Reward Bias Substitution: Single-Axis Bias Mitigations Redirect Optimization Pressure

• URL: http://arxiv.org/abs/2605.27996v2
• Date: Fri, 29 May 2026 02:24:56 GMT
• Title: Reward Bias Substitution: Single-Axis Bias Mitigations Redirect Optimization Pressure
• Abstract summary: Single-axis mitigations of reward-model biases can rotate optimization pressure onto correlated proxies rather than eliminate it.<n>The failure is enabled by a measurement-versus-optimization gap between audit and policy-induced distributions.<n>We formalize mitigation outcomes into a regime taxonomy and prove that successful mitigation, bias substitution, and overcorrection produce identical observables.
• Score: 23.021647176206972
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Single-axis mitigations of reward-model biases (e.g., reducing proxy reliance on length, sycophancy, or style) can rotate optimization pressure onto correlated proxies rather than eliminate it, a failure mode we call reward bias substitution. The failure is enabled by a measurement-versus-optimization gap between audit and policy-induced distributions during mitigation evaluation and policy training. We formalize mitigation outcomes into a regime taxonomy and prove that successful mitigation, bias substitution, and overcorrection produce identical observables under any audit-distribution scoring, including ranking accuracy and win-rate, even when granted oracle access to the true reward. Across published preference-learning mitigation work, no method we survey reports the evidence needed to certify successful mitigation. Augmenting evaluation with policy-induced distributions while tracking multiple biases provably closes the gap, and we translate this into actionable prescriptions for mitigation methods and benchmarks. We demonstrate bias substitution in language model RLHF, where a length penalty during GRPO training compresses responses as intended yet redirects optimization pressure onto confidence calibration, driving the policy into overconfidence while factual free-form accuracy falls. We also show a published length-debiasing operator that zeroes reward-length correlation on the audit distribution but reintroduces bias under best-of-N selection on three of four SOTA reward models, and a length-sycophancy coupling whose direction reverses under human-LLM judge disagreement.

