mlsanityData to 7 Oct 2026

Great Models Think Alike and this Undermines AI Oversight

arXiv 2502.04313 ↗First posted 6 Feb 2025cs.LG cs.AI cs.CLcited by 53

On the map: LLM reasoning → Reasoning benchmarks and evaluation

Abstract (arXiv)

As Language Model (LM) capabilities advance, evaluating and supervising them at scale is getting harder for humans. There is hope that other language models can automate both these tasks, which we refer to as ''AI Oversight''. We study how model similarity affects both aspects of AI oversight by proposing Chance Adjusted Probabilistic Agreement (CAPA): a metric for LM similarity based on overlap in model mistakes. Using CAPA, we first show that LLM-as-a-judge scores favor models similar to the judge, generalizing recent self-preference results. Then, we study training on LM annotations, and find complementary knowledge between the weak supervisor and strong student model plays a crucial role in gains from ''weak-to-strong generalization''. As model capabilities increase, it becomes harder to find their mistakes, and we might defer more to AI oversight. However, we observe a concerning trend -- model mistakes are becoming more similar with increasing capabilities, pointing to risks from correlated failures. Our work underscores the importance of reporting and correcting for model similarity, especially in the emerging paradigm of AI oversight.

Key facts. Cited by 53 papers in the map (data through 7 Oct 2026). Placed in Reasoning benchmarks and evaluation, one of 17 directions in LLM reasoning. Builds on 0 of its 106 references that are in the map. First posted on arXiv 6 Feb 2025.

Builds on

0 of 106 references.

No references here yet.

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 16; 2026: 37.

53 papers.

Sort:

By area:

