Pith. sign in

REVIEW 34 cited by

A Long Way to Go: Investigating Length Correlations in RLHF

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

arxiv 2310.03716 v2 pith:CA3OVUO5 submitted 2023-10-05 cs.CL cs.LG

classification cs.CLcs.LG
keywords rlhflengthmodelsrewardfindimprovementsbiasespreference
verification ladder T0 review T1 audit T2 compute T3 formal

Signed reviews

No signed human review yet.

0 comments
read the original abstract

Great success has been reported using Reinforcement Learning from Human Feedback (RLHF) to align large language models, with open preference datasets enabling wider experimentation, particularly for "helpfulness" in tasks like dialogue and web question answering. Alongside these improvements, however, RLHF also often drives models to produce longer outputs. This paper demonstrates, on three diverse settings, that optimizing for response length is, much more than previously thought, a significant factor behind RLHF. Studying the strategies RL optimization uses to maximize reward, we find improvements in reward to largely be driven by increasing response length, instead of other features. Indeed, we find that even a purely length-based reward reproduces most downstream RLHF improvements over supervised fine-tuned models. Testing a comprehensive set of length-countering interventions, we identify the dominant source of these biases to be reward models, which, by studying training dynamics, we find are non-robust and easily influenced by length biases in preference data.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 34 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. More Convincing, Not More Correct: Self-Play Reward Hacking of Reference-Free LLM Judges

    cs.LG 2026-07 accept novelty 7.0 of 10

    Self-play against reference-free LLM judges drives judge pass rates to 0.94 while true accuracy stays at 0.20, a reward-hacking basin that transfers across judge families and is prevented only by requiring the judge t...

  2. Attention Limited Reward Learning

    cs.AI 2026-07 conditional novelty 7.0 of 10

    Limited attention makes pairwise preference labels non-identifiable for reward, can reverse Bradley-Terry rankings, and bounds learning by attended information rather than raw label count.

  3. Preference learning made easy: Everything should be understood through win rate

    cs.LG 2025-02 conditional novelty 7.0 of 10

    Under two axioms (preference-consistency and prevalence-consistency), the only distribution-grounded evaluation for preference learning is h-win rate, and most popular alignment methods can be classified by whether th...

  4. Copilot Arena: A Platform for Code LLM Evaluation in the Wild

    cs.SE 2025-02 conditional novelty 7.0 of 10

    An in-IDE pairwise-preference platform for code LLMs reveals that real developer choices rank models differently than static coding benchmarks.

  5. Disentangling Length Bias In Preference Learning Via Response-Conditioned Modeling

    cs.LG 2025-02 conditional novelty 7.0 of 10

    Training reward models and DPO policies on response-conditioned preference pairs, where a length constraint is added to or withheld from the same prompt-response pair, reduces length bias and improves length instructi...

  6. Style over Substance: A Shortcut Audit of Emotion-Description Preference Evaluation

    cs.CV 2026-07 conditional novelty 6.5 of 10

    A content-blind probe using only description length and generator identity matches finetuned 7B multimodal judges on EmoPrefer, showing benchmark scores are reachable without video grounding.

  7. GCPO: Diagnosing and Constraining Subspace Geometry in Rollout RL for LLMs

    cs.LG 2026-08 conditional novelty 6.0 of 10

    A dimension-corrected overlap statistic shows that transient excursions into pretrained principal subspaces precede LLM rollout-RL degradation, and a hard bilateral projection method (GCPO) that prevents such excursio...

  8. RepoProbe: Benchmarking Architecture-Aware Repository Comprehension with Checklists

    cs.SE 2026-08 conditional novelty 6.0 of 10

    RepoProbe uses GitHub Discussions and checklist-based verification to measure repository-level code understanding in LLMs, finding a persistent gap between fluent explanations and evidence-grounded correctness plus a ...

