Evaluating Multilingual Retrieval-

Augmented Generation Systems: A

Comprehensive Framework for
Argument-Level Verification and Citation
Attribution in Hebrew
The Paradigm of Structured Retrieval-Augmented
Generation Evaluation
The rapid proliferation of large language models (LLMs) has revolutionized information-seeking
applications, yet their tendency to generate fluent but factually unsupported content
necessitates rigorous grounding mechanisms. Retrieval-Augmented Generation (RAG)
architectures address this by conditioning the generative process on externally retrieved,
verifiable corpora. However, as RAG systems evolve from generating monolithic, continuous
prose to outputting highly structured, discrete information—such as specific arguments mapped
to precise citations—the evaluation methodologies must similarly evolve.
Evaluating a specialized RAG system that ingests queries in Hebrew, retrieves relevant Hebrew
documents, and outputs a discrete list of arguments alongside specific citations for each
argument presents a multidimensional assessment challenge. The ground truth for this system
is equally structured, comprising a verified list of arguments and their corresponding citations.
Traditional natural language processing metrics, which rely on lexical overlap and n-gram
matching, are wholly inadequate for this architecture. These legacy metrics fail to capture
semantic equivalence, ignore the functional utility of citations, and degrade significantly when
applied to morphologically rich and highly inflected languages like Hebrew.
A rigorous, literature-backed evaluation framework for this specific architecture must decouple
the generative text assessment (the arguments) from the attribution assessment (the citations),
while simultaneously addressing the cross-lingual and linguistic complexities of employing LLMs
as automated evaluators. This report synthesizes leading research in atomic fact verification,
LLM-as-a-Judge topologies, fine-grained citation attribution, and Hebrew-native language
modeling to propose an exhaustive, domain-specific evaluation pipeline.

The Evolution of RAG Evaluation and the Shift to

Atomic Factuality
The evaluation of generative models traditionally relied on metrics such as BLEU, ROUGE, or
METEOR, which quantify the surface-level string overlap between a generated output and a
reference text. The literature overwhelmingly indicates that these reference-based metrics fail to
capture the semantic nuance, logical coherence, and factual accuracy required in complex
information-seeking tasks. In the context of RAG systems, a model might generate a response
that shares very few lexical tokens with a reference answer but is nonetheless completely
factually correct and perfectly grounded in the retrieved context.
Consequently, the academic focus has shifted toward dedicated RAG evaluation strategies that
isolate and inspect each subsystem independently. Frameworks such as RAGAS (Retrieval-
Augmented Generation Assessment) formalize this disaggregated approach by separating the
evaluation of the retrieval apparatus from the generation apparatus. RAGAS employs
specialized LLM critics to compute continuous scores for distinct dimensions of the pipeline.
RAGAS Evaluation Focus Component Mathematical or Primary Objective
Dimension Conceptual Definition
Context Precision Retriever Evaluates the ranking Ensures that the most
quality of retrieved highly relevant
contexts, often using documents appear at
metrics like Normalized the absolute top of the
Discounted Cumulative retrieved context
Gain (NDCG@k). window.
Context Recall Retriever The proportion of Ensures the generative
relevant documents or model receives all
ground-truth concepts necessary information
that are successfully to construct a complete
captured within the answer.
retrieved context
window.
Answer Relevancy Generator Measures the semantic Ensures the system
distance between the directly addresses the
generated response user's specific query
and the original user without digression.
prompt, penalizing
tangential or incomplete
answers.
Faithfulness Generator The ratio of claims in Quantifies the model's
the generated response resistance to
that can be logically hallucination by
inferred from the ensuring absolute
retrieved context to the grounding in the
total number of claims provided evidence.
generated.
While the RAGAS framework provides a robust foundation, its standard application is designed
for continuous, long-form text. The specific architecture under review operates differently: it
natively outputs a discrete list of arguments. This structural characteristic requires a paradigm
shift toward atomic factuality.

