.
f +o+v f +o/v f
len(Tpred ) + α · len(Tpred ) + β · len(Tpred ) For function-type sequences, we use stepwise multi-level
Racc (ai ) = , evaluation to assess alignment with the correct solution.
max(len(Tpred ), len(Tgt ))
(7) Implementation Details We utilize Qwen2-VL-2B and
f +o+v
where Tpred is the subset of transformation steps Qwen2-VL-7B [59] as the backbone models for our exper-
with complete matches (w/ function, object, and value), iments. Our implementation is built on the open-source
f +o/v
Tpred are the subsets with partial and only-function frameworks Open-R1 [25] and vLLM [33], ensuring repro-
matches (w/ function and object, or w/ function and ducibility and scalability. All experiments were conducted
f
value), Tpred is the subset with only-function matches. α on a cluster of servers, each equipped with 8×A800 GPUs.
and β are the weighting coefficients for partial matches. For further details, see the supplementary materials.
This formulation ensures nuanced evaluation, rewarding Training Paradigms and Baselines To assess the per-
both exact and partially correct responses while allowing formance and generalization of different training strategies,
flexible adjustment of partial match contributions. we compare: (1) SFT-based methods—ANS-SFT, which

5
Table 1. Results on three visual reasoning tasks. The best results among different training paradigms are highlighted in bold, while the
second-best results are underlined. “ID” denotes in-domain test data, and “OOD” denotes out-of-domain test data.

Visual Counting Structure Perception Spatial Transformation

Method
Clevr-Math Super-Clevr GeoMath Geometry3k TRANCE TRANCE-L TRANCE-R
AVG AVG AVG
ID OOD ID OOD ID OOD
Proprietary Models
GPT-4o-2024-08-06 [26] 68.10 34.31 51.20 50.18 43.49 46.83 42.55 28.67 29.76 35.88
Gemini-1.5-Pro [55] 61.80 37.50 49.65 50.12 48.38 49.45 26.22 18.76 19.88 22.77
Open-Source Models
Qwen2.5-VL-3B-Instruct [5] 75.90 39.30 57.60 36.75 37.44 37.09 8.57 8.26 8.31 8.42
Phi-3.5-Vision-4B-Instruct [1] 21.40 15.20 18.30 36.83 50.25 43.54 7.42 2.45 4.02 5.33
Llava-OneVision-7B [35] 69.70 29.10 49.40 77.63 43.66 60.64 10.00 8.33 8.74 9.27
Qwen2.5-VL-7B-Instruct [5] 74.60 35.20 54.90 44.00 45.61 44.80 19.63 13.12 13.42 16.45
InternVL-2.5-8B [6] 93.50 35.30 64.40 63.00 47.32 51.60 7.19 6.62 6.63 6.91
Llama-3.2-11B-Vision [44] 10.30 9.50 9.90 13.75 20.85 17.30 8.22 8.40 9.03 8.47
Pixtral-12B [3] 42.60 22.90 32.75 30.38 36.09 33.23 7.35 5.03 5.22 6.42
Qwen2VL-2B-Instruct
Zero-Shot 82.40 32.00 57.20 25.86 20.63 23.25 3.78 4.60 4.67 4.35
+ ANS-SFT 96.20 39.20 67.70 51.34 22.50 36.92 77.39 49.24 50.33 58.99
+ CoT-SFT 85.50 46.50 66.00 43.05 25.25 34.15 64.37 43.19 42.86 50.14
+ Reason-RFT-Zero 98.40 44.80 71.60 47.68 32.50 40.09 42.13 34.07 33.41 33.74
+ Reason-RFT 96.80 51.20 74.00 49.03 33.13 41.08 74.61 64.05 64.08 67.58
Qwen2VL-7B-Instruct
Zero-Shot 98.60 42.10 70.35 43.30 43.88 43.59 13.53 12.72 12.78 13.01
+ ANS-SFT 95.00 33.90 64.45 51.34 25.38 38.36 82.19 54.29 54.83 63.77
+ CoT-SFT 87.30 42.40 64.85 50.49 33.00 41.75 81.31 47.90 47.80 59.00
+ Reason-RFT-Zero 99.40 53.00 76.20 55.00 54.75 54.88 67.67 57.20 56.15 56.68
+ Reason-RFT 95.60 51.00 73.30 59.27 49.25 54.26 79.97 59.36 58.61 65.98

Table 2. More detailed results on Super-Clevr. “Encountered” impact of model scale. In addition, we also select the most
refers to the types of questions the model has previously seen dur- advanced open-source models [1, 3, 5, 6, 35, 44] and the
ing Clevr-Math training, while “UnEncountered” denotes the com- proprietary models [26, 55] available as baselines to evalu-
plicated types that the model has not encountered (i.e. questions
ate the performance of different training paradigms.
with mixture of addition and subtraction).

Methods Encountered UnEncountered Total

4.2. Results on In-Domain Tasks
Qwen2VL-2B-Instruct To evaluate the In-Domain (ID) performance of Reason-
Zero-Shot 42.67 0.00 32.00 RFT against different training methods and baseline mod-
+ ANS-SFT 51.07 5.20 39.20 els across visual reasoning tasks, we conducted extensive
+ CoT-SFT 49.73 36.80 46.50 training and validation on 2B/7B models for three tasks.
+ Reason-RFT-Zero 58.00 5.20 44.80
The results in Tab. 1 show: (1) Visual Counting RL-based
+ Reason-RFT 60.00 28.40 51.20
methods outperform all open-source and proprietary base-
Qwen2VL-7B-Instruct
line models as well as SFT-based methods in both 2B and
Zero-Shot 54.53 4.80 42.10 7B models, with Reason-RFT-Zero in 7B nodel achieving
+ ANS-SFT 42.53 8.00 33.90
the best performance; (2) Structure Perception RL-based
+ CoT-SFT 45.33 33.60 42.40
+ Reason-RFT-Zero 63.60 21.20 53.00 methods surpass SFT-based approaches in 7B model, while
+ Reason-RFT 56.13 35.60 51.00 ANS-SFT performs best in 2B model. CoT-SFT shows
limited improvement, as enforced reasoning supervision in-
hibits cognitive enhancement. Additionally, Reason-RFT in
fine-tunes on answer generation, and CoT-SFT, which uses 7B model outperforms all proprietary and most open-source
supervised learning with chain-of-thought (CoT) reasoning; models except InternVL-2.5-8B [6] and Llava-OneVision-
and (2) RL-based methods—Reason-RFT-Zero, which ap- 7B [35]; (3) Spatial Transformation SFT-based methods
plies reinforcement learning without reasoning activation achieve the highest performance, with Reason-RFT achiev-
phase, and Reason-RFT, which uses partial CoT data for ing comparable performance while totally surpassing all
reasoning activation before RL training. For comprehensive baseline models. However, Reason-RFT-Zero performs
experiments, we use Qwen2-VL-Instruct [59] as the base poorly, likely due to its difficulty adapting to function-
model, evaluating both 2B and 7B variants to explore the based tasks, whereas Reason-RFT benefits from reason-

6
 ANS-SFT COT-SFT Reason-RFT-ZERO Reason-RFT
ing activation, reaching a higher performance ceiling. In

ACC/%

ACC/%
Activation Activation
summary, Reason-RFT achieves performance comparable Zero-Shot Zero-Shot
to SFT-based methods in ID tasks while outperforming all
baselines in Visual Counting and Spatial Transformation,
and most of baselines in Structure Perception.

4.3. Results on Out-of-Domain Generalization

Sample/# Sample/#
(a) (b)
To validate the Out-of-Domain (OOD) performance of

ACC/%

ACC/%
Activation Activation
Reason-RFT across different training methods and base- Zero-Shot Zero-Shot

line models in visual reasoning tasks, we conducted com-

prehensive experiments on 2B/7B models for three tasks.
Results in Tab. 1 and Tab. 2 show that: (1) Visual Count-
ing RL-based methods demonstrate superior generaliza-
tion over SFT-based methods in both 2B and 7B models,
with Reason-RFT surpassing ANS-SFT by 12% (2B) and (c) Sample/# (d) Sample/#

17% (7B) while outperforming all open-source and propri-

etary baselines. Notably, compared with purely RL-based Figure 3. Results of different methods on the Spatial Transfor-
method (i.e., Reason-RFT-Zero), CoT-SFT and Reason- mation task across training processes. (a) Evaluation results for
RFT can activate model capabilities on unseen complicated 2B model on ID task, (b) Evaluation results for 7B model on ID
questions, achieving strong results despite its poor zero-shot task, (c) Evaluation results for 2B model on OOD task, (d) Evalu-
ation results for 7B model on OOD task.
performance before in-domain training, as shown in Tab. 2.
This also highlights the limitations of Reason-RFT-Zero
due to cognitive constraints, especially in smaller mod- 4.5. Training Insights
els (i.e. 2B models); (2) Structure Perception RL-based
methods consistently outperform SFT-based methods, with Transient Adaptation Gap This phenomenon refers to the
Reason-RFT achieving the best result in 2B model (8% temporary performance drop observed during the initial
higher than CoT-SFT) while Reason-RFT-Zero performs phase of training process with Reason-RFT-Zero. As the
strongest in 7B model and Reason-RFT remains competi- model transitions from directly outputting answers to gen-
tive (16% higher than CoT-SFT), with SFT-based methods erating structured reasoning processes, it experiences a brief
showing limited impact especially in the 7B model; (3) Spa- adaptation difficulty, leading to a sharp performance de-
tial Transformation RL-based methods surpass SFT-based cline followed by gradual recovery. As shown in Fig. 4 (a),
approaches in both 2B and 7B models while significantly which illustrates the training process of Reason-RFT-Zero
outperforming all baseline models, with Reason-RFT (2B) on the CLEVR-MATH, both ID and OOD test performances
demonstrating remarkable OOD generalization exceeding exhibit this sharp drop and recovery within the first 100
GPT-4o [26] by 35%. In summary, Reason-RFT surpasses steps. We further investigate this phenomenon through a
all open-source and proprietary baselines as well as alter- case study, shown in Fig. 4 (b), which demonstrates that
native training methods, demonstrating remarkable perfor- forcing the model to output its reasoning process can lead
mance in visual reasoning generalization capabilities. to incorrect answers to originally correct question.
Greedy Reward Stratification This phenomenon de-
4.4. Assessment on Training Efficiency scribes the model’s tendency during Reason-RFT-Zero
training to prioritize easier rewards (e.g., Format Reward)
To demonstrate the data efficiency of Reason-RFT during over harder ones (e.g., Accuracy Reward). As shown in
training, we trained all methods on TRANCE and recorded Fig. 5, the model’s Reasoning Token Length initially de-
intermediate and validation results, as shown in Fig. 3. De- creases, then gradually increases before stabilizing. This
tailed results are provided in the Appendix. Key findings behavior coincides with the Format Reward reaching its ini-
include: (1) Reason-RFT exhibits strong data efficiency. In tial peak and the Accuracy Reward entering its rapid growth
the ID tasks of the 2B model, Reason-RFT achieves 70% of phase. We infer that the model simplifies its outputs early
the performance of Reason-RFT-Zero using only 3% of the on to quickly adapt to the structured response format, re-
training data (1,600 samples), reaching 82.5% with 9% of ducing Reasoning Token Length. Once the Format Reward
the data. (2) This efficiency generalizes to OOD tasks. In is maximized, the model shifts focus to improving accuracy,
the 7B model, Reason-RFT achieves over 92% of the per- increasing Reasoning Token Length.
formance of Reason-RFT-Zero using just 3% of the training Reasoning Redundancy This phenomenon refers to
data, demonstrating robust generalization capabilities. the significant difference in Reasoning Token Length be-