  9. SMOPD: Multi-Reward Reinforcement Learning via Specialize-and-Merge Online Policy Distillation

    cs.LG 2026-08 conditional novelty 6.0 of 10

    SMOPD trains one reward-specialized teacher per reward and merges them via token-level online distillation into a balanced student, reporting gains over GDPO on tool-calling format compliance and safety alignment.

  10. Grounded Chess Reasoning in Language Models via Master Distillation

    cs.AI 2026-03 unverdicted novelty 6.0 of 10

    Master Distillation turns opaque chess-engine search into natural-language CoT, lifting a 4B model (C1) from near-zero to 48.1% accuracy that beats its teacher and most larger LMs.

  11. Mitigating Reward Hacking in RLHF via Bayesian Non-negative Reward Modeling

    cs.LG 2026-02 conditional novelty 6.0 of 10

    BNRM learns rewards as sparse non-negative factor combinations and shows reduced length bias and better OOD alignment than standard BT reward models.

  12. AALC: Large Language Model Efficient Reasoning via Adaptive Accuracy-Length Control

    cs.CL 2025-06 conditional novelty 6.0 of 10

    AALC, an accuracy-aware length reward, reduces reasoning-model response length by over 50% on math benchmarks while preserving or improving accuracy.

  13. Med-U1: Incentivizing Unified Medical Reasoning in LLMs via Large-scale Reinforcement Learning

    cs.CL 2025-06 conditional novelty 6.0 of 10

    A pure reinforcement learning recipe with mixed rule-based rewards and length control improves Qwen2.5-based models across diverse medical QA formats.

  14. Debiasing Online Preference Learning via Preference Feature Preservation

    cs.LG 2025-06 conditional novelty 6.0 of 10

    PFP preserves the distribution of human preference features during online preference learning by mapping each new instruction to a feature set, converting it to system prompts, and training with DPO.

  15. Multi-Domain Explainability of Preferences

    cs.CL 2025-05 conditional novelty 6.0 of 10

    A concept-discovery plus hierarchical regression pipeline explains human, LLM-judge, and reward-model preferences at local and global levels across eight domains.

  16. MPO: Multilingual Safety Alignment via Reward Gap Optimization

    cs.CL 2025-05 conditional novelty 6.0 of 10

    MPO reduces jailbreak success in multilingual LLMs by regressing target-language reward gaps onto the English reward gap, outperforming DPO and related methods while preserving utility.

  17. Bias Fitting to Mitigate Length Bias of Reward Model in RLHF

    cs.LG 2025-05 conditional novelty 6.0 of 10

    FiMi-RM reduces length bias in RLHF reward models by fitting a nonlinear length-reward curve and training the reward model to be uncorrelated with that fitted curve.

  18. Scalable Chain of Thoughts via Elastic Reasoning

    cs.LG 2025-05 conditional novelty 6.0 of 10

    Training reasoning models with a fixed two-part token budget, thinking plus solution, makes them robust to truncated thinking and more concise overall.

  19. A RAG-Based Multi-Agent LLM System for Natural Hazard Resilience and Adaptation

    cs.CL 2025-04 conditional novelty 6.0 of 10

    WildfireGPT, a multi-agent RAG system with user profiling, outperforms ChatClimate and Perplexity AI in location-specific wildfire data analysis and evidence-based recommendations across ten expert case studies.

  20. Self-Consistency of the Internal Reward Models Improves Self-Rewarding Language Models

    cs.AI 2025-02 conditional novelty 6.0 of 10

    Forcing a self-rewarding LLM's generative judge and implicit DPO reward to agree on preference labels substantially improves its AlpacaEval 2.0 alignment score.

  21. Online Preference Alignment for Language Models via Count-based Exploration

    cs.LG 2025-01 conditional novelty 6.0 of 10

    COPO adds a pseudo-count exploration bonus to DPO-style preference optimization and reports improved online RLHF results on Zephyr and Llama-3.