FActScore and the Atomic Fact Paradigm

Evaluating long-form text is inherently complex because a single sentence can contain a
mixture of supported facts, unsupported claims, and entirely hallucinated information, rendering
binary judgments of quality inadequate. To resolve this, researchers introduced the FActScore
(Factual precision in Atomicity Score) methodology, which evaluates factuality by decomposing
generation into the smallest possible units of information.
The FActScore methodology operates on three fundamental assumptions regarding atomic
facts. First, whether an atomic fact is supported by a knowledge source must be undebatable
and objective. Second, every atomic fact carries an equal weight of importance within the
evaluation. Third, the pieces of information residing in the knowledge base do not conflict or
overlap with each other in a manner that precludes verification.
The standard FActScore pipeline consists of two distinct stages: Atomic Fact Generation (AFG),
which utilizes an LLM to parse a complex sentence into multiple simple claims, and Atomic Fact
Validation (AFV), which verifies each extracted claim against a trusted knowledge source. The
primary advantage of the evaluated Hebrew RAG system is that it structurally bypasses the
highly error-prone AFG step. Because the system's output and the ground truth are already
formatted as discrete lists of arguments, the data is natively atomized. The evaluation
framework can therefore proceed directly to a specialized form of Atomic Fact Validation,
mapping the generated arguments against the ground truth arguments.

LLM-as-a-Judge for N-to-M Argument Matching

The task of validating the generated list of arguments against the ground truth list of arguments
cannot be resolved through deterministic algorithmic matching. A generated argument in
Hebrew might utilize entirely different vocabulary or syntactic structures than the ground truth
argument while conveying the exact same semantic meaning. The literature dictates that
evaluating semantic equivalence requires the implementation of the LLM-as-a-Judge paradigm,
utilizing a powerful instruction-tuned model to evaluate the outputs based on natural language
rubrics.

The Bipartite Matching Problem

When the RAG model generates N arguments and the ground truth (GT) contains M arguments,
the LLM judge must execute an N-to-M mapping to determine both the comprehensiveness and
the accuracy of the response. This process fundamentally measures recall and precision at the
semantic argument level.
To evaluate Argument Recall (Coverage), the framework must determine whether the
semantic core of every expected argument was successfully generated. The LLM judge iterates
through the M ground truth arguments. For each ground truth argument, the judge scans the
entire list of N generated arguments to identify if the necessary information is present. The
evaluation relies on a defined support level, typically categorizing the match as full entailment,
partial entailment, or contradiction.
To evaluate Argument Precision (Correctness), the framework must ensure that the model
did not generate extraneous, hallucinated, or factually incorrect arguments. The LLM judge
iterates through the N generated arguments. For each generated argument, the judge compares
it against the complete set of M ground truth arguments. If a generated argument contradicts the
ground truth or introduces domain-specific claims absent from the reference material, the
precision score is severely penalized.
Matching Vector Evaluation Target LLM Judge Instruction Mathematical
Paradigm Representation
GT \rightarrow Argument Recall "Does the generated \frac{\sum \text{Entailed
Generated list of arguments GT Arguments}}{M}
contain the semantic
equivalent of this
specific ground truth
argument?"
Generated \rightarrow Argument Precision "Is this specific \frac{\sum \
GT generated argument text{Supported
factually supported and Generated Arguments}}
verified by the provided {N}
ground truth list?"
Matching Vector Evaluation Target LLM Judge Instruction Mathematical
Paradigm Representation
Generated \rightarrow Contextual Faithfulness "Can this specific \frac{\sum \
Context generated argument be text{Contextually
logically inferred solely Grounded Arguments}}
from the retrieved {N}
context, regardless of
the GT?"

Architecting the Evaluation Prompts

The reliability of the LLM-as-a-Judge methodology is heavily dependent on the architecture of
the evaluation prompt. The literature distinguishes between reference-less and reference-based
evaluation. For this system, the evaluation is strictly reference-based, utilizing the ground truth
arguments as the definitive anchor.
Research into frameworks like G-Eval and Directed Acyclic Graph (DAG) metrics demonstrates
that providing a single, monolithic prompt asking for a score from 1 to 5 yields highly
inconsistent results. Instead, the prompt must enforce strict, deterministic logic. Furthermore,
the CLEAR (Error Analysis via LLM-as-a-Judge Made Easy) framework highlights that forcing
the LLM to generate instance-level textual feedback—a Chain-of-Thought (CoT) explanation—
before outputting a final verdict significantly stabilizes the evaluation and provides actionable
diagnostic data. The prompt must explicitly define the criteria for "entailment," "partial overlap,"
and "contradiction," ensuring the judge operates within a tightly bounded cognitive framework.