7
 ACC/%
Table 3. Results of different format reward strategies on the Spatial
Prompt: You are a
visual counting expert.
Given the <image> in
Transformation task.
front of you, if I
remove all the blocks,
then how many objects
are there in the table? ID OOD
Setting AVG
TRANCE TRANCE-L TRANCE-R
Direct Response: The final answer is 5.
Response Token Length Accuracy Reward Qwen2VL-2B-Instruct
Transient Adaptation Gap Reasoning Response: <think> Initially, there are 5 objects in
Reason-RFT-ZERO Reason-RFT the Reason-RFT-ZERO Reason-RFT
image: a yellow cylinder, a brown cube, a purple cylinder,
200
1
1.0
In-Domain a cyan sphere, and a gray cylinder. The brown cube is a block, Reason-RFT-Zero 42.13 34.07 33.41 33.74
so we need to remove it. The final answer is 5 – 1 = 4 </think> + visual tokens 42.01 36.05 35.97 38.01
Out-of-Domain
0.8
0.8 <answer> 4 </answer>
Steps
Reason-RFT 74.61 64.05 64.08 69.33
(a) 0.6
0.6 (b) + visual tokens 71.99 60.13 59.87 65.99
100
90 Figure 4. Illustration of the Transient Adaptation Gap. (a) 0.4
0.4 Qwen2VL-7B-Instruct
80
Almost the Same Accuracy
70 shows
Redundancy a sharp
Reasoning Token drop and recovery in both ID and OOD test per-
0.2
0.2 Reason-RFT-Zero 67.67 57.2 56.15 62.17
60 train/global_step train/global_step
train/global_step
formances within the early training steps during training on the
500 1k 1.5k 500
500 1k
1k 1.5k
1.5k
+ visual tokens 70.28 59.52 57.01 64.27
CLEVR-MATH (a) task. (b) presents a case(b)study of the prediction Reason-RFT 79.97 59.36 58.61 69.48
+ visual tokens 79.85 58.71 57.98 69.09
result on the 10-th step for further investigation.
Table 4. Results of different accuracy reward strategies on the
Response Token Length Format Reward Accuracy Reward
90 1.0 1.0
Spatial Transformation task.
80
0.9
70
Maximum Point 0.8
60 0.8
ID OOD
50 0.7 0.6 Setting α β AVG
TRANCE TRANCE-L TRANCE-R
40 0.6
0.4 Rapid Growth Point
Minimum Point
0.5 Qwen2VL-2B-Instruct
30 train/global_step train/global_step 0.2 train/global_step
0.4
100 200 300 400 500 20 40 60 80 100 20 40 60 80 100
Baseline 0 0 74.61 64.05 64.08 69.33
(a) 0.50 0.25 79.18 56.36 55.45 67.54
(a) (b) (c)
(b) -0.25 -0.50 73.69 64.41 64.72 69.13
Figure 5. Analysis of Greedy Reward Stratification. The Qwen2VL-7B-Instruct
model’s Reasoning Token Length initially decreases, then gradu-
Baseline 0 0 79.97 59.36 58.61 69.48
ally increases before stabilizing during Reason-RFT-Zero training. (a) 0.50 0.25 80.89 53.20 52.61 66.90
(a) illustrates the curve of Reasoning Token Length, highlighting (b) -0.25 -0.50 75.03 64.83 63.18 69.52
the Minimum Point. (b) shows the curve of Format Reward, in-
dicating the Maximum Point. (c) displays the curve of Accuracy Table 5. Results of various mixed CoT activation datasets of
Reward, marking the Rapid Growth Point. This trend aligns with Reason-RFT on the Structure Perception task.
the Format Reward reaching its initial peak and the Accuracy Re-
ward entering its rapid growth phase. Setting CoT Activation Data
ID OOD
AVG
GeoMath Geometry3k
Baseline GeoMath-only data 59.27 49.25 54.26
Response Token Length Accuracy Reward

Reason-RFT-ZERO Reason-RFT Reason-RFT-ZERO Reason-RFT (a) Mixed Specific-Domain data 50.61 45.35 48.02
1
1.0
200 (b) Mixed General-Domain data 42.51 40.25 41.38
0.8
0.8

0.6
0.6

100
90
80
0.4
0.4
Almost the Same Accuracy
During the Reasoning Reinforcement phase, the model re-
70 Redundancy Reasoning Token
60 train/global_step
0.2
0.2
train/global_step
train/global_step
tains these lengthy chains due to the lack of penalties or
1k 1k
500 1.5k 500
500 1k 1.5k
1.5k
incentives for response length. In contrast, Reason-RFT-
(a) (b)
Zero, which lacks CoT data, stabilizes at a shorter Reason-
Figure 6. Analysis of Reasoning Redundancy. (a) illustrates
the curves of Reasoning Token Length for Reason-RFT-Zero and
ing Token Length through exploration. We hypothesize that
Reason-RFT during
Response Tokentraining
Length on the Structure
Format Reward Perception
Accuracytask,
Reward
Reason-RFT’s longer reasoning chains may be unnecessary
90 1.0
highlighting
80 the significant difference
0.9 in Reasoning
1.0
Token Length for the current task difficulty, introducing redundant com-
70
between these
60 two paradigms 0.8in the later Maximum Point 0.8
steps of training. (b) putational overhead. Experiments limiting Reasoning To-
0.7 0.6
50
shows the curves of Accuracy Reward, 0.6
indicating that the accuracy ken Length during inference for Reason-RFT-trained mod-
40 0.4 Rapid Growth Point
Minimum Point
rates of both paradigms converge0.5 and remain nearly identical. els show stable performance until a certain threshold, par-
30 train/global_step train/global_step 0.2 train/global_step
0.4
100 200 300 400 500 20 40 60 80 100 20 40 60 80 100
tially confirming Reasoning Redundancy in specific tasks.
(a) (b) (c)
tween models trained under different paradigms. For ex-
ample, in the Structure Perception task, Reason-RFT and
4.6. Exploration on Reward Design
Reason-RFT-Zero achieve comparable final test accuracy, Exploration on Format Reward In DeepSeek-R1 [15], the
but Reason-RFT exhibits notably longer Reasoning Token Format Reward mandates the use of <think> tags for
Length, as shown in Fig. 6. This occurs because the CoT structuring reasoning processes in pure-textual tasks. To
data used in Reason-RFT’s Reasoning Activation phase, of- enhance visual reasoning tasks, we propose integrating vi-
ten distilled from stronger models (e.g., GPT-4o), leads the sual observations through captioning, extending the format
model to learn longer reasoning chains during CoT-SFT. reward with <summary> and <caption> tokens along-

8
side <think>. As shown in Tab. 3, this extension signifi- for assessing model performance across diverse scenar-
cantly improved the performance of Reason-RFT-Zero, but ios. Extensive experiments demonstrate the effectiveness
not Reason-RFT. We attribute this discrepancy to Reason- of Reason-RFT, providing valuable insights for advancing
RFT’s prior exposure to Reasoning Activation, which likely visual reasoning research and introducing a new paradigm
learned the caption-like guidance from a limited CoT rea- in multimodal learning.
soning samples, resulting in reduced effectiveness of addi-
tional tag guidance. Conversely, Reason-RFT-Zero, being References
purely RL-based, benefits more from effective tag guidance, [1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadal-
highlighting its greater potential for improvement. lah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree,
Exploration on Accuracy Reward We investigate the Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3
Accuracy Reward for the challenging Spatial Transforma- technical report: A highly capable language model locally
tion task, which requires predicting novel transformation on your phone. arXiv preprint arXiv:2404.14219, 2024. 6
sequences. The reward design in Eq. 7 includes coeffi- [2] Bo Adler, Niket Agarwal, Ashwath Aithal, Dong H Anh,
cients α and β for partial matches, influencing tolerance Pallab Bhattacharya, Annika Brundyn, Jared Casper, Bryan
for incomplete outputs. We test three configurations: (1) Catanzaro, Sharon Clay, Jonathan Cohen, et al. Nemotron-
α = 0, β = 0 (full matches only), (2) α = 0.50, β = 0.25 4 340b technical report. arXiv preprint arXiv:2406.11704,
(full and partial matches), and (3) α = −0.25, β = −0.50 2024. 3
(full matches with partial penalties). Experimental results [3] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna,
Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky,
on 2B and 7B models (Tab. 4) revealed: (1) Rewards for par-
Diogo Costa, Baudouin De Monicault, Saurabh Garg,
tial matches improved in-domain performance but harmed Theophile Gervet, et al. Pixtral 12b. arXiv preprint
out-of-domain generalization, suggesting “soft rewards” arXiv:2410.07073, 2024. 6
may limit adaptability; (2) Penalties for partial matches en- [4] Saeed Amizadeh, Hamid Palangi, Alex Polozov, Yichen
hanced performance on out-of-domain, indicating “hard re- Huang, and Kazuhito Koishida. Neuro-symbolic visual rea-
wards” are more effective for serialized tasks. soning: Disentangling. In International Conference on Ma-
chine Learning, pages 279–290. Pmlr, 2020. 2, 3
4.7. Exploration on COT Activation Data [5] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin
To investigate the impact of differently composed CoT ac- Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun
tivation data on Reason-RFT, we construct two distinct Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhao-
datasets: a mixed domain-specific dataset, which integrates hai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren
Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen
relevant yet distinct data from in-domain tasks, and a mixed
Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Jun-
general-domain dataset, encompassing a broader range of yang Lin. Qwen2.5-vl technical report. arXiv preprint
visual reasoning tasks (e.g., graph topology, visual puzzles). arXiv:2502.13923, 2025. 6
Using these datasets, we perform Reason-RFT training on [6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo
Structure Perception task, with the results detailed in Tab. 5. Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou
From this, two key points emerge: (1) As the proportion of Zhu, Lewei Lu, et al. Internvl: Scaling up vision founda-
in-domain training data decreases, the model’s performance tion models and aligning for generic visual-linguistic tasks.
on specific tasks declines; (2) Models trained on more di- In IEEE/CVF conference on computer vision and pattern
verse visual reasoning domain data may also exhibit a re- recognition, pages 24185–24198, 2024. 6
duction in domain-specific performance. [7] Minkyu Choi, Harsh Goel, Mohammad Omama, Yunhao
Yang, Sahil Shah, and Sandeep Chinchali. Towards neuro-
5. Conclusion symbolic video understanding. In European Conference on
Computer Vision, pages 220–236. Springer, 2024. 2, 3
In this paper, we propose Reason-RFT, a novel reinforce- [8] Tianzhe Chu, Yuexiang Zhai, Jihan Yang, Shengbang Tong,
ment fine-tuning framework that enhances the generaliza- Saining Xie, Dale Schuurmans, Quoc V Le, Sergey Levine,
tion capabilities of visual reasoning models. By integrating and Yi Ma. Sft memorizes, rl generalizes: A compara-
supervised fine-tuning (SFT) with Chain-of-Thought (CoT) tive study of foundation model post-training. arXiv preprint
arXiv:2501.17161, 2025. 3
reasoning activation data and Group Relative Policy Opti-
[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo
mization (GRPO)-based reinforcement learning, Reason-
Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe
RFT effectively mitigates key challenges such as overfit- Franke, Stefan Roth, and Bernt Schiele. The cityscapes
ting and cognitive rigidity, thereby improving cross-domain dataset for semantic urban scene understanding. In IEEE
transferability and real-world applicability. To support sys- conference on computer vision and pattern recognition,
tematic evaluation, we reconstruct a comprehensive dataset pages 3213–3223, 2016. 2
covering visual counting, structure perception, and spa- [10] Yuhao Dong, Zuyan Liu, Hai-Long Sun, Jingkang Yang,
tial transformation tasks, establishing a robust benchmark Winston Hu, Yongming Rao, and Ziwei Liu. Insight-v: Ex-