  22. When Can Proxies Improve the Sample Complexity of Preference Learning?

    cs.LG 2024-12 reject novelty 6.0 of 10

    Under four structural conditions linking proxy and true preference policies, the true policy is a low-dimensional adapter of the proxy policy, reducing the number of true preference samples needed.

  23. UAlign: Leveraging Uncertainty Estimations for Factuality Alignment on Large Language Models

    cs.CL 2024-12 conditional novelty 6.0 of 10

    UAlign improves LLM factuality alignment by adding predicted confidence and semantic entropy as input features to prompts and the reward model, helping the model answer known questions and refuse unknown ones.

  24. Self-Generated Critiques Boost Reward Modeling for Language Models

    cs.CL 2024-11 conditional novelty 6.0 of 10

    Self-generated, label-filtered critiques jointly trained with reward prediction improve preference accuracy and data efficiency without a stronger teacher model.

  25. Mapping out the Space of Human Feedback for Reinforcement Learning: A Conceptual Framework

    cs.LG 2024-11 conditional novelty 6.0 of 10

    A conceptual framework classifies human feedback to RL agents along nine dimensions and seven quality criteria, unifying human-centered, interface-centered, and model-centered design perspectives.

  26. Establishing Reliability Metrics for Reward Models in Large Language Models

    cs.AI 2025-04 conditional novelty 5.0 of 10

    RETA measures reward model reliability as the average oracle quality of the top η fraction of ranked responses, divided by the average response quality.

  27. Beyond Reward Hacking: Causal Rewards for Large Language Model Alignment

    cs.LG 2025-01 conditional novelty 5.0 of 10

    Causal reward modeling applies MMD regularization to RLHF reward models to make reward scores statistically independent of spurious features, reducing measured length, sycophancy, concept, and demographic biases in ex...

  28. Utility-inspired Reward Transformations Improve Reinforcement Learning Training of Language Models

    cs.LG 2025-01 conditional novelty 5.0 of 10

    Applying a utility-inspired, threshold-based transformation to individual rewards before summing them improved the harmlessness of an RLHF-trained 2B language model without reducing helpfulness.

  29. CoT-based Synthesizer: Enhancing LLM Performance through Answer Synthesis

    cs.CL 2025-01 conditional novelty 5.0 of 10

    A small CoT-based synthesizer model trained on candidate-response analysis improves LLM reasoning accuracy, including cases where all sampled candidate answers are incorrect.

  30. Interpreting Language Reward Models via Contrastive Explanations

    cs.LG 2024-11 conditional novelty 5.0 of 10

    Reward model preferences can be explained by generating counterfactual and semifactual answer variations along 15 hand-picked evaluation attributes and measuring which attribute changes flip the model's preference.

  31. Reward Modeling for Reinforcement Learning-Based LLM Reasoning: Design, Challenges, and Evaluation

    cs.LG 2026-02 conditional novelty 4.0 of 10

    A taxonomy-driven survey arguing that reward design is the central mechanism shaping reliable LLM reasoning, with maps of reward paradigms, reward-hacking failure modes, and benchmark pitfalls.

  32. Meta-Thinking in LLMs via Multi-Agent Reinforcement Learning: A Survey

    cs.AI 2025-04 conditional novelty 4.0 of 10

    The paper surveys existing work on LLM meta-thinking and argues that multi-agent reinforcement learning is a promising missing ingredient for building self-correcting language models.

  33. Open Problems in Machine Unlearning for AI Safety

    cs.LG 2025-01 conditional novelty 4.0 of 10

    Machine unlearning is not a comprehensive solution for AI safety; it is best suited to data removal, while capability control faces fundamental and unresolved challenges.

  34. Reinforcement Learning from Human Feedback

    cs.LG 2025-04 unverdicted novelty 2.0 of 10

    The book introduces the origins, mathematical setup, and optimization stages of RLHF including reward modeling, reinforcement learning, rejection sampling, and direct alignment algorithms.

Pith tools