Mitigating Judge Bias and Inconsistency

Despite their utility, LLM judges are susceptible to inherent biases, including verbosity bias
(favoring longer answers), positional bias (favoring answers presented first), and calibration
drift. The literature identifies several critical mechanisms for mitigating these vulnerabilities and
ensuring rigorous evaluation.
The most effective mitigation strategy is Ensemble Judging. Research from the FACTS (Factual
Accuracy Check) framework by DeepMind demonstrates that utilizing a multi-judge ensemble—
for example, deploying GPT-4o, Claude 3.5 Sonnet, and a specialized local model
simultaneously—and aggregating their verdicts via majority vote dramatically reduces single-
vendor bias and aligns much closer to human expert annotations.
Furthermore, addressing imperfect sensitivity and specificity requires calibration. Evaluators
construct a small, manually annotated calibration dataset of 100 to 200 examples. The LLM
judge is run against this dataset, and its verdicts are compared to the human ground truth using
statistical metrics such as Cohen's Kappa for categorical agreement and Pearson correlation for
numerical alignment. If the LLM judge exhibits systemic leniency or strictness, the evaluation
prompt is adaptively refined until the judge's baseline aligns with human consensus.

Advanced Metrics for Citation Attribution

In the evaluated system, the model outputs not only a list of arguments but also specific
citations appended to each argument. In high-stakes, information-critical domains, the
attribution of a claim to a specific, verifiable source is as important as the claim itself. The
literature reveals that evaluating citation quality requires a distinct set of mathematical metrics
and LLM judge paradigms.
The ALCE Benchmark and Fundamental Citation Mathematics
The ALCE (Automatic LLMs' Citation Evaluation) framework serves as the standard for
evaluating retrieval-augmented text generation. ALCE defines attribution across multiple
dimensions, moving beyond simple factual correctness to explicitly measure the utility and
accuracy of the citations themselves. ALCE utilizes Natural Language Inference (NLI) models to
compute two critical mathematical metrics: Citation Recall and Citation Precision.
Citation Recall (R_2) measures the proportion of the generated text that is fully supported by
the specific documents cited inline. In the context of the argument-list system, this metric
evaluates whether the specific citation attached to an argument actually contains the evidence
necessary to prove that argument. The NLI model treats the cited passage as the premise and
the generated argument as the hypothesis. A score of +w_2 is assigned if the argument is
entailed by the premise, and -w_2 otherwise.
Citation Precision (R_3) addresses the problem of citation hallucination and redundancy. It
measures the percentage of citations generated that are genuinely relevant and necessary to
support the argument. If an LLM generates a valid argument but appends three citations to it,
and only one of those citations actually contains the relevant information, the Citation Precision
score is penalized. This ensures the model is not merely appending random retrieved
documents to create an illusion of grounding.