9
 ploring long-chain visual reasoning with multimodal large and Yu Liu. Mapfusion: A novel bev feature fusion network
language models. arXiv preprint arXiv:2411.14432, 2024. 3 for multi-modal map construction. Information Fusion, page
[11] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Ab- 103018, 2025. 2
hishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil [23] Xin Hong, Yanyan Lan, Liang Pang, Jiafeng Guo, and
Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The Xueqi Cheng. Transformation driven visual reasoning.
llama 3 herd of models. arXiv preprint arXiv:2407.21783, In IEEE/CVF Conference on computer vision and pattern
2024. 3 recognition, pages 6903–6912, 2021. 3, 5, 14
[12] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wan- [24] Yingdong Hu, Fanqi Lin, Tong Zhang, Li Yi, and Yang
jun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Gao. Look before you leap: Unveiling the power of gpt-
Xu, Zhenguo Li, et al. G-llava: Solving geometric prob- 4v in robotic vision-language planning. arXiv preprint
lem with multi-modal large language model. arXiv preprint arXiv:2311.17842, 2023. 3
arXiv:2312.11370, 2023. 3, 5, 13 [25] Huggingface. open-r1: Fully open reproduction of deepseek-
[13] Artur d’Avila Garcez, Marco Gori, Luis C Lamb, Luciano r1. [Link]
Serafini, Michael Spranger, and Son N Tran. Neural- r1, 2025. [Online; accessed: 2025-01-24]. 5, 14
symbolic computing: An effective methodology for princi- [26] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perel-
pled integration of machine learning and reasoning. arXiv man, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-
preprint arXiv:1905.06088, 2019. 2, 3 hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.
[14] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, arXiv preprint arXiv:2410.21276, 2024. 6, 7
Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi [27] Yuheng Ji, Yue Liu, Zhicheng Zhang, Zhao Zhang, Yuting
Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning Zhao, Gang Zhou, Xingwei Zhang, Xinwang Liu, and Xi-
capability in llms via reinforcement learning. arXiv preprint aolong Zheng. Advlora: Adversarial low-rank adaptation of
arXiv:2501.12948, 2025. 3 vision-language models. arXiv preprint arXiv:2404.13425,
[15] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, 2024. 2
Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi [28] Yuheng Ji, Huajie Tan, Jiayu Shi, Xiaoshuai Hao, et al.
Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning Robobrain: A unified brain model for robotic ma-
capability in llms via reinforcement learning. arXiv preprint nipulation from abstract to concrete. arXiv preprint
arXiv:2501.12948, 2025. 2, 8, 13 arXiv:arXiv:2502.21257, 2025. 2, 3
[16] Tanmay Gupta and Aniruddha Kembhavi. Visual program- [29] Justin Johnson, Bharath Hariharan, Laurens Van
ming: Compositional visual reasoning without training. In Der Maaten, Judy Hoffman, Li Fei-Fei, C Lawrence Zitnick,
IEEE/CVF Conference on Computer Vision and Pattern and Ross Girshick. Inferring and executing programs for
Recognition, pages 14953–14962, 2023. 2, 3 visual reasoning. In IEEE international conference on
[17] Peng Hao, Chaofan Zhang, Dingzhe Li, Xiaoge Cao, Xi- computer vision, pages 2989–2998, 2017. 3
aoshuai Hao, Shaowei Cui, and Shuo Wang. Tla: Tactile- [30] Mehran Kazemi, Hamidreza Alvari, Ankit Anand, Jialin Wu,
language-action model for contact-rich manipulation. arXiv Xi Chen, and Radu Soricut. Geomverse: A systematic evalu-
preprint arXiv:2503.08548, 2025. 3 ation of large models for geometric reasoning. arXiv preprint
[18] Xiaoshuai Hao, Yi Zhu, Srikar Appalaraju, Aston Zhang, arXiv:2312.12241, 2023. 3
Wanqian Zhang, Bo Li, and Mu Li. Mixgen: A new multi- [31] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon
modal data augmentation. In IEEE/CVF winter conference Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is
on applications of computer vision, pages 379–389, 2023. 3 worth a dozen images. In European Conference on Computer
[19] Xiaoshuai Hao, Ruikai Li, Hui Zhang, Dingzhe Li, Rong Vision, pages 235–251, 2016. 3
Yin, Sangil Jung, Seung-In Park, ByungIn Yoo, Haimei [32] Komal Kumar, Tajamul Ashraf, Omkar Thawakar,
Zhao, and Jing Zhang. Mapdistill: Boosting efficient Rao Muhammad Anwer, Hisham Cholakkal, Mubarak
camera-based hd map construction via camera-lidar fusion Shah, Ming-Hsuan Yang, Phillip HS Torr, Salman Khan,
model distillation. In European Conference on Computer Vi- and Fahad Shahbaz Khan. Llm post-training: A deep
sion, pages 166–183, 2024. 2 dive into reasoning large language models. arXiv preprint
[20] Xiaoshuai Hao, Mengchuan Wei, Yifan Yang, Haimei Zhao, arXiv:2502.21321, 2025. 3
Hui Zhang, Yi Zhou, Qiang Wang, Weiming Li, Lingdong [33] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng,
Kong, and Jing Zhang. Is your HD map constructor reliable Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao
under sensor corruptions? In Advances in Neural Informa- Zhang, and Ion Stoica. Efficient memory management for
tion Processing Systems, 2024. large language model serving with pagedattention. In ACM
[21] Xiaoshuai Hao, Hui Zhang, Yifan Yang, Yi Zhou, Sangil SIGOPS 29th Symposium on Operating Systems Principles,
Jung, Seung-In Park, and ByungIn Yoo. Mbfusion: A new 2023. 5, 15
multi-modal bev feature fusion method for hd map construc- [34] Nathan Lambert, Jacob Morrison, Valentina Pyatkin,
tion. In IEEE International Conference on Robotics and Au- Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester
tomation, pages 15922–15928, 2024. James V Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, et al.
[22] Xiaoshuai Hao, Yunfeng Diao, Mengchuan Wei, Yifan Yang, T\” ulu 3: Pushing frontiers in open language model post-
Peng Hao, Rong Yin, Hui Zhang, Weiming Li, Shu Zhao, training. arXiv preprint arXiv:2411.15124, 2024. 3

10
[35] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Rueckert. Medvlm-r1: Incentivizing medical reasoning ca-
Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Zi- pability of vision-language models (vlms) via reinforcement
wei Liu, et al. Llava-onevision: Easy visual task transfer. learning. arXiv preprint arXiv:2502.19634, 2025. 2
arXiv preprint arXiv:2408.03326, 2024. 6 [48] Maria MP Petrou and Costas Petrou. Image processing: the
[36] Dingzhe Li, Yixiang Jin, Yuhao Sun, Hongze Yu, Jun Shi, fundamentals. John Wiley & Sons, 2010. 2
Xiaoshuai Hao, Peng Hao, Huaping Liu, Fuchun Sun, Jian- [49] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christo-
wei Zhang, et al. What foundation models can bring for pher D Manning, Stefano Ermon, and Chelsea Finn. Direct
robot learning in manipulation: A survey. arXiv preprint preference optimization: Your language model is secretly a
arXiv:2404.18201, 2024. 2 reward model. Advances in Neural Information Processing
[37] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Ko- Systems, pages 53728–53741, 2023. 3
rtylewski, Wufei Ma, Benjamin Van Durme, and Alan L [50] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao
Yuille. Super-clevr: A virtual benchmark to diagnose domain Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li,
robustness in visual reasoning. In IEEE/CVF conference Y Wu, et al. Deepseekmath: Pushing the limits of mathe-
on computer vision and pattern recognition, pages 14963– matical reasoning in open language models. arXiv preprint
14973, 2023. 3, 5, 13 arXiv:2402.03300, 2024. 3
[38] Adam Dahlgren Lindström and Savitha Sam Abra- [51] Wenhao Shi, Zhiqiang Hu, Yi Bin, Junhua Liu, Yang Yang,
ham. Clevr-math: A dataset for compositional lan- See-Kiong Ng, Lidong Bing, and Roy Ka-Wei Lee. Math-
guage, visual and mathematical reasoning. arXiv preprint llava: Bootstrapping mathematical reasoning for multimodal
arXiv:2208.05358, 2022. 2, 3, 5, 13 large language models. arXiv preprint arXiv:2406.17294,
[39] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao 2024. 3, 5, 13
Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu
[52] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu,
Zhang, Chong Ruan, et al. Deepseek-v3 technical report.
Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui,
arXiv preprint arXiv:2412.19437, 2024. 3
Yu-Xiong Wang, Yiming Yang, et al. Aligning large multi-
[40] Pan Lu, Ran Gong, Shibiao Jiang, Liang Qiu, Siyuan Huang,
modal models with factually augmented rlhf. arXiv preprint
Xiaodan Liang, and Song-Chun Zhu. Inter-gps: Interpretable
arXiv:2309.14525, 2023. 3
geometry problem solving with formal language and sym-
[53] Dı́dac Surı́s, Sachit Menon, and Carl Vondrick. Vipergpt:
bolic reasoning. arXiv preprint arXiv:2105.04165, 2021. 5,
Visual inference via python execution for reasoning. In
13
IEEE/CVF International Conference on Computer Vision,
[41] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei
pages 11888–11898, 2023. 3
Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and
Ashwin Kalyan. Learn to explain: Multimodal reasoning via [54] Yingbo Tang, Shuaike Zhang, Xiaoshuai Hao, Pengwei
thought chains for science question answering. Advances in Wang, Jianlong Wu, Zhongyuan Wang, and Shanghang
Neural Information Processing Systems, pages 2507–2521, Zhang. Affordgrasp: In-context affordance reasoning for
2022. 3 open-vocabulary task-oriented grasping in clutter. arXiv
preprint arXiv:2503.00778, 2025. 3
[42] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li,
Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel [55] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell,
Galley, and Jianfeng Gao. Mathvista: Evaluating mathemat- Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent,
ical reasoning of foundation models in visual contexts. arXiv Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking
preprint arXiv:2310.02255, 2023. 3 multimodal understanding across millions of tokens of con-
[43] Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang text. arXiv preprint arXiv:2403.05530, 2024. 6
Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng [56] Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu
Chen, and Dongmei Zhang. Wizardmath: Empowering Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang
mathematical reasoning for large language models via rein- Du, Chonghua Liao, et al. Kimi k1. 5: Scaling reinforcement
forced evol-instruct. arXiv preprint arXiv:2308.09583, 2023. learning with llms. arXiv preprint arXiv:2501.12599, 2025.
3 2
[44] Meta AI. Llama 3 at connect 2024: Vision for edge and [57] Omkar Thawakar, Dinura Dissanayake, Ketan More, Ritesh
mobile devices, 2024. Accessed: 2025-02-15. 6 Thawkar, Ahmed Heakl, Noor Ahsan, Yuhao Li, Mohammed
[45] OpenAI. Learning to reason with llms. https : Zumri, Jean Lahoud, Rao Muhammad Anwer, et al. Llamav-
//[Link]/index/learning- to- reason- o1: Rethinking step-by-step visual reasoning in llms. arXiv
with-llms/, 2024. Accessed: 2025-03-02. 2 preprint arXiv:2501.06186, 2025. 2
[46] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Car- [58] Ke Wang, Houxing Ren, Aojun Zhou, Zimu Lu, Sichun Luo,
roll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Weikang Shi, Renrui Zhang, Linqi Song, Mingjie Zhan, and
Agarwal, Katarina Slama, Alex Ray, et al. Training lan- Hongsheng Li. Mathcoder: Seamless code integration in
guage models to follow instructions with human feedback. llms for enhanced mathematical reasoning. arXiv preprint
Advances in neural information processing systems, pages arXiv:2310.03731, 2023. 3
27730–27744, 2022. 3 [59] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan,
[47] Jiazhen Pan, Che Liu, Junde Wu, Fenglin Liu, Jiayuan Zhu, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin
Hongwei Bran Li, Chen Chen, Cheng Ouyang, and Daniel Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui

11
 Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Jun- Geoffrey Irving. Fine-tuning language models from human
yang Lin. Qwen2-vl: Enhancing vision-language model’s preferences. arXiv preprint arXiv:1909.08593, 2019. 3
perception of the world at any resolution. arXiv preprint
arXiv:2409.12191, 2024. 5, 6, 14
[60] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu,
Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and
Quoc V Le. Finetuned language models are zero-shot learn-
ers. arXiv preprint arXiv:2109.01652, 2021. 3
[61] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten
Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al.
Chain-of-thought prompting elicits reasoning in large lan-
guage models. Advances in neural information processing
systems, pages 24824–24837, 2022. 3
[62] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten
Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al.
Chain-of-thought prompting elicits reasoning in large lan-
guage models. Advances in neural information processing
systems, pages 24824–24837, 2022. 3
[63] Guowei Xu, Peng Jin, Li Hao, Yibing Song, Lichao Sun, and
Li Yuan. Llava-o1: Let vision language models reason step-
by-step. arXiv preprint arXiv:2411.10440, 2024. 2, 3
[64] Jihan Yang, Shusheng Yang, Anjali W Gupta, Rilyn Han,
Li Fei-Fei, and Saining Xie. Thinking in space: How mul-
timodal large language models see, remember, and recall
spaces. arXiv preprint arXiv:2412.14171, 2024. 2
[65] Simon Zhai, Hao Bai, Zipeng Lin, Jiayi Pan, Peter Tong,
Yifei Zhou, Alane Suhr, Saining Xie, Yann LeCun, Yi Ma,
et al. Fine-tuning large vision-language models as decision-
making agents via reinforcement learning. Advances in Neu-
ral Information Processing Systems, pages 110935–110971,
2025. 3
[66] Li-Ming Zhan, Bo Liu, Lu Fan, Jiaxin Chen, and Xiao-Ming
Wu. Medical visual question answering via conditional rea-
soning. In ACM International Conference on Multimedia,
pages 2345–2354, 2020. 2
[67] Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu,
Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan
Zhang, Jingkang Yang, Chunyuan Li, et al. Lmms-eval: Re-
ality check on the evaluation of large multimodal models.
arXiv preprint arXiv:2407.12772, 2024. 5
[68] Lingfeng Zhang, Xiaoshuai Hao, Qinwen Xu, Qiang Zhang,
Xinyao Zhang, Pengwei Wang, Jing Zhang, Zhongyuan
Wang, Shanghang Zhang, and Renjing Xu. Mapnav: A novel
memory representation via annotated semantic maps for
vlm-based vision-and-language navigation. arXiv preprint
arXiv:2502.13451, 2025. 3
[69] Mingyu Zhang, Jiting Cai, Mingyu Liu, Yue Xu, Cewu Lu,
and Yong-Lu Li. Take a step back: Rethinking the two stages
in visual reasoning. In European Conference on Computer
Vision, pages 124–141. Springer, 2024. 2, 3
[70] Renrui Zhang, Xinyu Wei, Dongzhi Jiang, Ziyu Guo,
Shicheng Li, Yichi Zhang, Chengzhuo Tong, Jiaming Liu,
Aojun Zhou, Bin Wei, et al. Mavis: Mathematical visual
instruction tuning with an automatic data engine. arXiv
preprint arXiv:2407.08739, 2024. 3
[71] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B
Brown, Alec Radford, Dario Amodei, Paul Christiano, and

