Rubric Dropout: A Simple Way to Mitigate Reward Hacking in Rubric-as-Reward RL

• URL: http://arxiv.org/abs/2608.11669v1
• Date: Wed, 12 Aug 2026 05:29:21 GMT
• Title: Rubric Dropout: A Simple Way to Mitigate Reward Hacking in Rubric-as-Reward RL
• Abstract summary: We find that the two scores diverge during training.<n>A judge with a fixed bias would shift the gold curve by a constant, not send it down while the training score rises.<n>We propose Dropout, a one-line fix borrowed from neuron dropout.
• Score: 11.857468467125505
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Reinforcement learning against rubrics, lists of criteria graded by an LLM judge, has become a standard way to post-train language models on tasks with no deterministic answer. The rubric, however, is a fixed proxy for quality, never a complete description of it, and a policy trained against it long enough will learn to exploit the difference. We measure this directly. Training Qwen3-8B with Group Relative Policy Optimization (GRPO) on medical and science rubrics and grading out-of-distribution (OOD) benchmarks with both the training judge and a stronger gold judge, we find that the two scores diverge during training. The training judge's score keeps climbing while the gold judge's score peaks and then falls, by 3 points on HealthBench-Hard and by 22 points on ResearchQA. A judge with a fixed bias would shift the gold curve by a constant, not send it down while the training score rises, so the divergence is reward hacking, not judge noise. We propose Rubric Dropout, a one-line fix borrowed from neuron dropout. At every step, we randomly drop a subset of the rubric's criteria before computing the reward, so the policy never optimizes the same rubric twice. The dropped subset is shared across each rollout group, so GRPO's group-relative advantages stay comparable, and evaluation always uses the full rubric. Comparing no dropout against dropout at 30% and 50% on both benchmark pairs, dropout raises the OOD gold score at every matched checkpoint (+1 to +2 points on HealthBench-Hard, +6 to +7 points on ResearchQA), lowers the two hacking measures we track, and costs nothing in domain. Sweeping the dropout fraction shows a broad 30-50% sweet spot, while the natural alternative, reweighting criteria by how useful they are to training, performs worse than no intervention at all in our setting.