ALiiCE and Positional Fine-Grained Evaluation

Traditional citation evaluation frameworks often operate at the sentence or paragraph level,
which obscures nuances when a single complex sentence contains multiple atomic claims
requiring different citations. The ALiiCE (Automatic LLM’s Positional Fine-grained Citation
Evaluation) framework was developed to address this limitation.
ALiiCE utilizes a dependency tree-based approach to parse complex, sentence-level claims into
atomic sub-claims. Once parsed, it evaluates the citation quality at the atomic level and
introduces the Coefficient of Variation of Citation Positions (CVCP) to measure the dispersion
and precise placement of citation markers within a text.
The system under review possesses a distinct architectural advantage: it bypasses the need for
ALiiCE's complex dependency tree parsing because its output is natively structured as a list of
discrete arguments mapped to specific citations. The evaluation can directly apply the core
entailment formulation of fine-grained citation evaluation without the overhead of syntactic
parsing. The mathematical judgment of entailment is formulated as \Psi(\mathcal{H}, \
mathcal{S}) = 1 if the Hypothesis (the text of the specific cited document) entails the Statement
(the generated argument), and 0 otherwise.
Citation Metric Evaluation Focus NLI Formulation Interpretation
Citation Recall Claim Support Does Cited Doc C_i Measures if the model
entail Argument A_i? correctly identified the
source of its generated
knowledge.
Citation Precision Citation Utility Is Cited Doc C_i Penalizes the model for
necessary to entail over-citing irrelevant or
Argument A_i? tangential documents.
Citation Coverage System What percentage of Ensures the model is
Comprehensiveness total generated fully attributing its
arguments possess at outputs, leaving no
least one valid citation? "orphan" arguments.
TRACE: Differentiating Correctness from Causal Faithfulness
Recent advancements in attribution evaluation have introduced the TRACE (Trustworthy
Retrieval-Aligned Citation Evaluation) framework, which identifies a critical vulnerability in
standard citation metrics. TRACE delineates a strict boundary between Citation Correctness
and Citation Faithfulness.
Citation Correctness is a post-hoc semantic assessment; it asks whether the cited document
entails the argument. However, LLMs frequently engage in a phenomenon known as "post-
rationalization." In this scenario, the model retrieves a document, completely ignores it,
generates a factual argument from its own internal parametric memory, and then appends a
citation marker to the output. If the cited document happens to contain the fact, traditional
metrics will score the citation as "correct."
Citation Faithfulness, however, asks whether the cited source genuinely contributed to the
generation of the content in a causal manner. TRACE mandates that evaluation frameworks
audit the causal dependency of the argument on the retrieved context. To evaluate causal
faithfulness, the framework constructs citation-support matrices and compares the generated
citations against an ablation baseline, such as prompting the model without the retrieved
context, to determine if the attribution represents genuine synthesis or merely parametric
regurgitation.

Defining Granular Citation Support Levels

Evaluating the relationship between a cited document and a generated argument is rarely a
binary proposition. The literature emphasizes the necessity of utilizing multi-category support
levels to accurately capture the nuance of citation quality. When the LLM judge evaluates a
citation, it must classify the relationship into one of three distinct categories:
First, Full Support (FS) occurs when the cited text fully corroborates every detail, entity, and
relational claim present within the generated argument. Second, Partial Support (PS) occurs
when the citation corroborates the core premise of the argument, but the argument contains
supplementary details, temporal constraints, or specific entities that are absent from the cited
text. Third, No Support (NS) occurs when the cited text is completely irrelevant to the
argument, or explicitly contradicts the claims made in the argument. By implementing this
granular classification system, the evaluation framework can assign fractional credit for partial
citations, providing a much more accurate reflection of the system's attribution capabilities than
a binary pass/fail metric.

The Complexities of Multilingual and Hebrew LLM

Evaluation
The architectural evaluation described thus far—utilizing LLM judges for N-to-M argument
mapping and fine-grained citation attribution—is highly complex. Implementing this entirely in
Hebrew introduces extreme methodological challenges, as the vast majority of evaluation
benchmarks, NLI models, and established LLM-as-a-Judge paradigms are deeply optimized for
the English language.

The Multilingual Performance Gap and Tokenization Inefficiencies

The literature consistently highlights a pronounced "multilingual performance gap." When
powerful LLMs are deployed as automated judges to evaluate non-English text, their
performance and reliability degrade significantly. Empirical studies assessing the reliability of
multilingual LLM-as-a-judge frameworks reveal severe inconsistencies across languages, with
average Fleiss' Kappa scores (measuring inter-rater reliability) plummeting to approximately 0.3
for low-resource or highly complex languages.
This degradation is fundamentally rooted in tokenization architectures. Standard LLM tokenizers
are disproportionately calibrated for English. When processing highly agglutinative and
morphologically dense languages like Hebrew—where single words frequently encode complex
syntactic relationships, prepositions, and grammatical gender—the tokenization process
becomes highly inefficient. This leads to fragmented semantic representations, diluting the
context window and severely impairing the LLM judge's ability to perform precise semantic
equivalence mapping or citation verification. Furthermore, strict length penalties or evaluation
criteria calibrated on English data often artificially penalize Hebrew outputs, leading to systemic
judge bias.