12
Appendix multihop-subtraction. To evaluate out-of-domain general-
ization, we extended CLEVR-Math by enhancing the di-
This supplementary material provides additional details on versity of objects through the incorporation of 3D assets
the proposed method and experimental results that could not from Super-CLEVR [37], which leads to the creation of
be included in the main manuscript due to page limitations. Super-CLEVR-Math, an advanced benchmark with 1K test
Specifically, this appendix is organized as follows.
samples designed to assess model generalization under in-
• Sec. A provides more details on the evaluation of reason- creased complexity. These test samples are also divided into
ing tasks and discusses how we collected, filtered, and four task types: addition, subtraction, addition-subtraction,
reconstructed a high-quality dataset. and subtraction-multihop. Notably, the addition-subtraction
• Sec. B outlines the models and training processes, provid- type introduces a novel category consisting both addition
ing more detailed experimental specifics. and subtraction, which is not present in CLEVR-Math, fur-
• Sec. C presents comprehensive experimental results. ther elevating the benchmark’s challenge and diversity.
• Sec. D includes more visualization cases. Reward Design Following the reward methodology of
DeepSeek-R1 [15], we define two distinct reward functions:
A. Details of Evaluation Reasoning Tasks Format Reward and Accuracy Reward. The Format Reward
A.1. Visual Counting is assigned a value of 1 if the response adheres to the
predefined template structure, specifically in the form of
Task Definition Visual Counting is a multi-modal reason- <think>...</think><answer>...</answer>;
ing task that evaluates the integration of language, visual, otherwise, it is assigned a value of 0. The Accuracy Reward
and mathematical capabilities by solving basic arithmetic is assigned a value of 1 if the numerical counting result in
problems in visual scenes featuring 3D blocks with diverse the response is correct; otherwise, it is assigned a value
attributes like color, size, material, and shape. The task in- of 0. This dual-reward mechanism ensures both structural
cludes four different types: 1) Subtraction Counting after compliance and numerical accuracy in model responses.
removing specific object subsets, 2) Addition Counting af-
ter inserting objects with specified quantities and attributes, A.2. Structure Perception
3) Adversarial Designed as trick questions where actions
are performed on one object, but the query targets an unaf- Task Definition Structure Perception represents a complex
fected object, and 4) Multi-Hop Counting after sequential class of visual mathematical reasoning tasks, which focuses
steps by addition or subtraction action. This task challenges on assessing the model’s capacity to determine geomet-
models to perform attribute-based reasoning in dynamic vi- ric structure relationships and perform calculations involv-
sual contexts, emphasizing cross-modal understanding and ing angles, lengths, areas, and other geometric properties.
reasoning capabilities. Some examples are shown in Fig. 7. The task includes problems such as identifying congruent
or similar shapes, calculating perimeters and areas, deter-
mining angles between lines or shapes, and solving prob-
In-Domain Out-of-Domain
lems related to geometric transformations (e.g., rotations,
translations, and reflections). By combining mathematical
rigor with visual reasoning, this task challenges models to
demonstrate a deep understanding of geometric principles
in both abstract and real-world scenarios. Some examples
are shown in Fig. 8.
Dataset Preparation For the in-domain dataset, we uti-
Prompt: You are a visual counting lized GeoMath-8K, a dataset specifically designed for ge-
Prompt: You are a visual counting
expert. Given the <image> in front
expert. Given the <image> in front
of you, if I remove all the small
of you, if I add two large buses, and ometric problem-solving, which is constructed based on
metal balls, and then remove all
then remove all the bikes, now how
the yellow objects, now how many
many buses are there in the table?
Math360K [51] and Geo170K [12]. To ensure data quality,
objects are there in the table?
we employed GPT-4o to filter out incorrect samples and re-
moved those with answers that were neither numerical nor
Figure 7. The sample of Visual Counting. included in the provided options, thereby streamlining the
validation process during training and testing. This refine-
Dataset Preparation For in-domain dataset, we refined ment process resulted in a curated dataset consisting of 4.5K
the original dataset from CLEVR-Math [38] by filtering training samples and 820 test samples. For out-of-domain
out low-quality or incorrect samples using GPT-4o, result- evaluation, we selected 800 samples from Geometry3K [40]
ing in a clean dataset comprising 35K training samples (including 400 multiple-choice and 400 open-ended ques-
and 1K test samples. These samples are categorized into tions) to comprehensively assess the model’s generalization
four specific types: subtraction, addition, adversarial, and capabilities on geometry reasoning.

13
 In-Domain Out-of-Domain from front, left, and right perspectives. To ensure high data
quality, we implemented a rigorous filtering process: (1)
Angle ? Angle ?
removing samples containing occluded or invisible objects
in either the initial or final states, (2) eliminating redun-
122°
X
dant actions within the transformation sequences, and (3)
consolidating multi-step displacement actions, which col-
lectively ensure the uniqueness and correctness of the solu-
tions. The refined dataset consists of 60K training samples
Volume ? Area ?
and 6K test samples. For the training set, we constructed the
S1
S1 Trans-Center-60K dataset using the Center-Center configu-
ration, which pairs front-view initial and final state images.
For in-domain evaluation, we derived the Trans-Center-6K
dataset from the 6K test samples under the same Center-
Figure 8. The sample of Structure Perception. Center configuration. To evaluate out-of-domain gener-
alization, we constructed two additional datasets: Trans-
Left-6K and Trans-Right-6K, leveraging the Center-Left
Reward Design We maintain the same Format Reward
and Center-Right configurations to systematically assess the
as used in the Visual Counting task above. The Accu-
racy Reward is extended to support the evaluation of both model’s generalization capabilities in spatial reasoning un-
multiple-choice questions and mathematical expressions, der diverse viewpoint conditions.
ensuring comprehensive assessment across various problem Reward Design For the Format Reward, we adopted the
types. Specifically, mathematical reward type is designed same formulation as used in the Visual Counting task. As
for Structure Perception tasks involving numerical answers, for the Accuracy Reward, a specialized design was devel-
such as floating-point values or LaTeX-formatted expres- oped to evaluate the sequence of transformation functions.
sions. It uses a tolerance-based evaluation to account for Function-based type is designed for Spatial Transformation
minor numerical deviations. The accuracy reward Racc (ai ) tasks requiring a sequence of transformation functions. The
is defined as: accuracy reward Racc (ai ) evaluates the alignment between
    the predicted sequence Tpred and the ground truth Tgt , com-
1 |apred − agt | − ϵ1 × |agt | puted as:
Racc (ai ) = cos π × + 1 , (8)
2 (ϵ2 − ϵ1 ) × |agt |
f +o+v f +o/v f
len(Tpred ) + α · len(Tpred ) + β · len(Tpred )
where apred is the predicted answer, agt is the ground truth, Racc (ai ) = ,
max(len(Tpred ), len(Tgt ))
ϵ1 is the tolerance threshold for an exact match (e.g., 0.05), (9)
and ϵ2 is the upper bound for partial rewards (e.g., 0.20). If f +o+v
where Tpred is the subset of transformation steps with
|apred − agt | < ϵ1 × |agt |, the reward is 1 (exact match); f +o/v
complete matches (w/ function, object, and value), Tpred
if |apred − agt | > ϵ2 × |agt |, the reward is 0 (incorrect).
are the subsets with partial and only-function matches (w/
This formulation ensures smooth transitions between full f
function and object, or w/ function and value), Tpred is the
and partial rewards, enabling fair numerical evaluation.
subset with only-function matches. α and β are the weight-
A.3. Spatial Transformation ing coefficients for partial matches. This formulation en-
sures nuanced evaluation, rewarding both exact and par-
Task Definition Spatial Transformation is a spatial-visual
tially correct responses while allowing flexible adjustment
reasoning task designed to infer single-step or multi-step
of partial match contributions.
transformation actions by analyzing the initial and final
System Prompts For the Spatial Transformation task,
visual states from multiple perspectives (e.g., center, left,
we designed two versions of the system prompt. The
right). The task utilizes transformation functions, including
first version specifies the answer output format using the
change size, change color, change material,
<think> and <answer> tags, while the second version
change shape, and change position, to modify ob-
includes additional outputs <summary> and <caption>
ject properties such as size, color, material, shape, and
for experiments on exploration of format reward design in
position using predefined values. This task evaluates the
main paper. These two versions are illustrated in Fig. 10
model’s ability to reason about spatial relationships and ob-
and Fig. 11, respectively.
ject transformations across diverse viewpoints in dynamic
visual scenarios. Some examples are shown in Fig. 9.
B. Details of Models and Training
Dataset Preparation We generated 100K samples us-
ing the environment and configuration from Trance [23], We utilize Qwen2-VL-2B and Qwen2-VL-7B [59] as the
with each sample comprising initial object attributes, front- backbone models for our experiments. Our implementation
view image of initial state, and images of final state captured is built on the open-source frameworks Open-R1 [25] and

14
 Answer: 0 0
<think> Alright, the changes in the objects from the 1 2
Out-of-Domain Views
initial to the final state are as follows: the large red
metal cube in the initial state becomes a medium . . . 2
3 3
and keeping the other objects unchanged. </think> 1
<answer> ChangeSize (5, medium), 4 4
5 5
ChangeMaterial (3, rubber),
6 6
ChangePosition (2, front_left),
ChangePosition (1, behind) </answer>

Function List Final Left-View

ChangeColor (obj, value)

ChangePosition (obj, value)

Final Right-View
ChangeShape (obj, value) Initial Center-View Final Center-View

Prompt: You are a spatial visual reasoning expert. Given the

ChangeMaterial (obj, value)
center-view <image> of initial state, the center-view <image> (or
the left-view <image> or the right-view <image>) of final state,
ChangeSize (obj, value) and the {Function List} to be used, you should determine a
sequence of transformations that can achieve the change of states.

Figure 9. The sample of Spatial Transformation.

vLLM [33], ensuring reproducibility and scalability. All imately 70% of the performance of Reason-RFT-Zero with
experiments were conducted on a cluster of servers, each only 3% of the training data (1,600 samples), and 82.5%
equipped with 8×A800 GPUs. For the Visual Counting with just 9Robust Generalization to OOD Tasks: In the
task and Spatial Transformation task, we trained the mod- 7B model, Reason-RFT achieves over 92% of Reason-
els for 1 epoch each on their respective training datasets, RFT-Zero’s performance using just 3% of the training data,
ensuring sufficient exposure to task-specific patterns while showcasing its strong generalization capabilities. Com-
avoiding overfitting. For the Structure Perception task, due parison Across Methods: Reason-RFT consistently outper-
to its GeoMath training dataset consisting of a relatively forms other methods, particularly in data-constrained sce-
small number of training samples (a total of 4,500), we ex- narios, indicating its suitability for applications with limited
tended the training duration to 6 epochs to allow the models data availability. Performance Saturation: As training sam-
to fully capture the underlying structural and geometric re- ple size increases, some methods experience performance
lationships. In the Reason-RFT training pipeline, all models plateaus, suggesting that beyond a certain point, additional
underwent an initial CoT activation stage with 1,600 sam- data yields diminishing returns.
ples before proceeding to the RL phase. In conclusion, the evaluation of performance across dif-
ferent training samples not only highlights the strengths of
Reason-RFT in terms of data efficiency and generalization
C. More Experiment Results
but also provides critical insights into the performance dy-
Performance at Different Training Samples Fig. 12 and namics of various methods. These findings are essential for
Fig. 13 illustrate the In-Domain (ID) and Out-of-Domain practitioners aiming to maximize performance while effec-
(OOD) performance of all methods across three visual rea- tively managing training resources.
soning tasks, evaluated at various training sample sizes.
This analysis helps us understand how each method scales D. Visualization
with training data. More detail evaluation results for each In this section, we present additional visualization results,
subset of three tasks are in Tab. 6 - Tab. 13. see Fig. 14 - Fig. 22. Reason-RFT demonstrates superior
We systematically varied the number of training sam- performance over CoT-SFT in terms of logical consistency,
ples, from minimal to substantial, allowing us to identify reasoning quality, and correctness. CoT-SFT’s flaws stem
performance thresholds and data efficiency for each method from incorrect assumptions and misinterpretations, high-
in both ID and OOD contexts. Key findings from this anal- lighting the importance of accurate problem interpretation
ysis include: Data Efficiency of Reason-RFT: Reason-RFT and reasoning in visual reasoning tasks.
demonstrates exceptional data efficiency, achieving approx-

15
 System Prompt for Spatial Transformation Task

'''Your need to complete the spatial visual reasoning task according to the following rules.

### Task Description:

Given the image of the initial state, the image of the final state, and the attributes of the initial objects, you should determine a transformation that
can achieve the change of states.

The **attributes of the initial objects** are provided as a list of tuples in the following format:
**('object_id', 'shape', 'size', 'color', 'material')**
Each tuple represents an object and its properties in the initial state.

The transformation should be a sequence of functions with a length ranging from 1 to 4, where each function is represented as **'func(object_id,
value)'**.

### Available functions and values:

1. **'change_size(object_id, value)'** - Changes the object to a new size relative to its initial size.
- Possible values: `['small', 'medium', 'large']`

2. **'change_color(object_id, value)'** - Changes the object to a new color relative to its initial color.
- Possible values: `['yellow', 'gray', 'cyan', 'blue', 'brown', 'green', 'red', 'purple']`

3. **'change_material(object_id, value)'** - Changes the object to a new material relative to its initial material.
- Possible values: `['glass', 'metal', 'rubber']`

4. **'change_shape(object_id, value)'** - - Changes the object to a new shape relative to its initial shape.
- Possible values: `['cube', 'sphere', 'cylinder']`

5. **'change_position(object_id, value)'** - Moves the object to a new position relative to its initial location.
- Possible values: `['front', 'behind', 'left', 'right', 'front_left', 'front_right', 'behind_left', 'behind_right']`
- 'front' means moving forward along the object's initial direction.
- 'behind' means moving backward along the object's initial direction.
- 'left' means moving to the left of the object's initial orientation.
- 'right' means moving to the right of the object's initial orientation.
- 'front_left' means moving diagonally toward the front and left of the initial location.
- 'front_right' means moving diagonally toward the front and right of the initial location.
- 'behind_left' means moving diagonally toward the behind and left of the initial location.
- 'behind_right' means moving diagonally toward the behind and right of the initial location.