  1. 2025Correlated Errors in Large Language Modelscited by 42
  2. 2025Do LLM Evaluators Prefer Themselves for a Reason?cited by 38
  3. 2025Answer Matching Outperforms Multiple Choice for Language Model Evaluationcited by 37
  4. 2025How Benchmark Prediction from Fewer Data Misses the Markcited by 26
  5. 2025When Your AI Agent Succumbs to Peer-Pressure: Studying Opinion-Change Dynamics of LLMscited by 14
  6. 2026Stop Automating Peer Review Without Rigorous Evaluationcited by 12
  7. 2025PrivacyBench: A Conversational Benchmark for Evaluating Privacy in Personalized AIcited by 11
  8. 2025Discrepancies are Virtue: Weak-to-Strong Generalization through Lens of Intrinsic Dimensioncited by 10
  9. 2025PiCSAR: Probabilistic Confidence Selection And Ranking for Reasoning Chainscited by 9
  10. 2026The Hot Mess of AI: How Does Misalignment Scale With Model Intelligence and Task Complexity?cited by 8
  11. 2026Consensus is Not Verification: Why Crowd Wisdom Strategies Fail for LLM Truthfulnesscited by 8
  12. 2026Are Diversity Metrics Measuring Diversity? A Capability-Controlled Audit of Majority-Vote Gain in LLM Ensemblescited by 7
38 more
  1. 2025Quantitative LLM Judgescited by 6
  2. 2025Mitigating Self-Preference by Authorship Obfuscationcited by 6
  3. 2026When More Sampling Hurts: The Modal Ceiling and Correlation Ceiling of Test-Time Scalingcited by 6
  4. 2026Seven simple steps for log analysis in AI systemscited by 5
  5. 2026JEV vs. LLMs as Rubric Judges: Cheaper, Faster, and Wrong in the Same Placescited by 5
  6. 2025Weak-to-Strong Generalization via Bregman Bias-Variance Decompositioncited by 3
  7. 2025Contrastive Decoding Mitigates Score Range Bias in LLM-as-a-Judgecited by 3
  8. 2025Decentralized Arena: Towards Democratic and Scalable Automatic Evaluation of Language Modelscited by 2
  9. 2026Knowledge Divergence and the Value of Debate for Scalable Oversightcited by 2
  10. 2026Intersectional Sycophancy: How Perceived User Demographics Shape False Validation in Large Language Modelscited by 2
  11. 2026State-dependent error correlations shape voting thresholds in committees of AI agentscited by 2
  12. 2025What if I ask in textitalia lingua? Measuring Functional Similarity Across Languagescited by 1
  13. 2025Does Weak-to-strong Generalization Happen under Spurious Correlations?cited by 1
  14. 2025Teaming LLMs to Detect and Mitigate Hallucinationscited by 1
  15. 2026Dependence-Aware Label Aggregation for LLM-as-a-Judge via Ising Modelscited by 1
  16. 2026The Subjectivity of Monoculturecited by 1
  17. 2026Cutting LLM Evaluation Costs with SySRs: A Bandit Algorithm that Provably Exploits Model Similaritycited by 1
  18. 2026CIAware-Bench: Benchmarking Control Intervention Awareness Across Frontier LLMscited by 1
  19. 2026Hitting a Moving Target: Test-Time Adaptation for AI Text Detection under Continual Distribution Shiftcited by 1
  20. 2026Harnessing Disagreement: Detecting Correlated Agreement Blindness in Multi-Agent Triagecited by 1
  21. 2026Language Models Agree With Each Other, Not With Readerscited by 1
  22. 2026Who Judges Matters: Measuring Family-Conditioned Preference in LLM-as-Judge Panelscited by 1
  23. 2026Agreement Overstates Evidence: Error Dependence in LLM Judge Consensuscited by 1
  24. 2026Groundability, Not Scale Alone: When Weak Reviewers Can Audit Strong Coding Agentscited by 1
  25. 2026On the Blessing of Pre-training in Weak-to-Strong Generalizationcited by 0
  26. 2026The Periodic Table of LLM Reasoning: A Structured Survey of Reasoning Paradigms, Methods, and Failure Modescited by 0
  27. 2026Post-Training Recipe, More Than Model Family, Shapes Multi-Agent LLM Conversational Behaviorcited by 0
  28. 2026The Consistency Dilemma in LLMs: Generator-Evaluator Agreement and Vulnerability to Mistakescited by 0
  29. 2026Builder, Defender, Breaker: Measurable Independence and Bounded Autonomy When Generative Models Build, Defend and Test Softwarecited by 0
  30. 2026Sixteen models, fewer than two voices: measuring ensemble dispersion where no answer is uniquely correctcited by 0
  31. 2026Floor, Ceiling, and the Fusion Gap: How Much of Crowd Reading Attention Can Machines Predict?cited by 0
  32. 2026Sharding Prevents LLM Oversight Failures and Adversarial Exploitationcited by 0
  33. 2026Evaluator Ensembles Under Reward Hacking: Covariance Geometry and Finite-Search Guaranteescited by 0
  34. 2026Mapping and Measuring the Behavioral Evolution of Large Language Modelscited by 0
  35. 2026Inferred Generative-Process Diversity Predicts Correlated Failure Across Language Modelscited by 0
  36. 2026Why Better Models Can Create Riskier Systems: Evidence from LLM Agents in Financial Marketscited by 0
  37. 2026Debate-to-Skill: Capability-Bound Process Supervision for Industrial Query-to-Agent Annotationcited by 0
  38. 2026Theory of Scene: Breaking the Symmetry Trap in Multi-Agent LLM Coordinationcited by 0

50 most cited of 53 All 53 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2024Explanation, Debate, Align: A Weak-to-Strong Framework for Language Model Generalization
  2. 2023Large Language Model Alignment: A Survey
  3. 2025Alignment and Safety in Large Language Models: Safety Mechanisms, Training Paradigms, and Emerging Challenges
  4. 2025Self-Correction Bench: Uncovering and Addressing the Self-Correction Blind Spot in Large Language Models
  5. 2025Extracting and Understanding the Superficial Knowledge in Alignment
  6. 2025On the Reasoning Capacity of AI Models and How to Quantify It
  7. 2023The Alignment Problem in Context
  8. 2026Are Dilemmas and Conflicts in LLM Alignment Solvable? A View from Priority Graph