Related papers

• FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience [18.64380706071327]
  FlowBalance is a verifier-grounded self-improvement method that learns a normalized distribution over complete responses.<n>For each on-policy trajectory, a frozen training-time view of the same policy uses privileged context to produce token-level log-probability gains.<n>Our analysis establishes within-group contrast preservation, a minimum-change reverse-KL characterization, monotonic verifier control of target reward, and an exact correction against false-positive self-guidance on rejected responses.
  arXiv  Detail & Related papers  (2026-09-03T00:47:11Z)
• Conditional Equivalence of DPO and RLHF: Implicit Assumption, Failure Modes, and Provable Alignment [51.18269946911088]
  Direct Preference Optimization (DPO) has emerged as a popular alternative to Reinforcement Learning from Human Feedback (RLHF)<n>We show that DPO and RLHF optimize fundamentally different objectives in such cases.<n>We introduce Constrained Preference Optimization (CPO), augmenting RLHF with constraints for provable alignment.<n>Our theoretical analysis establishes when DPOs' guarantees hold and provides solutions preserving simplicity with provable alignment.
  arXiv  Detail & Related papers  (2026-05-20T07:26:22Z)
• Online Conformal Prediction with Corrupted Feedback [42.94410959330529]
  Online conformal prediction (OCP) addresses this challenge through adaptively updated prediction sets.<n>These guarantees hinge on the assumption of perfect feedback about the coverage of past prediction sets.<n>In practice, the observed miscoverage indicator may be corrupted by noise, communication failures, or adversarial manipulation.
  arXiv  Detail & Related papers  (2026-05-19T21:34:37Z)
• Interactive Critique-Revision Training for Reliable Structured LLM Generation [18.00222080273147]
  We propose DPA-GRPO, a paired-action training method for a two-player generator--verifier game with structured verifier interventions.<n>We analyze the unregularized game and show that positive probability on strictly lower-reward intervention or revision actions creates a profitable unilateral deviation.<n>Experiments on TaxCalc TY24 show that DPA-GRPO improves structured decision accuracy over zero-shot generation and generator-only RL baselines.
  arXiv  Detail & Related papers  (2026-05-08T17:00:38Z)
• SiMPO: Measure Matching for Online Diffusion Reinforcement Learning [52.46919717963149]
  We introduce SiMPO, a simple and unified framework that generalizes reweighting scheme in diffusion RL with general monotonic functions.<n>SiMPO revisits diffusion RL via a two-stage measure matching lens.<n>We provide geometric interpretations to illustrate how negative reweighting actively repels the policy from suboptimal actions.
  arXiv  Detail & Related papers  (2026-03-10T22:01:13Z)
• DARC: Disagreement-Aware Alignment via Risk-Constrained Decoding [59.16244104797919]
  We propose **Disagreement-Aware Alignment via Risk-Constrained Decoding (DARC)**.<n>DARC frames response selection as distributionally robust, risk-sensitive decision making.<n>Experiments on alignment benchmarks show that DARC reduces disagreement and tail risk while maintaining competitive average quality.
  arXiv  Detail & Related papers  (2026-03-09T09:21:29Z)
• How RLHF Amplifies Sycophancy [23.213056717401418]
  Large language models often exhibit increased sycophantic behavior after preference-based post-training.<n>We identify an explicit amplification mechanism that causally links optimization against a learned reward to bias in the human preference data used for alignment.<n>We propose a training-time intervention designed to neutralize the amplification mechanism itself.
  arXiv  Detail & Related papers  (2026-02-01T03:46:14Z)
• Reference-Free Rating of LLM Responses via Latent Information [53.463883683503106]
  We study the common practice of asking a judge model to assign Likert-scale scores to free-text responses.<n>We then propose and evaluate Latent Judges, which derive scalar ratings from internal model signals.<n>Across a broad suite of pairwise and single-rating benchmarks, latent methods match or surpass standard prompting.
  arXiv  Detail & Related papers  (2025-09-29T12:15:52Z)
• Counterfactual Reward Model Training for Bias Mitigation in Multimodal Reinforcement Learning [0.5204229323525671]
  We present a counterfactual reward model that introduces causal inference with multimodal representation learning to provide an unsupervised, bias-resilient reward signal.<n>We evaluated the framework on a multimodal fake versus true news dataset, which exhibits framing bias, class imbalance, and distributional drift.<n>The resulting system achieved an accuracy of 89.12% in fake news detection, outperforming the baseline reward models.
  arXiv  Detail & Related papers  (2025-08-27T04:54:33Z)
• Proximalized Preference Optimization for Diverse Feedback Types: A Decomposed Perspective on DPO [19.5712961932773]
  We revisit direct preference optimization (DPO) and demonstrate that its loss theoretically admits a decomposed reformulation.<n>We introduce PRoximalized PReference Optimization (PRO), a unified method to align with diverse feeback types.
  arXiv  Detail & Related papers  (2025-05-29T10:23:22Z)
• Provably Mitigating Overoptimization in RLHF: Your SFT Loss is Implicitly an Adversarial Regularizer [52.09480867526656]
  We identify the source of misalignment as a form of distributional shift and uncertainty in learning human preferences.<n>To mitigate overoptimization, we first propose a theoretical algorithm that chooses the best policy for an adversarially chosen reward model.<n>Using the equivalence between reward models and the corresponding optimal policy, the algorithm features a simple objective that combines a preference optimization loss and a supervised learning loss.
  arXiv  Detail & Related papers  (2024-05-26T05:38:50Z)
• Where is the Grass Greener? Revisiting Generalized Policy Iteration for Offline Reinforcement Learning [81.15016852963676]
  We re-implement state-of-the-art baselines in the offline RL regime under a fair, unified, and highly factorized framework. We show that when a given baseline outperforms its competing counterparts on one end of the spectrum, it never does on the other end.
  arXiv  Detail & Related papers  (2021-07-03T11:00:56Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.