### Output Format

You should first thinks about the reasoning process internally and then provides the user with the answer. The **reasoning process** and
**answer** are enclosed within specific tags:

- **Reasoning process**: Provide a chain-of-thought, logical explanation of the problem. This should outline step-by-step reasoning, enclosed
within `<think>...</think>`

- **Final answer (sequence of functions only)**: Enclosed within `<answer>...</answer>`

Now, it's your turn!

{Question} Output the thinking process in <think> </think> and final answer in <answer> </answer> tags.
'''

Figure 10. The system prompt used in Spatial Transformation task.

16
 System Prompt for Spatial Transformation Task (Add <summary> <caption> in FORMAT)

'''Your need to complete the spatial visual reasoning task according to the following rules.

### Task Description:

Given the image of the initial state, the image of the final state, and the attributes of the initial objects, you should determine a transformation that
can achieve the change of states.

The **attributes of the initial objects** are provided as a list of tuples in the following format:
**('object_id', 'shape', 'size', 'color', 'material')**
Each tuple represents an object and its properties in the initial state.

The transformation should be a sequence of functions with a length ranging from 1 to 4, where each function is represented as **'func(object_id,
value)'**.

### Available functions and values:

1. **'change_size(object_id, value)'** - Changes the object to a new size relative to its initial size.
- Possible values: `['small', 'medium', 'large']`

2. **'change_color(object_id, value)'** - Changes the object to a new color relative to its initial color.
- Possible values: `['yellow', 'gray', 'cyan', 'blue', 'brown', 'green', 'red', 'purple']`

3. **'change_material(object_id, value)'** - Changes the object to a new material relative to its initial material.
- Possible values: `['glass', 'metal', 'rubber']`

4. **'change_shape(object_id, value)'** - - Changes the object to a new shape relative to its initial shape.
- Possible values: `['cube', 'sphere', 'cylinder']`

5. **'change_position(object_id, value)'** - Moves the object to a new position relative to its initial location.
- Possible values: `['front', 'behind', 'left', 'right', 'front_left', 'front_right', 'behind_left', 'behind_right']`
- 'front' means moving forward along the object's initial direction.
- 'behind' means moving backward along the object's initial direction.
- 'left' means moving to the left of the object's initial orientation.
- 'right' means moving to the right of the object's initial orientation.
- 'front_left' means moving diagonally toward the front and left of the initial location.
- 'front_right' means moving diagonally toward the front and right of the initial location.
- 'behind_left' means moving diagonally toward the behind and left of the initial location.
- 'behind_right' means moving diagonally toward the behind and right of the initial location.

### Output Format

You should first thinks about the reasoning process internally and then provides the user with the answer. The **reasoning process** and
**answer** are enclosed within specific tags:

- **Summary process**: Summary how you will approach the problem and explain the steps you will take to reach the answer, enclosed within
`<summary>...</summary>`

- **Caption process**: Provide a detailed description of the image, particularly emphasizing the aspects related to the question, enclosed within
`<caption>...</caption>`

- **Reasoning process**: Provide a chain-of-thought, logical explanation of the problem. This should outline step-by-step reasoning, enclosed
within `<think>...</think>`

- **Final answer (sequence of functions only)**: Enclosed within `<answer>...</answer>`

Now, it's your turn!

{Question} Output the summary process in <summary> </summary>, caption process in <caption>...</caption>, thinking process in <think>
</think> and final answer in <answer> </answer> tags.
'''

Figure 11. The system prompt used in Spatial Transformation task w/ <summary> and <caption> tags in format.

17
 Visual-Counting Structure-Perception Spatial-Transformation

c
In-Domain
Out-of-Domain

ANS-SFT COT-SFT Reason-RFT-ZERO Reason-RFT Zero-Shot Activation

Figure 12. Results of all methods on Qwen2VL-2B-Instruct, ID and OOD performance at different training checkpoints.

Visual-Counting Structure-Perception Spatial-Transformation

In-Domain
Out-of-Domain

ANS-SFT COT-SFT Reason-RFT-ZERO Reason-RFT Zero-Shot Activation

Figure 13. Results of all methods on Qwen2VL-7B-Instruct, ID and OOD performance at different training checkpoints.

18
 Visual Counting
Methods Steps Clevr-Math (ID)
adversarial sub-multi addition subtraction AVG
Zero-Shot - 93.60 84.00 55.60 96.40 82.40
100 83.60 56.40 91.20 81.60 78.20
200 69.20 67.60 91.60 82.00 77.60
400 81.60 65.60 90.80 84.80 80.70
600 72.40 73.20 92.40 89.20 81.80
ANS-SFT
800 78.40 77.20 82.80 90.40 82.20
1200 85.60 78.00 91.60 95.60 87.70
1600 92.80 82.40 94.80 96.80 91.70
2187 95.20 92.80 97.60 99.20 96.20
100 49.20 40.00 82.00 69.20 60.10
200 65.20 55.60 88.00 76.40 71.30
400 66.00 57.20 90.00 79.60 73.20
600 67.20 59.20 87.20 82.80 74.10
CoT-SFT
800 77.60 61.60 92.40 85.20 79.20
1200 76.80 70.00 91.20 93.60 82.90
1600 80.80 66.80 91.60 92.00 82.80
2187 83.20 71.20 93.20 94.40 85.50
100 92.80 88.80 94.40 96.00 93.00
200 95.60 91.60 95.60 97.60 95.10
400 92.00 87.60 84.00 96.40 90.00
600 94.40 92.80 93.60 96.00 94.20
Reason-RFT-Zero
800 96.40 96.40 96.00 98.80 96.90
1200 98.40 95.60 100.00 99.60 98.40
1600 96.40 94.80 98.80 99.60 97.40
2500 98.40 95.60 99.60 100.00 98.40
100 89.60 73.20 93.60 95.60 88.00
200 89.20 78.00 95.20 96.40 89.70
400 92.80 82.40 95.20 97.60 92.00
600 94.80 86.00 96.80 97.20 93.70
Reason-RFT
800 96.80 88.40 96.80 98.80 95.20
1200 94.80 86.00 96.40 98.80 94.00
1600 94.40 91.60 97.20 99.60 95.70
2500 98.40 92.80 96.80 99.20 96.80

Table 6. Complete experimental results of Qwen2VL-2B-Instruct on the Clevr-Math test set after training on Clevr-Math. “sub-multi”
donates the subtraction-multihop task.

19
 Visual Counting
Methods Steps Super-Clevr (OOD)
addition subtraction add-sub sub-multi AVG Encountered UnEncountered
Zero-Shot - 10.40 54.40 0.00 63.20 32.00 42.67 0.00
100 51.20 37.60 11.60 39.20 34.90 42.67 11.60
200 38.40 55.60 8.40 15.60 29.50 36.53 8.40
400 40.80 45.20 5.60 35.20 31.70 40.40 5.60
600 41.20 61.60 8.00 35.60 36.60 46.13 8.00
ANS-SFT
800 49.20 50.40 7.20 26.00 33.20 41.87 7.20
1200 44.00 53.20 5.60 38.80 35.40 45.33 5.60
1600 48.80 53.60 6.00 26.00 33.60 42.80 6.00
2187 49.60 62.00 5.20 41.60 39.60 51.07 5.20
100 47.20 50.00 28.80 25.60 37.90 40.93 28.80
200 56.00 52.40 38.00 34.00 45.10 47.47 38.00
400 55.20 57.20 22.40 30.40 41.30 47.60 22.40
600 58.40 55.20 24.00 35.60 43.30 49.73 24.00
CoT-SFT
800 57.60 47.60 26.80 41.60 43.40 48.93 26.80
1200 58.00 54.40 35.60 32.40 45.10 48.27 35.60
1600 53.20 58.40 33.20 40.40 46.30 50.67 33.20
2187 53.60 58.80 36.80 36.80 46.50 49.73 36.80
100 46.00 65.20 6.80 58.80 44.20 56.67 6.80
200 48.80 66.00 9.20 57.60 45.40 57.47 9.20
400 42.00 71.20 8.40 50.80 43.10 54.67 8.40
600 47.20 65.20 7.60 47.60 41.90 53.33 7.60
Reason-RFT-Zero
800 56.40 69.20 6.80 55.20 46.90 60.27 6.80
1200 52.00 73.60 7.20 59.20 48.00 61.60 7.20
1600 51.60 71.60 6.40 54.80 46.10 59.33 6.40
2500 49.60 71.20 5.20 53.20 44.80 58.00 5.20
100 59.20 57.60 38.00 41.60 49.10 52.80 38.00
200 59.60 64.40 39.20 42.00 51.30 55.33 39.20
400 61.60 64.00 39.20 37.20 50.50 54.27 39.20
600 66.80 67.20 32.00 46.00 53.00 60.00 32.00
Reason-RFT
800 66.00 65.60 34.00 39.20 51.20 56.93 34.00
1200 67.20 65.20 33.60 40.80 51.70 57.73 33.60
1600 63.60 66.00 33.20 44.80 51.90 58.13 33.20
2500 68.00 67.20 28.40 44.80 52.10 60.00 28.40

Table 7. Complete experimental results of Qwen2VL-2B-Instruct on the Super-Clevr test set after training on Clevr-Math. “add-sub”
donates the addition-subtraction task, while “sub-multi” donates the subtraction-multihop task.

20
 Visual Counting
Methods Steps Clevr-Math (ID)
adversarial sub-multi addition subtraction AVG
Zero-Shot - 99.60 98.40 97.60 98.80 98.60
100 69.20 54.00 81.20 69.20 68.40
200 61.20 50.00 82.40 75.60 67.30
400 69.20 63.60 89.20 77.60 74.90
600 70.40 54.00 90.40 81.20 74.00
ANS-SFT
800 80.00 74.00 91.20 89.20 83.60
1200 86.80 79.20 94.40 91.20 87.90
1600 90.40 84.40 95.20 92.00 90.50
2187 96.80 89.20 96.80 97.20 95.00
100 81.60 63.60 91.20 83.60 80.00
200 80.00 64.00 92.00 88.80 81.20
400 72.40 66.00 88.80 79.60 76.70
600 77.60 66.00 94.40 89.20 81.80
CoT-SFT
800 78.40 65.20 94.00 87.20 81.20
1200 79.60 76.80 92.40 88.00 84.20
1600 86.40 78.00 92.80 93.20 87.60
2187 87.20 78.80 93.60 89.60 87.30
100 98.00 94.40 98.80 99.60 97.70
200 99.60 93.20 99.20 100.00 98.00
400 99.60 95.20 99.60 98.80 98.30
600 98.00 98.40 100.00 99.60 99.00
Reason-RFT-Zero
800 99.60 98.40 99.60 98.80 99.10
1200 100.00 98.00 99.60 99.20 99.20
1600 99.60 97.60 100.00 99.20 99.10
2500 99.60 98.40 100.00 99.60 99.40
100 88.80 79.20 95.60 94.40 89.50
200 92.00 80.00 96.40 95.20 90.90
400 94.40 84.40 96.00 95.60 92.60
600 92.80 84.00 96.40 97.60 92.70
Reason-RFT
800 92.80 85.20 96.80 96.40 92.80
1200 94.80 89.60 97.20 97.60 94.80
1600 94.80 86.40 97.60 97.20 94.00
2500 96.80 88.40 99.20 98.00 95.60

Table 8. Complete experimental results of Qwen2VL-7B-Instruct on the Clevr-Math test set after training on Clevr-Math. “sub-multi”
donates the subtraction-multihop task.

