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%).
256
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.
Analysis of 500k ChatGPT logs shows over one-third of conversations generate fiction, dominated by power users with repetitive and niche patterns.
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.
OR-Space is a benchmark for LLM agents performing full-lifecycle optimization tasks across Build, Revise, and Explain modes in executable multi-artifact workspaces.
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.
LogDx-CI benchmark shows hybrid grep+tail reducers achieve top diagnosis quality at low cost, agent loops shrink quality variance across reducers, and cross-family LLM summarizers outperform same-family pairs.
MuCRASP prunes VLMs in a CoT-aware manner, outperforming baselines by preserving reasoning quality at 30-50% compression rates on models like Qwen2.5-VL-7B.
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 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.
LLMs show severe staleness after training cutoffs and recency bias on historical German statutes; RAG with version filtering mitigates both better than web search.
A multi-agent pipeline iteratively refines topology optimization outputs to match natural language preferences for branched structures, achieving 60% success rate across replicates in cantilever and phone-stand tasks.
citing papers explorer
-
Studying Lobby Influence in the European Parliament
NLP comparison of lobby papers and MEP speeches discovers influence links validated indirectly via retweets and meetings, achieving AUC 0.77 and ideological alignment in aggregate analysis.
-
MAmmoTH: Building Math Generalist Models through Hybrid Instruction Tuning
MAmmoTH models trained via hybrid CoT-PoT instruction tuning on MathInstruct outperform prior open-source LLMs by 16-32% average accuracy on nine math datasets, reaching 33% and 44% on MATH for 7B and 34B scales.
-
Textbooks Are All You Need II: phi-1.5 technical report
phi-1.5 is a 1.3B parameter model trained on synthetic textbook data that matches the reasoning performance of models five times larger on natural language, math, and basic coding tasks.
-
ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate
Multi-agent debate among LLMs yields more reliable text evaluations than single-agent prompting by simulating collaborative human judgment.
-
Large Language Models are not Fair Evaluators
LLMs show strong position bias when scoring model outputs, allowing easy manipulation of rankings, but calibration with multiple evidence, position balancing, and selective human input reduces this bias to better match human judgments.
-
A French OSCE Dialogue Dataset and Controllable Virtual Patient System for Clinical Training
Introduces a French OSCE dialogue dataset of 240 interactions and a modular LLM-based controllable virtual patient generation system with multi-level LLM-as-Judge evaluation for clinical skills training.
-
TRACE: A taxonomy-grounded synthetic dataset for teaching-program generation and session interpretation in Applied Behavior Analysis
TRACE is a taxonomy-grounded synthetic instruction-tuning dataset with 2,999 examples for ABA teaching-program generation and multi-session behavioral interpretation, released with code, provenance, and stratified splits.
-
Greedy or not, here I come: Language production under vocabulary constraints in humans and resource-rational models
Humans produce language more like greedy local choices than globally optimal planning when vocabulary is tightly constrained, with skilled speakers showing more revision.
-
Tokenizer Fertility and Zero-Shot Performance of Foundation Models on Ukrainian Legal Text: A Comparative Study
Tokenizer fertility varies 1.6x across models on Ukrainian legal text, Qwen uses 60% more tokens than Llama-family models, zero-shot outperforms few-shot by up to 26 points, and pre-war classifiers lose 27.9 points on invasion-era decisions.
-
Why Expert Alignment Is Hard: Evidence from Subjective Evaluation
Expert alignment in subjective LLM evaluations is difficult because expert judgments are heterogeneous, partly tacit, dimension-dependent, and temporally unstable.
-
Cross-Model Consistency of AI-Generated Exercise Prescriptions: A Repeated Generation Study Across Three Large Language Models
Three LLMs exhibit distinct consistency profiles in repeated exercise prescription generation, with GPT-4.1 producing unique but semantically stable outputs while Gemini 2.5 Flash achieves high similarity through text duplication.
