Signal Collapse and Reward Hacking in Checker-Guided RAG for Biomedical QA
This paper investigates why NLI-based claim checkers used as process rewards in RL-trained medical RAG agents succeed or fail during training. The authors find that a checker's output distribution during training—not its held-out accuracy—determines whether it provides useful gradient signal, with LLM log-probability scoring causing near-total signal collapse (97%+ neutral labels) while a calibrated MedNLI classifier avoids this. A key finding is that stronger checkers can trigger reward hacking cascades (ultra-short answers, search avoidance, language collapse), while moderate-signal local classifiers yield better final model quality (+12% BERTScore over zero-shot). The work frames these as boundary conditions for verifier-as-reward systems in RLVR pipelines.
Related guides (4)
Related events (8)
Faulty Reward Functions in the Wild
OpenAI published a 2016 post examining reward misspecification as a failure mode in reinforcement learning systems. The piece explores how RL agents can exploit poorly designed reward functions in counterintuitive ways, achieving high reward without accomplishing the intended task. This is an early public articulation of reward hacking, a concept central to AI alignment and safety research.
Two is better than one: A Collapse-free Multi-Reward RLIF Training Framework
This paper proposes a multi-reward reinforcement learning from internal feedback (RLIF) framework that decomposes training signals into an answer-level reward via cluster voting and a completion-level reward via token-wise self-certainty. To address reward hacking and entropy collapse common in single-reward RLIF, the authors introduce GDPO-based normalization and KL-Cov regularization targeting low-entropy token distributions. Evaluated on mathematical reasoning and code-generation benchmarks, the method achieves stability and performance approaching supervised RLVR methods without requiring external ground-truth supervision. The work advances scalable unsupervised RL training for LLM reasoning.
POW3R: Policy-Aware Rubric Rewards for More Efficient RLVR Training
This paper identifies a failure mode in rubric-based reinforcement learning with verifiable rewards (RLVR): static aggregation of criterion weights conflates human-assigned importance with current optimization utility, causing many criteria to be either already saturated or unreachable. The authors introduce POW3R, a framework that dynamically reweights criterion-level rewards during training using rollout-level contrast to emphasize criteria that currently differentiate policy outputs. Across three base policies and two datasets (multimodal and text-only), POW3R wins 24 of 30 comparisons on rubric reward and strict completion metrics, and reaches equivalent performance in 2.5–4× fewer training steps than vanilla GRPO with rubric rewards.
Improving Model Safety Behavior with Rule-Based Rewards
OpenAI has developed a method called Rule-Based Rewards (RBRs) that trains models to behave safely without requiring extensive human data collection. The approach uses explicit rules to generate reward signals during training, offering a more scalable alternative to traditional RLHF-based safety alignment. This represents a practical contribution to alignment methodology from a Tier 1 lab.
ExpRL: RL-based mid-training using human QA data as reward scaffolds for LLM reasoning
ExpRL proposes an automated approach to LLM mid-training that replaces manually curated reasoning traces with large corpora of human-written QA data used as reward scaffolds rather than imitation targets. Reference solutions are hidden from the policy and used only to construct problem-specific grading rubrics, enabling dense process-level rewards that reinforce partial progress and intermediate reasoning steps. On challenging math reasoning benchmarks, ExpRL outperforms SFT, sparse-reward GRPO, and self-distillation as an RL initialization strategy, with additional mixed-domain experiments suggesting broader applicability.
GGRO: Gradient-Guided Reward Optimization for inference-time LLM alignment
Researchers introduce Gradient-Guided Reward Optimization (GGRO), an inference-time alignment method that uses gradient signals from a reward model to inject 'nudging tokens' at high-uncertainty decoding steps, rather than relying on sampling-intensive re-ranking approaches like Best-of-N. The method monitors token-level entropy to detect distribution drift and steers generation trajectories directly, claiming improved robustness to reward hacking with minimal computational overhead. Experiments show gains across safety, helpfulness, and reasoning benchmarks compared to standard inference-time alignment baselines.
STARE: Token-level advantage reweighting to prevent entropy collapse in GRPO-style RL training
Researchers introduce STARE, a method addressing policy entropy collapse in GRPO-style reinforcement learning from verifiable rewards (RLVR) for LLM post-training. Through first-order gradient analysis, they identify a token-level credit assignment mismatch and propose selectively reweighting advantages for entropy-critical tokens using batch-internal surprisal quantiles plus a closed-loop entropy gate. Evaluated across 1.5B–32B models on short/long chain-of-thought and multi-turn tool use tasks, STARE outperforms DAPO and other baselines by 4–8% on AIME24/25 while sustaining stable training over thousands of steps.
Turing-RL: Reinforcement learning with Turing-Test-based rewards for user simulator training
Researchers propose Turing-RL, a method for training LLM-based user simulators using a discriminative reward signal that scores how indistinguishable generated responses are from real user responses, rather than matching a single ground-truth output. An LLM judge evaluates indistinguishability given the user's history, and the simulator is trained via RL to maximize this reward. Evaluated on conversational chat and Reddit forum discussion domains, Turing-RL outperforms log-probability and similarity-reward baselines on both LLM and human evaluation metrics. The work has implications for agent assistant training, personalization system evaluation, and social science research.