Related papers

• RecurSE: Bounded Recursive Self-Evaluation for LLM Rubric Judges [48.831446371402414]
  RecurSE pairs a trainable judge with a synchronized policy-copy checker that audits the judge's reasoning against meta-rubrics to supply a scalar process reward.<n>We show that RecurSE achieves consistent generalization gains across held-out medical, pairwise, summarization, and professional benchmarks.
  arXiv  Detail & Related papers  (2026-08-25T08:35:04Z)
• CoRE: Consensus Rewards via Equilibrium for Test-Time Reinforcement Learning [6.071121358322323]
  emphCoRE improves the untrained base by $+21.7$ points on average versus $+20.4$ for majority-vote TTRL.<n>Treating the roll-out group as a graph rather than a ballot box turns a brittle vote into a calibrated, graded, self-supervised reward at no extra roll-out cost.
  arXiv  Detail & Related papers  (2026-08-10T09:06:03Z)
• Reasoning Arena: Trace Tournaments When Verifiable Rewards Fall Short [51.667769734342635]
  Reinforcement learning with verifiable rewards (RLVR) has become a leading paradigm for improving the reasoning ability of large language models.<n>We propose Reasoning Arena, an adaptive training framework that routes such non-diverse reward groups to a judge system instead of discarding them.<n>We show that Reasoning Arena consistently outperforms the RLVR baseline by 7.6% on average in competition mathematics and coding benchmarks.
  arXiv  Detail & Related papers  (2026-06-08T11:57:17Z)
• ARBOR: Online Process Rewards via a Reusable Rubric Buffer for Search Agents [48.80766702702854]
  LLM-based search agents are trained predominantly with outcome-only reward, leaving the search process itself unsupervised.<n>This signal degenerates on outcome-homogeneous groups where all sampled trajectories share the same correctness, yielding zero within-group advantage and no gradient.<n>We propose ARBOR (Adaptive Buffer for Online Reward), a reusable process-reward framework that maintains a rubric memory shared across queries.
  arXiv  Detail & Related papers  (2026-06-02T06:58:54Z)
• When RLHF Fails: A Mechanistic Taxonomy of Reward Hacking, Collapse, and Evaluator Gaming [0.0]
  We present an empirical failure-mode study of a compact RLHF pipeline with proximal policy optimization (PPO)<n>We classify matched transitions between checkpoints using the directions of the learned reward, judge scores, and average judge score.<n>A pre-transition logistic model predicts future row-level reward hacking with ROC-AUC 0.821, and row-level analysis finds localized reward hacking that checkpoint averages miss in 3 of 12 settings.
  arXiv  Detail & Related papers  (2026-06-02T06:55:52Z)
• Turning Bias into Bugs: Bandit-Guided Style Manipulation Attacks on LLM Judges [65.92623263645139]
  We introduce BITE, a black-box adversarial framework that learns semantics-preserving edits to mislead an LLM judge.<n>BITE achieves an attack success rate exceeding 65% and raises scores by 1-2 points on a 9-point scale.<n>Our findings expose a fundamental weakness in the LLM-as-a-judge paradigm and motivate robust, attack-aware evaluation.
  arXiv  Detail & Related papers  (2026-05-24T05:24:09Z)
• Not Every Rubric Teaches Equally: Policy-Aware Rubric Rewards for RLVR [15.763726386280723]
  POW3R is a policy-aware reward framework that preserves human weights and category balance as the objective.<n> POW3R uses rollout-level contrast to emphasize criteria that currently separate the policy's outputs.
  arXiv  Detail & Related papers  (2026-05-19T17:50:18Z)
• Step-wise Rubric Rewards for LLM Reasoning [72.17879367869503]
  Reinforcement Learning with Verifiable Rewards (RLVR) is widely used to improve reasoning in large language models.<n>We find 18.2% of steps in correct responses are wrong yet positively rewarded.<n>Across six mathematical reasoning benchmarks, SRaR improves average accuracy over RaR by 3.57 points.
  arXiv  Detail & Related papers  (2026-05-17T07:08:14Z)
• Reinforcement Learning with Verifiable yet Noisy Rewards under Imperfect Verifiers [90.50039419576807]
  Reinforcement Learning with Verifiable Rewards (RLVR) trains policies against automated verifiers to avoid costly human labeling.<n>To reduce vulnerability to verifier hacking, many RLVR systems collapse rewards to binary $0,1$ during training.<n>This choice carries a cost: it introduces textitfalse negatives (rejecting correct answers, FNs) and textitfalse positives (accepting incorrect ones, FPs)
  arXiv  Detail & Related papers  (2025-10-01T13:56:44Z)
• Compute as Teacher: Turning Inference Compute Into Reference-Free Supervision [26.922922043969958]
  We propose turning exploration into supervision through Compute as Teacher (CaT)<n>CaT synthesizes a single reference from a group of parallel rollouts and then optimizing toward it.<n>As a test-time procedure, CaT improves Gemma 3 4B, Qwen 3 4B, and Llama 3.1 8B.
  arXiv  Detail & Related papers  (2025-09-17T17:59:42Z)
• Rubrics as Rewards: Reinforcement Learning Beyond Verifiable Domains [9.917318870162365]
  Reinforcement Learning with Verifiable Rewards (RLVR) has proven effective for complex reasoning tasks with clear correctness signals such as math and coding.<n>Rubrics have recently been used in evaluation benchmarks to capture such judgments, but their potential as reward signals for on-policy post-training remains underexplored.<n>We introduce RaR, an on-policy reinforcement learning method that extends RLVR beyond verifiable domains by using rubric-based feedback.
  arXiv  Detail & Related papers  (2025-07-23T17:57:55Z)
• Large Language Models are not Fair Evaluators [60.27164804083752]
  We find that the quality ranking of candidate responses can be easily hacked by altering their order of appearance in the context. This manipulation allows us to skew the evaluation result, making one model appear considerably superior to the other. We propose a framework with three simple yet effective strategies to mitigate this issue.
  arXiv  Detail & Related papers  (2023-05-29T07:41:03Z)

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.