The authors create the first large-scale dataset and taxonomy of failure modes in multi-agent LLM systems to explain their limited performance gains.
super hub Mixed citations
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
Mixed citation behavior. Most common role is background (47%).
296
Pith papers citing it
Background
47% of classified citations
abstract
Evaluating large language model (LLM) based chat assistants is challenging due to their broad capabilities and the inadequacy of existing benchmarks in measuring human preferences. To address this, we explore using strong LLMs as judges to evaluate these models on more open-ended questions. We examine the usage and limitations of LLM-as-a-judge, including position, verbosity, and self-enhancement biases, as well as limited reasoning ability, and propose solutions to mitigate some of them. We then verify the agreement between LLM judges and human preferences by introducing two benchmarks: MT-bench, a multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our results reveal that strong LLM judges like GPT-4 can match both controlled and crowdsourced human preferences well, achieving over 80% agreement, the same level of agreement between humans. Hence, LLM-as-a-judge is a scalable and explainable way to approximate human preferences, which are otherwise very expensive to obtain. Additionally, we show our benchmark and traditional benchmarks complement each other by evaluating several variants of LLaMA and Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with human preferences are publicly available at https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.
hub tools
citation-role summary
background 15
method 10
dataset 4
baseline 1
citation-polarity summary
claims ledger
- abstract Evaluating large language model (LLM) based chat assistants is challenging due to their broad capabilities and the inadequacy of existing benchmarks in measuring human preferences. To address this, we explore using strong LLMs as judges to evaluate these models on more open-ended questions. We examine the usage and limitations of LLM-as-a-judge, including position, verbosity, and self-enhancement biases, as well as limited reasoning ability, and propose solutions to mitigate some of them. We then verify the agreement between LLM judges and human preferences by introducing two benchmarks: MT-be
authors
co-cited works
representative citing papers
ORPO performs preference alignment during supervised fine-tuning via a monolithic odds ratio penalty, allowing 7B models to outperform larger state-of-the-art models on alignment benchmarks.
LongBench is the first bilingual multi-task benchmark for long context understanding in LLMs, containing 21 datasets in 6 categories with average lengths of 6711 words (English) and 13386 characters (Chinese).
Gradient and greedy search over token suffixes produces universal, transferable adversarial prompts that elicit objectionable outputs from aligned models including black-box commercial systems.
FARS deployed at scale produced 166 AI/ML papers across 67 topics that received 282 structured human reviews indicating some review-worthy outputs alongside recurring failure modes.
The paper presents EMPATH, a new multilingual multi-turn benchmark for safety evaluation of emotional-support chatbots that uses separate auditor and judge models and releases its pipeline and rubrics.
CLQT is a new closed-loop, cost-aware benchmark that diagnoses LLM trading agent capabilities through strategy-consistent metrics and hash-verifiable trails rather than outcome rankings.
Experiments show prompt injection boosts resume rankings in LLM screening when rare and qualities homogeneous, but fails when common or qualities differ.
C3-Bench supplies a multi-domain dataset and LLM-based evaluation protocol that exposes systematic failures in existing change captioning models outside their training regimes.
MEMPROBE is a benchmark for direct recovery of hidden user states from LLM agent memory, showing task success and memory recovery as distinct capabilities with moderate recovery scores around 0.6.
ChartWalker provides a hierarchical knowledge graph construction method and structure-aware sampling to generate cross-chart RAG benchmarks, releasing ChartWalker-Bench that exposes performance gaps across RAG paradigms.
Analysis of 500k ChatGPT logs shows over one-third of conversations generate fiction, dominated by power users with repetitive and niche patterns.
Introduces the Power Systems Agent Benchmark with 41 task families across eight power engineering areas for executable evaluation of AI agents using deterministic feasibility checks.
User facts are internalized as surgical local edits to a hash-keyed Engram memory table with reasoning skill held in a shared adapter, claimed to match LoRA recall, improve indirect reasoning 5.6x on average, and compose across users with 33,000x smaller footprint than per-user adapters.
