REVIEW 22 cited by
LLM Critics Help Catch LLM Bugs
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
Reinforcement learning from human feedback (RLHF) is fundamentally limited by the capacity of humans to correctly evaluate model output. To improve human evaluation ability and overcome that limitation this work trains "critic" models that help humans to more accurately evaluate model-written code. These critics are themselves LLMs trained with RLHF to write natural language feedback highlighting problems in code from real-world assistant tasks. On code containing naturally occurring LLM errors model-written critiques are preferred over human critiques in 63% of cases, and human evaluation finds that models catch more bugs than human contractors paid for code review. We further confirm that our fine-tuned LLM critics can successfully identify hundreds of errors in ChatGPT training data rated as "flawless", even though the majority of those tasks are non-code tasks and thus out-of-distribution for the critic model. Critics can have limitations of their own, including hallucinated bugs that could mislead humans into making mistakes they might have otherwise avoided, but human-machine teams of critics and contractors catch similar numbers of bugs to LLM critics while hallucinating less than LLMs alone.
Forward citations
Cited by 22 Pith papers
-
Fantastic Adaptive Taxonomies and How to Use Them
AdaMAST induces a compact, evidence-grounded failure taxonomy from an agent's own traces and shows it improves agent-system search, runtime reflection, and trajectory selection over free-text and fixed-taxonomy baselines.
-
CodeJudgeBench: Benchmarking LLM-as-a-Judge for Coding Tasks
A 5,352-pair benchmark shows thinking LLMs outperform non-thinking ones as code judges, but position bias and formatting sensitivity remain strong.
-
A dataset of rated conceptual arguments
A multi-dimensional expert-rated dataset of 951 conceptual-argument critiques shows LLM judge performance tracks general model capability and is little helped by reasoning modes.
-
Judging Is Not Enumerating: Silent Omissions in LLM-Authored Acceptable Sets
LLMs under one-shot greedy decoding enumerate or materialize acceptable sets much worse than they judge membership, a gap that persists across scale, family, and generation and is dominated by omissions.
-
Beyond a Single Judge: The Evidence-Grounded, Social-Weighted Persona Panel for Generative UI Evaluation
Evidence-grounded persona panels with bounded-confidence deliberation raise GenUI judge–human correlation from 0.716 to 0.922, mostly from persona grounding rather than multi-prompt averaging.
-
Proof-or-Stop: Don't Trust the Agent, Trust the Evidence -- Loop Engineering for Verifiable Evidence-Gated Lifecycle Control
Proof-or-Stop gates autonomous coding lifecycle transitions on fresh, source-state-bound, mechanically verifiable evidence, cutting visible-pass/hidden-fail amplification from 31/1800 to 2/1800 injected cells in a 9,2...
-
Human-AI Complementarity: A Goal for Amplified Oversight
Confidence-based routing of fact-verification to humans, plus evidence-only AI assistance, beats either human or AI raters alone: 91.3% hybrid accuracy vs 87.7% for the AI rater.
-
ViseGPT: Towards Better Alignment of LLM-generated Data Wrangling Scripts and User Prompts
ViseGPT automatically converts user prompts into test cases and visualizes which steps of an LLM-generated data wrangling script pass or fail.
-
RefCritic: Training Long Chain-of-Thought Critic Models with Refinement Feedback
RefCritic uses reinforcement learning with two rule-based rewards, judgment accuracy and refinement accuracy, to train long-chain-of-thought critic models that improve LLM math reasoning and detect erroneous steps.
-
BMMR: A Large-Scale Bilingual Multimodal Multi-Discipline Reasoning Dataset
BMMR provides a 110k-question bilingual, multimodal, college-level dataset across 300 subjects where state-of-the-art models score at most about 50%.
-
Breakpoint: Scalable evaluation of system-level reasoning in LLM code agents
Breakpoint generates code-repair benchmarks by corrupting real GitHub functions and shows that frontier AI coding agents solve easy repairs but fail completely on tasks requiring coordinated, system-wide changes.
-
Quo Vadis, World Modeling?
An agent-centric reframing of world modeling, replacing physical state prediction with 'information transitions' organized into six proxy functions and three empowerment levels.
-
Code Monitor Red Teaming for Public-Test-Passing Code
Weak LLM verifiers miss most hidden bugs in public-test-passing code, and adversarial pressure makes that monitoring worse.
-
Dream-Coder 7B: An Open Diffusion Language Model for Code
An open 7B diffusion language model for code, adapted from Qwen2.5-Coder, reaches 21.4% pass@1 on LiveCodeBench and shows adaptive, non-left-to-right generation patterns.
-
OnGoal: Tracking and Visualizing Conversational Goals in Multi-Turn Dialogue with Large Language Models
OnGoal is an LLM chat interface that infers, merges, and evaluates user goals in real time and visualizes their progress, tested with 20 users on a writing task.
-
Let's Revise Step-by-Step: A Unified Local Search Framework for Code Generation with LLMs
The abstract claims a new local search framework for code generation, but the manuscript body is a different mathematics paper.
-
Are Today's LLMs Ready to Explain Well-Being Concepts?
AI judges can score explanations of well-being concepts, and small models fine-tuned with preference data score better than larger models, although judges and explainers are all AIs.
-
CRITICTOOL: Evaluating Self-Critique Capabilities of Large Language Models in Tool-Calling Error Scenarios
CRITICTOOL, a benchmark of 2,740 tool-calling error scenarios built from BFCL and T-Eval with GPT-4o-based error injection, finds most LLMs rarely recover from tool-use errors, with GPT-4o best at 69.01 overall and to...
-
A Survey of Automatic Evaluation Methods on Text, Visual and Speech Generations
A unified taxonomy and comparative meta-evaluation of automatic evaluation methods across text, vision, and speech generation, concluding that LLM-based evaluators dominate current practice.
-
Exchange of Perspective Prompting Enhances Reasoning in Large Language Models
A two-branch prompting method that exchanges answers between an original math question and a paraphrased version improves accuracy on several math benchmarks, but the gain is not separated from the extra compute or ru...
-
CRScore++: Reinforcement Learning with Verifiable Tool and AI Feedback for Code Review
A two-stage SFT+DPO pipeline using linter and code-smell tool outputs plus an LLM judge improves generated code review comments and appears to transfer from Python to Java and JavaScript, though the judge-based evalua...
-
Optimizing LLM-Based Multi-Agent System with Textual Feedback: A Case Study on Software Development
A locate-then-optimize prompt pipeline raises GPT-4-judged scores of a multi-agent software developer across five dimensions, but lacks error bars, code, and an evaluation metric independent of the optimization target.
Discussion (0). Continue with ORCID to comment.