Designing the Multilingual Judge Prompt Architecture

To mitigate the multilingual performance gap, researchers have identified counterintuitive
strategies for prompt engineering when evaluating non-English text. Studies examining prompt
language translation indicate that translating the evaluation rubrics, scoring criteria, and meta-
instructions into the target language (Hebrew) often degrades the judge's performance.
Instead, the optimal strategy involves providing the complex evaluation rubric and system
instructions entirely in English, while feeding the LLM judge the input arguments, ground truth
data, and retrieved context in Hebrew. By maintaining the instruction layer in English, the LLM
judge leverages its most robust and heavily trained reasoning pathways to comprehend the
strict logic and rules of the evaluation task, while simultaneously applying its cross-lingual
alignment capabilities to assess the semantic meaning of the Hebrew data payloads.

Leveraging Hebrew-Native Sovereign Models: DictaLM 3.0

While proprietary frontier models (such as GPT-4o or Claude 3.5 Sonnet) possess strong
multilingual capabilities, deploying them as automated judges for highly specialized, domain-
specific Hebrew RAG systems can introduce unpredictable edge-case failures. The literature
strongly suggests integrating Hebrew-native sovereign language models into the evaluation
pipeline.
Models such as DictaLM 3.0 represent the current frontier for open-weight Hebrew-capable
models. DictaLM 3.0 underwent massive continuous pre-training (CPT) on over 130 billion
Hebrew tokens, demonstrating superior comprehension of Hebrew morphology, orthographic
variation, and cultural context compared to baseline models.
Architectural analyses of these models reveal a critical phenomenon known as "expert routing
collapse" within Mixture-of-Experts (MoE) architectures. In standard pre-trained base models,
the processing of Hebrew tokens is often routed to a very narrow, sub-optimal subset of experts,
while English tokens utilize the full breadth of the network. Fine-tuning on Hebrew-enriched
corpora, as seen in the DictaLM architecture, forces a structural reorganization of the routing
mechanism, breaking the collapse and ensuring deep, language-agnostic processing.
Integrating a model like DictaLM 3.0 as one of the judges within a multi-model ensemble
ensures that subtle semantic nuances, domain-specific terminology, and morphological
variations within the generated Hebrew arguments are accurately evaluated, preventing false
negatives during the Argument Recall phase.

A Comprehensive Evaluation Blueprint for the Target

Architecture
Synthesizing the literature on RAG assessment, atomic factuality, fine-grained citation
evaluation, and multilingual modeling yields a highly structured, multi-phase evaluation blueprint
tailored specifically for a Hebrew system outputting lists of arguments and citations.

Phase 1: Generative Argument Evaluation (The N-to-M Map)

The initial phase ignores the citation markers and evaluates the text of the generated Hebrew
arguments against the text of the ground truth Hebrew arguments.
1. Evaluating Argument Coverage (Recall) The system must guarantee that no vital
information from the ground truth is omitted. The evaluation script iterates through the list of
ground truth arguments. For each ground truth argument, the LLM Judge scans the entire list of
generated arguments.
● Prompt Architecture: English instructions, Hebrew data variables.
● Judge Task: "Analyze the provided Ground Truth Argument. Review the list of Generated
Arguments. Does any Generated Argument or combination of Generated Arguments
contain the semantic equivalent of the Ground Truth Argument? Output 'FULL
ENTAILMENT', 'PARTIAL ENTAILMENT', or 'NO ENTAILMENT' with a mandatory chain-
of-thought justification."
● Metric Calculation: Aggregate the scores (e.g., Full = 1.0, Partial = 0.5) and divide by the
total number of ground truth arguments to yield a percentage score for system
comprehensiveness.
2. Evaluating Argument Correctness (Precision) The system must guarantee that it does not
generate hallucinated or contradictory arguments. The evaluation script iterates through the list
of generated arguments. For each generated argument, the LLM Judge cross-references the
complete set of ground truth arguments.
● Judge Task: "Analyze the provided Generated Argument. Is it factually supported and
verified by the provided list of Ground Truth Arguments? Output 'FULL SUPPORT',
'PARTIAL SUPPORT', or 'CONTRADICTION/UNSUPPORTED'."
● Metric Calculation: Aggregate the scores and divide by the total number of generated
arguments to yield a percentage score for system accuracy.