21
 Visual Counting
Methods Steps Super-Clevr (OOD)
addition subtraction add-sub sub-multi AVG Encountered UnEncountered
Zero-Shot - 46.80 75.20 4.80 41.60 42.10 54.53 4.80
100 57.60 41.20 5.60 46.40 37.70 48.40 5.60
200 42.00 38.80 8.00 33.60 30.60 38.13 8.00
400 37.20 46.40 5.20 31.60 30.10 38.40 5.20
600 32.00 44.80 12.40 19.20 27.10 32.00 12.40
ANS-SFT
800 38.80 38.00 6.80 37.20 30.20 38.00 6.80
1200 42.00 42.80 12.80 32.00 32.40 38.93 12.80
1600 36.40 48.40 11.20 17.20 28.30 34.00 11.20
2187 39.60 58.80 8.00 29.20 33.90 42.53 8.00
100 60.00 63.60 44.00 41.60 52.30 55.07 44.00
200 67.60 66.40 48.00 46.80 57.20 60.27 48.00
400 55.20 60.40 19.60 42.00 44.30 52.53 19.60
600 64.80 61.20 35.20 43.20 51.10 56.40 35.20
CoT-SFT
800 60.00 53.60 37.60 42.40 48.40 52.00 37.60
1200 51.20 56.00 35.20 39.60 45.50 48.93 35.20
1600 53.20 56.40 34.40 35.20 44.80 48.27 34.40
2187 51.60 51.60 33.60 32.80 42.40 45.33 33.60
100 58.80 82.80 24.00 62.40 57.00 68.00 24.00
200 56.00 83.20 18.80 50.00 52.00 63.07 18.80
400 62.40 79.60 22.80 37.60 50.60 59.87 22.80
600 61.20 85.20 17.20 49.20 53.20 65.20 17.20
Reason-RFT-Zero
800 52.80 86.80 20.40 52.00 53.00 63.87 20.40
1200 53.60 83.20 19.20 46.80 50.70 61.20 19.20
1600 61.20 84.80 18.40 43.20 51.90 63.07 18.40
2500 59.20 86.40 21.20 45.20 53.00 63.60 21.20
100 53.60 56.80 33.20 39.60 45.80 50.00 33.20
200 52.00 61.20 31.60 44.00 47.20 52.40 31.60
400 56.00 59.60 30.80 45.20 47.90 53.60 30.80
600 56.00 64.00 31.60 50.00 50.40 56.67 31.60
Reason-RFT
800 56.00 60.00 28.00 41.60 46.40 52.53 28.00
1200 66.00 65.60 38.00 50.40 55.00 60.67 38.00
1600 64.40 59.60 32.40 48.80 51.30 57.60 32.40
2500 62.80 60.80 35.60 44.80 51.00 56.13 35.60

Table 9. Complete experimental results of Qwen2VL-7B-Instruct on the Super-Clevr test set after training on Clevr-Math. “add-sub”
donates the addition-subtraction task, while “sub-multi” donates the subtraction-multihop task.

22
 Structure Perception
Methods Steps Geometry3k (OOD) GeoMath (ID)
CHOICE NON-CHOICE AVG CHOICE NON-CHOICE AVG
Zero-Shot - 40.25 1.00 20.63 35.57 20.31 25.86
100 35.25 16.25 25.75 58.72 29.89 40.37
200 33.25 17.50 25.38 56.38 35.44 43.05
400 30.75 17.00 23.88 64.77 35.06 45.86
600 - - - 73.83 38.12 51.10
ANS-SFT
800 32.75 16.00 24.38 72.15 36.40 49.39
1200 - - - 73.83 35.44 49.39
1600 29.00 16.00 22.50 74.83 37.36 50.98
1686 28.75 16.25 22.50 74.83 37.93 51.34
100 16.50 21.50 19.00 31.54 34.10 33.17
200 7.50 23.50 15.50 32.89 35.25 34.39
400 21.50 21.25 21.38 41.61 40.04 40.61
600 - - - 43.62 36.59 39.14
CoT-SFT
800 16.50 23.50 20.00 45.97 39.27 41.70
1200 - - - 53.02 40.04 44.76
1600 24.25 24.00 24.13 53.69 37.93 43.66
1686 26.75 23.75 25.25 51.34 38.31 43.05
100 32.25 17.75 25.00 41.61 31.23 35.00
200 33.00 18.50 25.75 48.99 35.06 40.12
400 41.50 23.50 32.50 52.68 34.87 41.34
600 37.00 22.75 29.88 60.74 37.55 45.98
Reason-RFT-Zero
800 42.25 25.00 33.63 62.42 40.42 48.42
1200 43.00 23.75 33.38 61.07 39.66 47.44
1600 42.75 22.25 32.50 63.09 38.31 47.32
1610 43.25 21.75 32.50 63.09 38.89 47.68
100 37.50 23.25 30.38 50.34 41.00 44.39
200 33.50 29.25 31.38 56.71 40.04 46.10
400 38.25 28.75 33.50 56.38 39.27 45.49
600 40.50 27.25 33.88 61.41 41.19 48.54
Reason-RFT
800 41.25 29.50 35.38 58.05 41.19 47.32
1200 40.25 31.00 35.63 61.74 42.34 49.39
1600 38.00 29.25 33.63 62.08 43.10 50.00
1610 36.75 29.50 33.13 60.74 42.34 49.03

Table 10. Complete experimental results of Qwen2VL-2B-Instruct on the Structure Perception task after training on GeoMath.

23
 Structure Perception
Methods Steps Geometry3k (OOD) GeoMath (ID)
CHOICE NON-CHOICE AVG CHOICE NON-CHOICE AVG
Zero-Shot - 45.25 23.00 34.13 61.07 38.12 46.46
100 38.50 18.25 28.38 64.77 34.87 45.74
200 32.50 22.75 27.63 69.46 35.25 47.68
400 - - - 72.48 40.42 52.07
600 32.25 18.00 25.13 73.49 39.27 51.71
ANS-SFT
800 - - - 75.50 37.93 51.58
1200 32.50 18.50 25.50 75.84 37.74 51.59
1600 32.50 18.25 25.38 75.84 37.36 51.34
1686 18.25 38.75 28.50 38.59 42.72 41.22
100 6.50 32.00 19.25 38.26 43.10 41.34
200 27.00 34.50 30.75 56.71 44.64 49.03
400 - - - 52.68 44.06 47.19
600 35.50 36.25 35.88 63.09 43.49 50.61
CoT-SFT
800 - - - 63.42 42.91 50.36
1200 29.50 37.50 33.50 64.09 44.06 51.34
1600 29.25 36.75 33.00 61.74 44.06 50.49
1686 58.50 41.75 50.13 56.71 45.98 49.88
100 59.00 44.25 51.63 63.42 45.21 51.83
200 62.00 43.00 52.50 70.47 45.40 54.51
400 - - - 70.13 46.74 55.24
600 64.75 45.25 55.00 70.47 49.23 56.95
Reason-RFT-Zero
800 - - - 66.11 46.17 53.42
1200 69.00 43.25 56.13 71.14 45.59 54.88
1600 66.25 43.25 54.75 69.80 46.55 55.00
1610 46.75 37.50 42.13 67.79 45.79 53.79
100 53.00 37.00 45.00 72.82 46.93 56.34
200 52.75 37.25 45.00 71.14 46.55 55.49
400 51.50 37.00 44.25 73.49 48.28 57.44
600 56.75 37.25 47.00 77.52 46.17 57.56
Reason-RFT
800 59.00 40.00 49.50 79.87 48.08 59.63
1200 56.00 39.50 47.75 74.50 49.62 58.66
1600 59.00 39.50 49.25 78.52 48.28 59.27
1610 59.00 39.50 49.25 78.52 48.28 59.27

Table 11. Complete experimental results of Qwen2VL-7B-Instruct on the Structure Perception task after training on GeoMath.

24
 Spatial Transformation
Method Steps TRANCE (ID) TRANCE-L (OOD) TRANCE-R (OOD)
Level-1 Level-2 Level-3 Level-4 AVG Level-1 Level-2 Level-3 Level-4 AVG Level-1 Level-2 Level-3 Level-4 AVG
GPT-4o / 47.28 42.96 40.87 39.08 42.55 23.16 30.56 30.73 30.22 28.67 24.38 31.74 31.13 31.77 29.76
Zero-Shot / 2.10 3.27 4.08 5.68 3.78 2.02 4.73 5.57 6.08 4.60 2.39 4.73 5.59 5.95 4.67
100 15.90 19.12 14.67 13.12 15.70 10.60 11.33 10.01 9.07 10.25 11.08 12.17 10.17 10.07 10.87
200 23.97 29.56 33.98 33.95 30.37 13.25 26.76 31.26 32.58 25.96 13.53 26.54 30.53 31.20 25.45
400 44.95 42.58 40.75 33.65 40.48 26.03 35.98 34.26 29.38 31.41 24.06 35.62 35.46 31.73 31.72
ANS-SFT 800 62.10 56.55 53.01 47.55 54.80 24.70 42.27 43.05 42.65 38.17 26.09 38.98 42.95 42.55 37.64
1600 80.70 75.68 68.34 64.60 72.33 34.38 49.00 53.83 53.10 47.58 35.84 49.28 51.23 52.50 47.21
3200 82.85 80.30 78.00 71.60 78.19 36.22 52.61 55.23 54.27 49.58 38.51 52.78 54.47 53.98 49.94
final 82.70 79.93 76.70 70.22 77.39 36.00 52.82 54.59 53.55 49.24 39.63 53.75 54.33 53.60 50.33
100 6.99 14.90 15.99 20.36 14.56 10.32 13.74 11.82 14.69 12.64 6.97 13.38 12.84 13.41 11.65
200 15.45 19.12 14.53 16.46 16.39 12.90 17.51 14.22 15.82 15.11 11.23 17.71 13.87 16.32 14.78
400 25.98 26.74 19.94 16.02 22.17 15.73 21.19 17.73 15.96 17.65 16.13 20.55 17.91 15.97 17.64
COT-SFT 800 43.85 43.19 41.77 37.84 41.66 22.88 38.29 37.54 35.29 33.50 22.74 35.81 37.60 36.23 33.10
1600 52.82 61.06 54.38 45.85 53.53 28.65 43.95 40.99 40.13 38.43 29.00 41.23 40.78 39.26 37.57

25
3200 61.40 69.15 65.32 62.28 64.54 28.67 45.97 50.06 52.10 44.20 31.19 45.87 45.92 51.35 43.58
final 67.47 67.52 62.78 59.70 64.37 28.87 44.41 49.16 50.30 43.19 30.20 44.77 47.15 49.33 42.86
100 8.44 17.96 20.69 26.22 18.33 8.53 17.42 21.16 25.05 18.04 8.08 18.12 21.09 25.70 18.25
200 9.59 18.76 22.97 28.73 20.01 9.49 20.08 23.19 27.00 19.94 9.72 18.97 22.82 28.50 20.00
400 12.35 21.47 27.01 26.25 21.77 11.10 21.47 25.73 25.30 20.90 10.54 21.19 25.44 25.60 20.69
Reason-RFT-Zero 800 18.47 32.08 32.77 27.85 27.79 15.40 29.12 30.93 27.38 25.71 15.52 27.75 31.50 27.88 25.66
1600 36.78 40.20 37.78 34.51 37.32 19.96 33.03 35.84 34.49 30.83 20.39 32.85 33.87 33.90 30.25
3200 43.72 46.89 44.07 40.50 43.80 18.67 34.11 37.85 39.69 32.58 18.08 34.01 37.27 40.27 32.41
final 46.21 45.01 44.53 42.11 44.47 18.33 34.57 37.94 40.57 32.85 18.28 33.45 37.44 40.34 32.38
100 53.52 55.47 58.91 53.35 55.31 31.84 47.02 50.62 50.39 44.97 31.29 46.08 48.61 49.85 43.96
200 54.97 59.77 63.67 59.46 59.47 35.72 49.28 54.85 54.67 48.63 36.74 52.27 53.68 54.16 49.21
400 63.80 66.97 68.47 64.70 65.99 39.74 55.94 61.27 57.94 53.72 41.10 56.35 59.16 57.22 53.46
Reason-RFT 800 64.33 68.13 66.88 63.15 65.62 47.40 61.64 63.00 58.45 57.62 46.76 60.60 61.40 59.91 57.17
1600 76.47 73.42 74.05 69.16 73.28 52.68 62.22 66.56 64.71 61.54 53.47 65.17 64.98 63.52 61.79
3200 72.88 74.85 75.77 72.45 73.99 52.58 63.60 68.51 66.01 62.68 52.67 65.60 67.60 65.66 62.88
final 74.10 74.52 76.68 73.12 74.61 53.49 65.72 69.64 67.34 64.05 54.95 66.25 68.32 66.80 64.08

Table 12. Complete experimental results of Qwen2VL-2B-Instruct on the Spatial Transformation task after training on TRANCE.
 Spatial Transformation
Method Steps TRANCE (ID) TRANCE-L (OOD) TRANCE-R (OOD)
Level-1 Level-2 Level-3 Level-4 AVG Level-1 Level-2 Level-3 Level-4 AVG Level-1 Level-2 Level-3 Level-4 AVG
GPT-4o / 47.28 42.96 40.87 39.08 42.55 23.16 30.56 30.73 30.22 28.67 24.38 31.74 31.13 31.77 29.76
Zero-Shot / 16.25 16.42 10.96 10.48 13.53 11.71 16.80 11.50 10.85 12.72 13.30 16.08 10.55 11.18 12.78
100 40.30 37.05 30.67 28.35 34.09 32.07 31.12 26.00 26.80 29.00 26.38 29.71 27.74 26.48 27.58
200 65.18 53.33 49.43 45.15 53.27 33.29 45.14 45.61 45.52 42.39 35.17 43.43 45.97 43.02 41.90
400 65.33 59.35 57.17 50.77 58.16 32.40 44.13 47.69 46.23 42.61 32.10 44.88 47.04 45.25 42.32
ANS-SFT 800 78.90 70.67 63.97 62.10 68.91 34.08 50.62 51.99 52.95 47.41 34.22 50.40 50.62 52.88 47.03
1600 78.50 76.12 73.80 66.25 73.67 38.85 52.97 57.93 56.05 51.45 37.77 53.57 56.45 55.92 50.93
3200 83.80 83.23 82.83 78.17 82.01 40.10 56.02 61.02 59.90 54.26 40.78 55.06 61.67 60.98 54.62
final 83.70 84.10 82.50 78.45 82.19 39.67 55.58 61.84 60.05 54.29 42.64 54.84 61.44 60.38 54.83
100 20.58 28.98 25.97 30.00 26.38 21.89 29.94 29.18 29.33 27.59 18.49 28.43 30.43 30.36 26.93
200 41.80 44.08 46.02 42.16 43.52 25.31 36.28 36.81 38.86 34.32 21.19 34.55 37.06 37.54 32.59
400 45.39 51.32 58.20 52.42 51.83 32.53 44.61 47.73 46.10 42.74 31.87 38.96 44.62 45.90 40.34
COT-SFT 800 54.87 61.97 62.20 59.93 59.74 30.19 46.01 50.01 52.88 44.77 29.91 45.02 49.60 52.83 44.34
1600 71.27 71.14 72.82 69.93 71.29 28.82 46.43 51.01 58.94 46.30 29.08 45.25 52.24 58.31 46.22