Wall-clock leaky-integrator monitors on variable-cadence agent streams exhibit a sharp cliff between constant-alarm and silent regimes, while sample-time CUSUM monitors remain invariant; transition detection avoids the trap.
RealMath-Eval benchmark shows LLM judges have an evaluation gap, performing worse on diverse real human math reasoning than on synthetic solutions due to greater error diversity and higher surprisal.
The paper defines Cherry-pick Override (CCO) as unauthorized directional commitment by LLM judges under mixed evidence and quantifies its prevalence (>84% on AVeriTeC conflicting subset) while testing intervention ladders and a two-channel reference probe.
MAC-Bench is a new adversarial benchmark that converts legal texts into executable scenarios via the SERV pipeline to measure procedural compliance in multi-agent LLM systems using CSR and MG metrics.
LLM judges exhibit high stability under neutral re-evaluation but substantial reversibility under targeted post-decision challenges, quantified via a new Evaluation Robustness Score (ERS).
CoEval generates task-specific benchmarks by rotating models through teacher, student, and judge roles, then weights questions by discriminative power and judges by panel consensus to recover accurate model rankings without labels.
Low-resource safety failures are action failures because the harmfulness representation transfers but the decision calibration does not; this is fixed by recalibrating a high-resource gate with 1-4 target-language examples.
RWGBench is a citation-centric benchmark for related work generation built from 40k CS papers and a 100-paper test set, with multi-dimensional metrics that better match human expert judgment than standard similarity scores.
Introduces (ε,q,t,A)-behavioral indistinguishability and shows via Qwen/Llama experiments that LoRA distillation boosts semantic similarity but leaves detectable behavioral differences under adversarial evaluation.
A Behavioral Specification interpretive layer improves representational accuracy for AI personalization by compressing user data into patterns, outperforming raw corpora and commercial memory systems on held-out behavioral predictions across 14 autobiographical corpora while reducing context cost.
citing papers explorer
-
ORPO: Monolithic Preference Optimization without Reference Model
ORPO performs preference alignment during supervised fine-tuning via a monolithic odds ratio penalty, allowing 7B models to outperform larger state-of-the-art models on alignment benchmarks.
-
LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding
LongBench is the first bilingual multi-task benchmark for long context understanding in LLMs, containing 21 datasets in 6 categories with average lengths of 6711 words (English) and 13386 characters (Chinese).
-
Universal and Transferable Adversarial Attacks on Aligned Language Models
Gradient and greedy search over token suffixes produces universal, transferable adversarial prompts that elicit objectionable outputs from aligned models including black-box commercial systems.
-
MEMPROBE: Probing Long-Term Agent Memory via Hidden User-State Recovery
MEMPROBE is a benchmark for direct recovery of hidden user states from LLM agent memory, showing task success and memory recovery as distinct capabilities with moderate recovery scores around 0.6.
-
AI Fiction in the Wild
Analysis of 500k ChatGPT logs shows over one-third of conversations generate fiction, dominated by power users with repetitive and niche patterns.
-
CoEval: Ranking Language Models for Custom Tasks Without Labeled Data or Trustworthy Benchmarks
CoEval generates task-specific benchmarks by rotating models through teacher, student, and judge roles, then weights questions by discriminative power and judges by panel consensus to recover accurate model rankings without labels.
-
Low-Resource Safety Failures Are Action Failures, Not Representation Failures
Low-resource safety failures are action failures because the harmfulness representation transfers but the decision calibration does not; this is fixed by recalibrating a high-resource gate with 1-4 target-language examples.
-
Beyond Recall: Behavioral Specification as an Interpretive Layer for AI Personalization
A Behavioral Specification interpretive layer improves representational accuracy for AI personalization by compressing user data into patterns, outperforming raw corpora and commercial memory systems on held-out behavioral predictions across 14 autobiographical corpora while reducing context cost.