Phase 2: Citation Attribution Evaluation

The second phase evaluates the precise citation markers appended to each generated
argument, implementing ALCE and TRACE methodologies to ensure verifiability.
1. Evaluating Citation Correctness and Faithfulness For every generated argument, the
evaluation isolates its appended citation markers. The script retrieves the exact text of the cited
documents.
● Judge Task: Provide the LLM judge with the Generated Argument and the raw text of the
Cited Document. "Analyze the Cited Document. Does the text within this document
contain sufficient and explicit evidence to fully entail the Generated Argument? Output
'FULL SUPPORT', 'PARTIAL SUPPORT', or 'NO SUPPORT'."
● Metric Calculation: This determines the true Citation Precision of the system, penalizing
the model for attaching citations that do not actually support the claim being made.
2. Evaluating Citation Necessity If a single argument is appended with multiple citations, the
judge evaluates the necessity of the cluster. If Citation A fully entails the argument, and Citation
B contains identical information, Citation B is flagged as redundant. If Citation C contains
irrelevant information, it is flagged as a hallucinated citation. This ensures the citation graph
remains dense and highly utilitarian.

Phase 3: Evaluation Infrastructure and Calibration

To operationalize this blueprint reliably, the engineering infrastructure must account for the non-
deterministic nature of LLM judges.
Infrastructure Requirement Implementation Strategy Impact on Evaluation Quality
Ensemble Judging Deploy a panel of three LLM Neutralizes the specific biases
judges (e.g., GPT-4o, Claude and systemic blind spots of any
3.5, and DictaLM 3.0) for every single commercial model,
evaluation point and utilize dramatically stabilizing the
majority voting. variance of the evaluation
scores.
Adaptive Calibration Maintain a golden dataset of Detects calibration drift or
100 human-annotated queries. prompt degradation. Ensures
Run the LLM judge panel the automated scores remain a
against this dataset weekly and trustworthy proxy for human
calculate Cohen's Kappa for expert judgment.
agreement.
Trace-Native Logging Attach the evaluation scores Contextualizes errors
and the LLM judge's CoT immediately. Allows developers
reasoning directly to the to see exactly which retrieved
execution trace of the specific chunk caused the model to
argument span, rather than hallucinate a specific citation or
exporting it to a disconnected argument.
database.
By implementing this comprehensive, multi-layered evaluation framework, developers can move
beyond superficial lexical metrics. This methodology rigorously interrogates the semantic
coverage of the generated arguments, the granular accuracy of the inline citations, and the
causal faithfulness of the entire retrieval pipeline, ensuring the deployment of a highly
trustworthy, multilingual RAG architecture.