26
3200 84.13 80.62 79.99 78.42 80.79 29.93 47.63 56.49 62.54 49.15 30.46 47.85 54.83 61.02 48.54
final 86.50 79.43 80.54 78.77 81.31 28.07 47.54 54.42 61.58 47.90 29.69 45.32 54.69 61.48 47.80
100 23.59 31.62 33.22 31.27 29.93 15.88 26.86 28.13 30.19 25.27 15.21 27.29 27.54 29.88 24.98
200 35.06 39.45 36.80 34.77 36.52 20.39 30.22 31.15 31.20 28.24 18.10 29.27 30.57 30.81 27.19
400 25.28 40.78 41.70 35.35 35.78 20.20 39.28 35.43 33.44 32.09 21.72 39.47 37.63 32.79 32.90
Reason-RFT-Zero 800 50.18 51.55 50.43 46.06 49.56 35.44 46.15 45.65 39.55 41.70 33.90 45.89 46.48 41.57 41.96
1600 59.60 61.90 57.30 55.36 58.54 43.95 55.03 52.96 50.60 50.64 41.28 56.46 51.64 49.08 49.62
3200 62.50 68.53 68.79 66.22 66.51 42.54 58.05 58.97 60.10 54.92 42.56 56.93 60.02 60.21 54.93
final 65.63 68.30 69.45 67.30 67.67 46.61 58.22 61.69 62.26 57.20 45.53 58.40 61.81 58.85 56.15
100 60.07 62.95 68.53 62.78 63.58 37.09 52.54 60.21 54.99 51.21 35.26 52.56 57.30 55.42 50.14
200 67.69 66.98 69.53 66.00 67.55 36.26 53.94 60.20 60.05 52.61 34.33 54.21 58.25 60.79 51.90
400 74.72 71.31 73.62 69.14 72.20 36.11 56.00 62.25 63.01 54.34 36.95 55.72 60.82 64.88 54.59
Reason-RFT 800 72.57 71.04 75.58 73.26 73.11 38.93 56.86 64.10 63.69 55.90 35.06 54.39 60.82 64.88 53.79
1600 79.18 76.70 76.59 73.26 76.43 37.25 53.44 66.38 66.76 55.96 35.01 55.39 66.02 66.67 55.77
3200 79.70 79.22 80.13 76.61 78.92 37.63 55.73 65.24 68.02 56.66 37.18 56.34 63.90 68.89 56.58
final 80.48 79.15 81.97 78.26 79.97 43.68 61.04 64.66 68.05 59.36 45.45 56.34 63.33 69.33 58.61

Table 13. Complete experimental results of Qwen2VL-7B-Instruct on the Spatial Transformation task after training on TRANCE.
 Question: Subtract all spheres. How many objects are left?

COT-SFT: Reason-RFT:
<think>Let‘s start by identifying all the objects in the <think>Let's start by identifying all the objects in the image.
image. We can see various shapes: spheres, cubes, and I see a total of 9 objects. The question asks us to subtract all
cylinders. First, let’s count the total number of objects the spheres.
in the image:1. Red sphere 2. Green sphere 3. Another There are 3 spheres in the image: a red sphere, a green
green sphere 4. Purple cube 5. Brown cube 6. Yellow sphere, and a green sphere. So, we need to subtract these 3
cube 7. Brown cylinder 8. Blue transparent cube. spheres from the total of 9 objects. 9 (total objects) - 3
(Miss one object) Now, let's identify and count the (spheres) = 6. Therefore, after subtracting the spheres, there
spheres: - Red sphere - Green sphere - Another green are 6 objects left.</think>
sphere. There are 3 spheres in total. Next, we subtract
the number of spheres from the total number of objects: <answer>6</answer>
Total objects = 8, Spheres = 3. Objects left after
subtracting spheres = Total objects - Spheres = 8 - 3 =
5. Oh, I see, after removing all the spheres, we are left
with 5 objects.</think>

<answer>5</answer>

Question: Add 7 small cars. How many cars are left?

COT-SFT: Reason-RFT:
<think>Let's start by identifying the number of cars in <think>Let's start by identifying the cars in the image. I see
the image. I see a blue SUV, a red airplane, a purple two cars: a blue SUV and a blue minivan. The question asks
dirt bike, and a blue small car. That makes a total of 4 to add 7 small cars to the existing cars. So, we begin with 2
cars initially. Now, according to the question, we need cars and add 7 more.
to add 7 small cars to the existing ones. So, let's 2 (existing cars) + 7 (additional cars) = 9 cars in total.
calculate the total number of cars after adding these 7 Therefore, after adding 7 small cars, there are 9 cars in
small cars. Initial number of cars = 4. Number of small total.</think>
cars to add = 7. Total number of cars = Initial number
of cars + Number of small cars to add. Total number of <answer>9</answer>
cars = 4 + 7 = 11. Therefore, after adding 7 small cars,
there are 11 cars in total.</think>

<answer>11</answer>

Figure 14. Visualization examples of Reason-RFT and CoT-SFT on Visual Counting task.

27
 Question: How many times larger is the total
number of eggs in the second group compared
to the first group?

Reason-RFT:
<think>First, let's count the number of egg trays in the
COT-SFT:
first group. There are 2 trays in the first group. Each
<think>First, let's count the number of egg trays tray contains 10 eggs, so the total number of eggs in
in the first group. There are 3 trays in the first the first group is 2 trays * 10 eggs/tray = 20 eggs.
group. Each tray holds 12 eggs, so the total Next, let's count the number of egg trays in the second
number of eggs in the first group is 3 trays * 12 group. There are 10 trays in the second group. Each
eggs/tray = 36 eggs. Next, let's count the number tray contains 10 eggs, so the total number of eggs in
of egg trays in the second group. There are 10 the second group is 10 trays * 10 eggs/tray = 100 eggs.
trays in the second group. Each tray holds 12 eggs, Now, let's compare the total number of eggs in the
so the total number of eggs in the second group is second group to the first group. The second group has
10 trays * 12 eggs/tray = 120 eggs. Now, let's find 100 eggs, and the first group has 20 eggs. To find how
the ratio of the total number of eggs in the second many times larger the second group is compared to
group to the first group: 120 eggs / 36 eggs = 10. the first group, we divide the total number of eggs in
Therefore, the second group has 10 times more the second group by the total number of eggs in the
eggs than the first group.</think> first group: 100 eggs / 20 eggs = 5.</think>

<answer>10</answer> <answer>5</answer>

Question: How many more purple circles are there

than green squares?

Reason-RFT:
<think>First, let‘s count the number of purple circles.
COT-SFT:
I see 1, 2, 3, 4, 5 purple circles. Next, let's count the
<think>First, let's count the number of green number of green squares. I see 1, 2, 3, 4, 5 green
squares. There are 4 green squares in the image. squares. Now, to find out how many more purple
Next, let's count the number of purple circles. circles there are than green squares, we subtract the
There are 5 purple circles in the image. To find the number of green squares from the number of purple
difference, we subtract the number of green circles: 5 (purple circles) - 5 (green squares) = 0. Oh,
squares from the number of purple circles: 5 I see, there are no more purple circles than green
(purple circles) - 4 (green squares) = 1.</think> squares.</think>

<answer>1</answer> <answer>0</answer>

Figure 15. Visualization examples of Reason-RFT and CoT-SFT on Structure Perception task.

28
 Question: Based on the image, determine
which number is directly opposite to number 4
on the die depicted. Use reasoning to arrive at
the correct answer from the choices provided.
Choices: A. 1 B. 2 C. 3 D. 6

COT-SFT: Reason-RFT:
<think>Let me think about the properties of a <think>Let's analyze the image. The numbers
standard six-sided die. On a six-sided die, each shown are 2, 3, and 4. To find the number opposite
number from 1 to 6 is opposite to another to 4, we need to consider the properties of a
number. The sum of the numbers on opposite standard six-sided die. On a die, opposite numbers
sides is always 7. Number 4 is opposite to add up to 7. Since 4 is opposite to 3, we can
number 2 because 4 + 2 = 6, and the sum of 6 conclude that the number opposite to 4 is 3.</think>
and 1 is 7. Therefore, the number directly
opposite to number 4 is 2.</think> <answer>C</answer>

<answer>B</answer>

Question: What is the total surface area of

this rectangular prism if its dimensions are 3
units by 4 units by 5 units?

COT-SFT: Reason-RFT:
<think>To find the total surface area of a <think>To find the total surface area of a rectangular
rectangular prism, we need to calculate the sum of prism, we need to calculate the sum of the areas of
the areas of all six faces. The formula for the all six faces. The formula for the total surface area is
surface area is 2*(length*width + width*height + 2*(length*width + width*height + height*length).
height*length). Let‘s calculate it step-by-step. 1. Let's calculate it step by step. First, calculate the area
Calculate the area of the face with dimensions 3 of the face with dimensions 3 units and 4 units: 3 * 4
units by 4 units: 3 * 4 = 12 square units. 2. = 12 square units. Next, calculate the area of the face
Calculate the area of the face with dimensions 4 with dimensions 4 units and 5 units: 4 * 5 = 20
units by 5 units: 4 * 5 = 20 square units. 3. square units. Then, calculate the area of the face with
Calculate the area of the face with dimensions 5 dimensions 5 units and 3 units: 5 * 3 = 15 square
units by 3 units: 5 * 3 = 15 square units. Now, add units. Now, add these areas together: 12 + 20 + 15 =
these areas together: 12 + 20 + 15 = 47 square units. 47 square units.

Therefore, the total surface area of the rectangular Finally, multiply by 2 to account for both pairs of
prism is 47 square units.</think> each face: 2 * 47 = 94 square units.</think>

<answer>47</answer> <answer>94</answer>

Figure 16. Visualization examples of Reason-RFT and CoT-SFT on Structure Perception task.

29
 Question: What the transformation has been applied to
the letter on the left to obtain the letter on the right, and
which option correctly describes this transformation?
Choices: (A) slide (B) flip (C) turn

COT-SFT: Reason-RFT:
<think>Let me think about the transformation applied <think>Let's analyze the image. The letter on the left is
to the letter on the left. The letter on the left is a grid the original image, and the letter on the right is the
of squares, and the letter on the right is also a grid of transformed image. The transformation involves flipping
squares. The orientation and shape of the letter remain the letter horizontally. This means the letter has been
the same, but the position of the letter has changed. mirrored along a vertical axis. Therefore, the correct
Oh, I see, this is a slide transformation because the option describing this transformation is a flip.</think>
letter has been moved horizontally. </think>
<answer>B</answer>
<answer>A</answer>

Question: Josiah kept track of how many kilometers he

walked during the past 7 days. Can you determine
which number of kilometers he walked most frequently?

COT-SFT: Reason-RFT:
<think>First, let's list the kilometers Josiah walked <think>First, let's list the kilometers Josiah walked
each day: 4, 3, 7, 4, 7, 3, 7. Now, let's count how each day: 4, 3, 7, 4, 7, 3, 7. Now, let's count how many
many times each number appears: 4 appears 3 times, times each distance appears: - 4 km appears 2 times. -
3 appears 2 times, and 7 appears 2 times. Oh, I see, 3 km appears 2 times. - 7 km appears 3 times. Oh, I
the number that appears the most frequently is the see, the highest frequency is 3, which corresponds to 7
mode. Since 4 appears 3 times, which is more than km.</think>
any other number, the mode is 4.</think>
<answer>7</answer>
<answer>4</answer>

Figure 17. Visualization examples of Reason-RFT and CoT-SFT on Structure Perception task.

30
 Question: How many main structural elements can be
identified in the suspension bridge, and what are their roles
in supporting the bridge?