-
Does Capability Transfer to Subjective Behavior -- and Would Our Instruments Tell Us? A Self-Evolving, Trust-by-Construction Evaluation Paradigm
Self-evolving rubric with anti-gaming fitness reveals that objective capability scaling fails to transfer to subjective LLM behaviors, with advice-restraint as the universal lowest dimension that can regress.
-
SomaliBench Eval: Measuring English-to-Somali Refusal Gaps in Open-Weight Language Models
SomaliBench finds large English-to-Somali refusal gaps (0.38 to 0.90) across Llama-3.1-8B, Gemma-2-9B, Qwen-2.5-7B, and Aya-23-8B, with many Somali responses being unclear rather than compliant.
-
ContextEcho: A Benchmark for Persona Drift in Long Agentic-Coding Sessions
ContextEcho benchmark shows persona drift occurs across 23 frontier models in long agentic-coding sessions, is not reliably reset by compaction, and can be restored by single-shot anchors with mode-dependent effects.
-
Asking For An Old Friend: Diagnosing and Mitigating Temporal Failure Modes in LLM-based Statutory Question Answering
LLMs show severe staleness after training cutoffs and recency bias on historical German statutes; RAG with version filtering mitigates both better than web search.
-
EditPropBench: Measuring Factual Edit Propagation in Scientific Manuscripts
EditPropBench evaluates LLM editors on propagating factual edits to dependent claims in synthetic scientific manuscripts, showing that even the strongest systems miss roughly 30% of required updates on hard cases.
-
Subliminal Steering: Stronger Encoding of Hidden Signals
Subliminal steering transfers complex behavioral biases and the underlying steering vector through fine-tuning on innocuous data, achieving higher precision than prior prompt-based methods.
-
CompliBench: Benchmarking LLM Judges for Compliance Violation Detection in Dialogue Systems
CompliBench uses simulation and adversarial flaw injection to create labeled dialogue data showing that top proprietary LLMs perform poorly at spotting guideline violations while fine-tuned smaller models outperform them and generalize to new domains.
-
FrontierFinance: A Long-Horizon Computer-Use Benchmark of Real-World Financial Tasks
FrontierFinance benchmark shows human financial experts outperform state-of-the-art LLMs by achieving higher scores and more client-ready outputs on realistic long-horizon tasks.
-
Synthetic Data Generation for Training Diversified Commonsense Reasoning Models
A two-stage synthetic data generation method creates the CommonSyn dataset, allowing LLMs fine-tuned on it to produce more diverse and higher-quality commonsense responses than vanilla or human-data-trained models.
-
VLegal-Bench: Cognitively Grounded Benchmark for Vietnamese Legal Reasoning of Large Language Models
VLegal-Bench supplies 10,450 expert-validated samples for evaluating LLMs on Vietnamese legal questions, retrieval, multi-step reasoning, and scenario solving.
-
TF1-EN-3M: Three Million Synthetic Moral Fables for Training Small, Open Language Models
The authors generate and publicly release the first large-scale open dataset of three million structured moral fables produced by small open language models together with a reproducible LLM-judge evaluation pipeline.
-
Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!
Fine-tuning aligned LLMs compromises safety guardrails even with minimal adversarial examples or benign data, creating new risks not covered by existing inference-time protections.
-
WizardLM: Empowering large pre-trained language models to follow complex instructions
WizardLM uses LLM-driven iterative rewriting to generate complex instruction data and fine-tunes LLaMA to reach over 90% of ChatGPT capacity on 17 of 29 evaluated skills.
-
Do Encoders Suffice? A Systematic Comparison of Encoder and Decoder Safety Judges for LLM Adversarial Evaluation
Fine-tuned ModernBERT-family encoders match LLM judges on F1, false negative rate, and precision-recall for harmful output detection across adversarial datasets and attack types while promising lower cost and latency.