Works cited

1. LLM Evaluation Metrics: The Ultimate LLM Evaluation Guide - Confident AI,
[Link]
2. How Reliable is Multilingual LLM-as-a-Judge? - ACL Anthology,
[Link] 3. Large language models in materials
science: assessing RAG evaluation frameworks through graphene synthesis - RSC Publishing,
[Link] 4. RAG evaluation: a technical
guide to measuring retrieval-augmented generation - Toloka AI, [Link]
evaluation-a-technical-guide-to-measuring-retrieval-augmented-generation/ 5. A Comparative
Evaluation of RAG Architectures for Cross-Domain LLM Applications: Design, Implementation,
and Assessment - IEEE Xplore,
[Link] 6. FActScore: Fine-grained
Atomic Evaluation of Factual Precision in Long Form Text Generation - arXiv,
[Link] 7. [2305.14251] FActScore: Fine-grained Atomic Evaluation
of Factual Precision in Long Form Text Generation - arXiv, [Link] 8.
OpenFActScore: Open-Source Atomic Evaluation of Factuality in Text Generation - arXiv,
[Link] 9. Reference-Guided Verdict: LLMs-as-Judges in Automatic
Evaluation of Free-Form QA - ACL Anthology, [Link]
10. LLM-as-a-Judge Simply Explained: The Complete Guide to Run LLM Evals at Scale,
[Link] 11. A
Comparative Analysis of Faithfulness Metrics and Humans in Citation Evaluation - Centrum
Wiskunde & Informatica, [Link] 12. LLM-as-a-judge: a complete
guide to using LLMs for evaluations - Evidently AI, [Link]
a-judge 13. Towards Fine-Grained Citation Evaluation in Generated Text: A Comparative
Analysis of Faithfulness Metrics - arXiv, [Link] 14. LLM-as-a-
Judge Evaluation with DeepEval, [Link] 15.
Debatable Intelligence: LLM Judges via Debate Speech Evaluation - Noy Sternlicht, [Link]
[Link]/Debatable-Intelligence-Web/ 16. CLEAR: Error Analysis via LLM-as-a-Judge
Made Easy | Proceedings of the AAAI Conference on Artificial Intelligence,
[Link] 17. Debatable Intelligence: Benchmarking
LLM Judges via Debate Speech Evaluation - arXiv, [Link] 18.
Evaluating the Effectiveness of LLM-Evaluators (aka LLM-as-Judge) - Eugene Yan,
[Link] 19. DeepMind FACTS Framework 2026: LLM
Factual Accuracy Guide - Galileo AI, [Link]
factual-accuracy 20. How to Correctly Report LLM-as-a-Judge Evaluations - arXiv,
[Link] 21. [2605.02520] Benchmarking Retrieval Strategies for
Biomedical Retrieval-Augmented Generation: A Controlled Empirical Study - arXiv,
[Link] 22. Generation-Time vs. Post-hoc Citation: A Holistic
Evaluation of LLM Attribution - arXiv, [Link] 23. ALCE | Notion,
[Link] 24. [2305.14627]
Enabling Large Language Models to Generate Text with Citations - arXiv,
[Link] 25. Training Language Models to Generate Text with Citations
via Fine-grained Rewards - arXiv, [Link] 26. ALiiCE: Evaluating
Positional Fine-grained Citation Generation - arXiv, [Link] 27.
ALiiCE: Evaluating Positional Fine-grained Citation Generation - ACL Anthology,
[Link] 28. Trustworthy Retrieval-Aligned Citation
Evaluation (TRACE) - Emergent Mind, [Link]
retrieval-aligned-citation-evaluation-trace 29. TRACE: A Framework for Analyzing and
Enhancing Stepwise Reasoning in Vision-Language Models - ACL Anthology,
[Link] 30. Towards Fine-Grained Citation Evaluation in
Generated Text: A Comparative Analysis of Faithfulness Metrics - ACL Anthology,
[Link] 31. Towards Reliable Multilingual LLMs-as-a-
Judge: An Empirical Study - arXiv, [Link] 32. Multilingual LLM
Evaluation - Emergent Mind, [Link]
large-language-models 33. The Origin of Multilingual Performance Gap: A Deep Dive into Multi-
Turn Conversational Agents - Lilt, [Link]
34. How Reliable is Multilingual LLM-as-a-Judge? - arXiv, [Link]
35. Adapting LLMs to Hebrew: Unveiling DictaLM 2.0 with Enhanced Vocabulary and Instruction
Capabilities - arXiv, [Link] 36. HEBATRON: A Hebrew-Specialized
Open-Weight Mixture-of-Experts Language Model Technical Report - arXiv,
[Link] 37. (PDF) Dicta-LM 3.0: Advancing The Frontier of Hebrew
Sovereign LLMs - ResearchGate, [Link]
LM_30_Advancing_The_Frontier_of_Hebrew_Sovereign_LLMs 38. HEBATRON: A Hebrew-
Specialized Open-Weight Mixture-of-Experts Language Model - arXiv,
[Link] 39. Dicta-LM 3.0: Advancing The Frontier of Hebrew Sovereign
LLMs - arXiv, [Link] 40. Mixture of Experts for Low-Resource LLMs -
arXiv, [Link]