Reason-RFT:
<think>Let me think about the main structural elements
visible in the image. First, I notice the two large towers
supporting the bridge. These towers are crucial for
providing the necessary height and strength to maintain the
bridge's structure. Next, I see the main span of the bridge,
COT-SFT:
which is the horizontal part connecting the towers. This
<think>Let me think about the main structural elements span is supported by the towers and is essential for carrying
visible in the image. First, I see the two large white the weight of the bridge and any traffic it carries. Finally,
towers, which are the main supports for the suspension there are additional support structures visible, such as the
bridge. These towers are tall and robust, providing the curved steel cables, which help distribute the weight evenly
necessary height and strength to maintain the tension in and maintain the bridge's stability. So, the main structural
the cables. So, there are two main structural elements: elements are the towers, the main span, and the curved steel
the towers. Oh, I see, the towers are crucial for cables. Counting these elements, I see two towers, one
supporting the weight of the bridge and ensuring its main span, and multiple curved steel cables. Therefore, the
stability. Without them, the bridge would not be able to main structural elements are three: the towers, the main
span the valley effectively.</think> span, and the curved steel cables.</think>

<answer>2</answer> <answer>3</answer>

Question: How many distinct colors of lights are used

in the Christmas tree decoration?

COT-SFT:
<think>Let me think about the colors described in the
image. The Christmas tree is decorated with lights that
are described as being red, blue, and white. I need to Reason-RFT:
count these distinct colors. First, I identify the color red, <think>Let me think. The image shows a Christmas tree
which is mentioned directly. Next, I see the color blue, decorated with lights. The description mentions that the
which is also mentioned. Finally, I observe the color lights are of different colors. I need to identify these colors.
white, which is described as shining brightly. So, there The description specifically mentions red and white lights.
are three distinct colors of lights used in the decoration: So, there are two distinct colors of lights used in the
red, blue, and white.</think> decoration.</think>

<answer>3</answer> <answer>2</answer>

Figure 18. Visualization examples of Reason-RFT and CoT-SFT on Structure Perception task.

31
 Question: Given the hyperbola shown in the image, can
we determine the equation of the hyperbola by analyzing
the relationship between the slope of the asymptote and
the distance from the center to the vertex?
Choices: A. Yes B. No

COT-SFT: Reason-RFT:
<think>Let's think about the properties of a <think>Let's analyze the image. The hyperbola is shown
hyperbola. The slope of the asymptote of a hyperbola with two asymptotes intersecting at the center. To find the
is given by the equation \\( m = \\frac{a}{b} \\), equation of the hyperbola, we need to consider the
where \\( a \\) and \\( b \\) are the coefficients from relationship between the slope of the asymptote and the
the equation of the hyperbola \\( \\frac{x^2}{a^2} - distance from the center to the vertex. The equation of a
\\frac{y^2}{b^2} = 1 \\). The distance from the hyperbola with its center at the origin and aligned with the
center to the vertex is \\( c \\), which is related to the coordinate axes is given by \\( \\frac{x^2}{a^2} -
semi-major and semi-minor axes. However, to find \\frac{y^2}{b^2} = 1 \\), where \\( a \\) and \\( b \\) are the
the equation of the hyperbola, we need more distances from the center to the vertices and asymptotes,
information than just the slope of the asymptote and respectively. The center to the vertices is given, which
the distance from the center to the vertex. We also suggests that \\( a^2 + b^2 \\) is given. The slope of the
need the values of \\( a \\) and \\( b \\), which are not asymptotes is given by \\( \\pm \\frac{b}{a} \\). In this
directly provided in the image. Therefore, the case, the slope is 3, which implies that \\( \\frac{b}{a} = 3
information given is insufficient to determine the \\). Therefore, the equation of the hyperbola can be
equation of the hyperbola.</think> determined by these relationships.</think>

<answer>B</answer> <answer>A</answer>

Question: What is the circumference of the circle?

COT-SFT: Reason-RFT:
<think>To find the circumference of thecircle, I need to <think>First, let‘s determine the radius of the circle.
know the diameter of thecircle. The diameter is the The problem states that the radius is 7. The formula
longest straight line that can be drawn across a circle, for the circumference of a circle is C = 2 π r, where r
passing through the center. Once I have the diameter, I is the radius. Substituting the given radius, r = 7, into
can use the formula for the circumference of a circle, the formula, we have C = 2 π (7). Calculating this
which is C = π d, where d is the diameter. Oh, I see, the gives C = 14 π. Therefore, the circumference of the
diameter of the circle is given as 7. So, I can calculate the circle is 14 π.</think>
circumference as C = π * 7. Using the value of π as
approximately 3.14159, I can calculate the circumference <answer>14 π </answer>
as 7 * 3.14159 = 21.99633. Rounding this to one decimal
place, the circumference is approximately 22.</think>

<answer>22</answer>

Figure 19. Visualization examples of Reason-RFT and CoT-SFT on Structure Perception task.

32
 Question: What is the length of side QP in the rhombus?

COT-SFT: Reason-RFT:
<think>Let‘s think about this step-by-step. The <think>Let‘s analyze the rhombus LMPQ. We know that
rhombus LMPQ is a type of quadrilateral where all all sides of a rhombus are equal in length. The side QP is
sides are equal in length. We are given that side LM is one of these equal sides. Since the problem states that
8 units long. Since all sides of a rhombus are equal, the side MP is 10, we can confirm that this is indeed the
the length of side QP must also be equal to the length length of side QP in the rhombus LMPQ. </think>
of side LM. Therefore, QP = LM = 8 units.</think>
<answer>10</answer>
<answer>8</answer>

Question: Consider the right triangle ACB, where

angle C is 90 degrees, angle A is 37 degrees, and the
length of AC is 4 units. What is the approximate
length of BC, expressed as 'k' in terms of the given
angles and length? (sin37.0 approximate 0.6,
cos37.0 approximate 0.8, tan37.0 approximate 0.75)
Choices: A: 2.4 B: 3.0 C: 3.2 D: 5.0
COT-SFT:
<think>Let‘s think through this step-by-step. We know that angle A is 37 degrees, and we have the tangent function,
which is the ratio of the opposite side (AB) to the adjacent side (AC). Given that tan(37°) = 0.75, we can write the
equation as AB = AC / tan(37°). Substituting the given value, AB = 4 / 0.75. Calculating this gives AB = 5.33. Now,
we need to find the hypotenuse BC using the Pythagorean theorem, which states that in a right triangle, the square
of the hypotenuse is equal to the sum of the squares of the other two sides. So, BC = \sqrt(AB^2+ AC^2).
Substituting the known values, BC = \sqrt(5.33^2+ 4^2). Calculating this gives BC = 7.34. Therefore, the
approximate length of BC is 7.34, which is not an exact match with the options. However, the closest approximation
from the choices given is 3.0, which is not correct based on our calculations. Let's re-evaluate the choices and the
reasoning provided.</think>

<answer>Based on the reasoning and calculations, the closest approximation from the choices given is 3.0, which is
not correct based on our calculations</answer>

Reason-RFT:

<think>Let's start by finding the length of BC using the tangent of angle A. We know that tan(angle A) = opposite /
adjacent = BC/AC. Given that angle A is 37 degrees, we can write: tan(37°) = BC/4. We also know that tan(37°) is
approximately 0.75. So, BC = 4 * 0.75 = 3. Therefore, the length of BC is approximately 3.0.</think>

<answer>B</answer>

Figure 20. Visualization examples of Reason-RFT and CoT-SFT on Structure Perception task.

33
 Question: The first image shows the central view of initial state,
the second image shows the left view of final state, and the
attributes of the initial objects are listed below:
(0, cube, medium, gray, rubber) (1, sphere, medium, cyan, metal)
(2, cube, small, gray, metal) (3, cube, small, brown, metal)
(4, cylinder, large, yellow, metal) (5, sphere, small, yellow, rubber)
What is the transformation process that can achieve the change
from initial state to final state?
Initial central-view image Final left-view image

COT-SFT:
<think>Let's analyze the initial and final states of the objects in the images. We have the attributes of the initial objects, and we need
to determine the transformations that lead to the final state.
1. **Object 0**: Initially, it‘s a medium gray rubber cube. In the final state, it appears to be a sphere. So, we need to change shape.
- Transformation: `change_shape(0, sphere)`
2. **Object 1**: Initially, it's a medium cyan metal sphere. In the final state, it seems to have moved to the right and increased in size.
- Transformation: `change_position(1, right)`, `change_size(1, large)`
3. **Object 2**: Initially, it's a small gray metal cube. In the final state, it appears unchanged.
4. **Object 3**: Initially, it's a small brown metal cube. In the final state, it appears unchanged.
5. **Object 4**: Initially, it's a large yellow metal cylinder. In the final state, it appears unchanged.
6. **Object 5**: Initially, it's a small yellow rubber sphere. In the final state, it appears unchanged.

Now, let's compile the transformations needed to achieve the final state from the initial state:
- `change_shape(0, sphere)`
- `change_position(1, right)`
- `change_size(1, large)`
These transformations match the original answer provided.</think>

<answer>change_shape(0, sphere), change_position(1, right), change_size(1, large)</answer>

Reason-RFT:
<think>Let's analyze the initial and final states to determine the transformations needed.
1. **Object 0**: Initially a medium gray rubber cube. In the final state, it appears to be a medium gray rubber sphere.
- Transformation: `change_shape(0, sphere)`
2. **Object 1**: Initially a medium cyan metal sphere. In the final state, it appears to be a large cyan metal sphere.
- Transformation: `change_size(1, large)`
3. **Object 2**: Initially a small gray metal cube. In the final state, it appears to be a small gray metal cube.
- Transformation: None needed as it remains unchanged.
4. **Object 3**: Initially a small brown metal cube. In the final state, it appears to be a large yellow metal cube.
- Transformation: `change_color(3, yellow)` and `change_size(3, large)`
5. **Object 4**: Initially a large yellow metal cylinder. In the final state, it appears to be a large yellow metal cylinder.
- Transformation: None needed as it remains unchanged.
6. **Object 5**: Initially a small yellow rubber sphere. In the final state, it appears to be a small yellow rubber sphere.
- Transformation: None needed as it remains unchanged.

By analyzing the changes, we can determine the sequence of transformations needed to achieve the final state from the initial state.
Therefore, the transformations are:
- `change_color(3, yellow)`
- `change_size(1, large)`
- `change_size(3, large)`
- `change_shape(0, sphere)`
These transformations match the original answer provided. </think>

<answer>change_color(3, yellow), change_size(1, large), change_size(3, large), change_shape(0, sphere)</answer>

Figure 21. Visualization examples of Reason-RFT and CoT-SFT on Spatial Transformation task.

34
 Question: The first image shows the central view of initial state,
the second image shows the right view of final state, and the
attributes of the initial objects are listed below:
(0, sphere, small, red, metal) (1, sphere, large, brown, metal)
(2, cube, large, blue, metal) (3, sphere, small, brown, metal)
(4, cylinder, small, blue, glass)
What is the transformation process that can achieve the change
from initial state to final state?
Initial central-view image Final right-view image

COT-SFT:
<think>Let's analyze the initial and final states of the objects in the images. We have the initial attributes of the objects, and we need
to determine the transformations that occurred.
1. **Object 0**: Initially a small red metal sphere. In the final image, it appears to have moved to the front left. So, we apply
`change_position(0, front_left)`.
2. **Object 1**: Initially a large brown metal sphere. In the final image, it seems to have moved behind the other objects. So, we
apply `change_position(1, behind)`.
3. **Object 2**: Initially a large blue metal cube. In the final image, it appears to have changed shape to a sphere. So, we apply
`change_shape(2, sphere)`.
4. **Object 3**: Initially a small brown metal sphere. In the final image, it appears to have changed color to yellow. So, we apply
`change_color(3, yellow)`.
5. **Object 4**: Initially a small blue glass cylinder. There doesn't seem to be any change in its attributes or position in final image.

Now, let's compile the transformations in the order they were identified:
- `change_color(3, yellow)`
- `change_position(0, front_left)`
- `change_shape(2, sphere)`
- `change_position(1, behind)`
These transformations match the original answer provided.</think>

<answer>change_color(3, yellow), change_position(0, front_left), change_shape(2, sphere), change_position(1, behind)</answer>

Reason-RFT:
<think>Let's analyze the initial and final states to determine the transformations needed.
1. **Object 3 (sphere, small, brown, metal):**
- In the initial state, it is a small brown metal sphere.
- In the final state, it appears to be a small yellow cylinder.
- Therefore, the transformations needed are: - Change its color from brown to yellow: `change_color(3, yellow)` - Change its
shape from a sphere to a cylinder: `change_shape(3, cylinder)`
2. **Object 1 (sphere, large, brown, metal):**
- In the initial state, it is a large brown metal sphere.
- In the final state, it has moved to the right.
- Therefore, the transformation needed is: - Change its position to the right: `change_position(1, right)`
3. **Object 2 (cube, large, blue, metal):**
- In the initial state, it is a large blue metal cube.
- In the final state, it appears to be a large blue metal sphere.
- Therefore, the transformation needed is: - Change its shape from a cube to a sphere: `change_shape(2, sphere)`

By applying these transformations, we can achieve the change from the initial state to the final state. The transformations are:
- `change_color(3, yellow)`
- `change_shape(3, cylinder)`
- `change_position(1, right)`
- `change_shape(2, sphere)`
These transformations match the original answer provided.</think>

<answer>change_color(3, yellow), change_shape(3, cylinder), change_position(1, right), change_shape(2, sphere)</answer>

Figure 22. Visualization examples of Reason-RFT and CoT-SFT on Spatial Transformation task.

35