-
Bridging the Reasoning Gap in Vietnamese with Small Language Models via Test-Time Scaling
SFT on localized Vietnamese math data improves explanation quality by 77% in a 1.7B model, with CoT+self-consistency outperforming ReAct-style test-time scaling.
-
AlphaEval: Evaluating Agents in Production
AlphaEval is a benchmark of 94 production-sourced tasks from seven companies for evaluating full AI agent products across six domains using multiple judgment methods, plus a framework to build similar benchmarks.
-
From Hallucination to Structure Snowballing: The Alignment Tax of Constrained Decoding in LLM Reflection
Enforcing structured reflection via Outlines-based constrained decoding on an 8B LLM triggers structure snowballing instead of better self-correction, producing near-perfect syntax but persistent semantic errors and revealing an alignment tax.
-
Breakdowns in Conversational AI: Interactional Failures in Emotionally and Ethically Sensitive Contexts
Mainstream conversational models show escalating affective misalignments and ethical guidance failures during staged emotional trajectories, organized into a taxonomy of interactional breakdowns.
-
PEFT-Factory: Unified Parameter-Efficient Fine-Tuning of Autoregressive Large Language Models
PEFT-Factory supplies a ready-to-use, extensible codebase that unifies 19 PEFT methods and evaluation pipelines for fine-tuning large autoregressive language models.
-
MARS-SQL: A multi-agent reinforcement learning framework for Text-to-SQL
MARS-SQL trains a multi-agent RL system with ReAct-style interaction and generative validation to produce SQL queries, reaching 77.84% execution accuracy on BIRD dev and 89.75% on Spider test.
-
WildFeedback: Aligning LLMs With In-situ User Interactions And Feedback
WildFeedback extracts preference pairs from in-situ user feedback in LLM conversations to fine-tune models for better alignment with real user preferences.
-
KG2Cypher: Data-Centric Pipeline for Building Enterprise Text-to-Cypher Systems
KG2Cypher generates validated synthetic Text-Cypher pairs from existing KGs, trains LoRA models with class-conditioned schema prompting, and reports execution-result F1 gains to 0.950 and 0.92 plus 95.2% exact match in Korean enterprise settings.
-
ImmigrationQA: A Source-Grounded Dataset and Small-Model Adaptation for U.S. Immigration Law
A new source-grounded QA dataset for U.S. immigration law is built from official documents and used to fine-tune a 3B model, yielding a 27% mean score improvement over the base model on a held-out sample.
-
Domain Adaptation and Reasoning Frameworks in Language Models: A Controlled Experiment with Historical Cosmology
Fine-tuning on historical cosmology data reshapes language model explanatory frameworks, leading to stance changes as a secondary effect from regime redistribution.
-
Refining and Reusing Annotation Guidelines for LLM Annotation
An iterative moderation framework refines and reuses annotation guidelines to improve LLM annotation accuracy on biomedical NER tasks across GPT, Gemini, and DeepSeek models.
-
Internalizing Tool Knowledge in Small Language Models via QLoRA Fine-Tuning
QLoRA fine-tuning on tool-use data enables 4B-parameter models to perform structured planning without tool catalogs in prompts, outperforming informed baselines on AssetOpsBench while reducing input length by 82.6%.
-
Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
Gemini 2.5 Pro and Flash models are presented as achieving frontier performance in reasoning, coding, and long-context multimodal tasks while spanning a cost-capability Pareto curve.
-
From Holistic Evaluation to Structured Criteria: Rubrics Across the Evolving LLM Landscape
The paper frames rubrics as a recurring structured-criteria approach that decomposes holistic judgments at evaluative, training, and intrinsic levels in LLM research.
-
Towards EnergyGPT: A Large Language Model Specialized for the Energy Sector
Fine-tuned LLaMA 3.1-8B variants for the energy sector outperform the base model on domain QA benchmarks, with LoRA delivering similar gains at lower training cost.
- Lessons from the Trenches on Reproducible Evaluation of Language Models