-
Towards Spec Learning: Inference-Time Alignment from Preference Pairs
Proposes compiling preference pairs into readable natural-language specifications for inference-time LLM alignment, claiming outperformance over DPO on dense-preference domains.
-
Same question, different history: language, national identity, and credit in large language models
Analysis of 11 LLMs on 21 disputed inventions across 12 languages and 75,896 responses finds query language systematically shifts credit toward lower-status claimants in their associated language while Anglophone figures remain stable.
-
The Metanym Game: A Self-Contained, Self-Consistent LLM Peer-Community Benchmark for Structural Intelligence
Proposes the Metanym Game as a self-contained LLM benchmark using peer ratings and SVD to extract generator and judge competence, with skills dissociating and correlation to GPQA Diamond at r=0.92.
-
Self-Preference Is Weak or Absent in Verifiable Instruction-Following Revision: A Four-Model Test Under Genuine Authorship
No detectable self-preference in LLMs accepting verified fixes to their own IFEval drafts versus neutral judges across four models and 85 comparisons (gap -5.1 pp, CI crossing zero).
-
OPD-Evolver: Cultivating Holistic Agent Evolver via On-Policy Distillation
OPD-Evolver uses on-policy self-distillation in fast interaction and slow attribution loops to build agents with holistic memory competence, outperforming prior systems by up to 11.5% and allowing a 9B model to compete with much larger ones.
-
M\"OVE: A Holistic LLM Benchmark for the German Public Sector
MÖVE presents a new German-language benchmark evaluating 39 LLMs on performance and governance criteria using ten public-administration datasets.
-
Substrate Asymmetry in User-Side Memory: A Diagnostic Framework
User memory in LLMs factors into three orthogonal axes where parametric adapters and retrieval show opposite strengths, with causal evidence from attention interventions and an alignment tax on RLHF models.
-
Are We Evaluating Knowledge or Phrasing? Mitigating MCQA Sensitivity with ParaEval
ParaEval reduces false performance gaps in MCQA benchmarks from over 2 points to below 1 point by scoring models on multiple paraphrases per answer option instead of single surface forms.
-
Catching One in Five: LLM-as-Judge Blind Spots in Production Multi-Turn Transaction Agents
Empirical study of a production multi-turn ordering agent finds LLM-as-judge recall below 25% for human-confirmed defects, missing cross-turn state issues due to limited rubric and routing.
-
From Player to Master: Enhancing Test-Time Learning of LLM Agents via Reinforcement Learning over Memory
MemoPilot trains memory updates for LLM agents via multi-turn GRPO on RPS and poker, achieving top Elo scores and outperforming baselines including DeepSeek-V3.2.
-
From Outliers to Errors: Auditing Pali-to-English LLM Translations with Multi-Reference Adjudication
A multi-reference audit framework for LLM translations of the Pali Canon uses embedding drift from a human reference centroid to triage candidates for LLM-judge adjudication, showing drift correlates with major error rates and model-specific differences in the high-drift tail.
-
Target-Side Paraphrase Augmentation for Sign Language Translation with Large Language Models
LLM-generated target-side paraphrases improve BLEU-4 from 9.56 to 10.33 on PHOENIX14T for a pose-based Transformer in sign language translation, with limits observed on other datasets.
-
Draft-OPD: On-Policy Distillation for Speculative Draft Models
Draft-OPD applies on-policy distillation via target-assisted generation and error replay to train speculative draft models, yielding over 5x lossless acceleration and gains over EAGLE-3 and DFlash.
-
Agreement Metrics for LLM-as-Judge Evaluation: What to Report and Why
For binary LLM judge validation, Pearson's r, Spearman's ρ, Kendall's τ_b, phi, and Matthews correlation all equal a single number on non-degenerate data, Cohen's κ supplies the extra signal on label-rate drift, and a reporting checklist is provided.
-
STREAM: A Data-Centric Framework for Mining High-Value Task-Oriented Dialogues from Streaming Media
Stream mines streaming media to create and release StreamDial, a dataset of 87,498 structured task-oriented dialogue sessions across automotive, restaurant, and hotel domains using persona construction, Conversational Blueprints, and RAG.
-
EmbGen: Teaching with Reassembled Corpora
EmbGen creates synthetic QA data by entity decomposition, embedding-based reassembly into clusters, and multi-level sampling with cluster-specific prompts, yielding up to 88.9% higher Binary Accuracy than baselines on heterogeneous datasets under fixed token budgets.
-
Transitivity Meets Cyclicity: Explicit Preference Decomposition for Dynamic Large Language Model Alignment
Introduces HRC model for game-theoretic decomposition of preferences into orthogonal transitive and cyclic components, paired with DSPPO for dynamic Nash-seeking alignment, reporting gains over BT and GPM baselines on RewardBench and downstream LLM evaluations.
-
Why Do Safety Guardrails Degrade Across Languages?
A latent variable IRT framework decouples four safety-driving factors across 61 model configurations and 10 languages using 1.9 million evaluations, revealing that safety is largely unidimensional and that high cross-lingual gaps cluster in physical harm prompts and lower-resource languages.
-
Agentic AI Translate: An Agentic Translator Prototype for Translation as Communication Design
Describes a conceptual agentic prototype for AI translation that operationalizes skopos theory and GEMBA-MQM verification into a four-stage cycle with user dialogue and memory for coherence.
-
PQR: A Framework to Generate Diverse and Realistic User Queries that Elicit QA Agent Failures
PQR framework generates diverse realistic queries to elicit QA agent failures, uncovering 23-78% more unhelpful responses than prior methods in e-commerce agent tests.
-
FINESSE-Bench: A Hierarchical Benchmark Suite for Financial Domain Knowledge and Technical Analysis in Large Language Models
FINESSE-Bench is a new hierarchical benchmark suite combining certification-style exams, trading tasks, and a Russian olympiad set to evaluate LLMs on financial competencies at multiple difficulty levels.
-
Dual Hierarchical Dialogue Policy Learning for Legal Inquisitive Conversational Agents
A dual hierarchical RL framework with two agents coordinates high-level dialogue strategy and low-level question generation to emulate judicial questioning and extract key information from Supreme Court arguments, outperforming baselines.
-
BioTool: A Comprehensive Tool-Calling Dataset for Enhancing Biomedical Capabilities of Large Language Models
BioTool dataset enables fine-tuning a 4B-parameter LLM to outperform GPT-5.1 in biomedical tool calling while improving downstream answer quality per human experts.
-
RLearner-LLM: Balancing Logical Grounding and Fluency in Large Language Models via Hybrid Direct Preference Optimization
RLearner-LLM achieves up to 6x gains in NLI entailment over standard fine-tuning by using an automated hybrid DPO pipeline that balances logic and fluency across multiple model sizes and domains.
-
LLMs Corrupt Your Documents When You Delegate
LLMs corrupt an average of 25% of document content during long delegated editing workflows across 52 domains, even frontier models, and agentic tools do not mitigate the issue.
-
DQA: Diagnostic Question Answering for IT Support
DQA maintains persistent diagnostic state and aggregates retrievals at the root-cause level to reach 78.7% success on 150 enterprise IT scenarios versus 41.3% for standard multi-turn RAG while cutting average turns from 8.4 to 3.9.
-
Response-Based Knowledge Distillation for Multilingual Jailbreak Prevention Unwittingly Compromises Safety
Distilling safe refusal behavior from OpenAI o1-mini into Llama-3, Gemma-2, and Qwen3 models via response-based LoRA on multilingual jailbreak data increases jailbreak success rates on MultiJail by up to 16.6 points.
-
Reading Between the Lines: The One-Sided Conversation Problem
The one-sided conversation problem is introduced with empirical results showing that future-turn access and utterance length improve missing-turn reconstruction while high-quality summaries are possible without reconstruction on standard dialogue datasets.