Pith. sign in

REVIEW 7 cited by

WARM: On the Benefits of Weight Averaged Reward Models

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 2401.12187 v1 pith:OHYL4NLG submitted 2024-01-22 cs.LG cs.AIcs.CL

classification cs.LGcs.AIcs.CL
keywords rewardwarmfine-tunedmodelsweightaveragedaveragingdistribution
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Aligning large language models (LLMs) with human preferences through reinforcement learning (RLHF) can lead to reward hacking, where LLMs exploit failures in the reward model (RM) to achieve seemingly high rewards without meeting the underlying objectives. We identify two primary challenges when designing RMs to mitigate reward hacking: distribution shifts during the RL process and inconsistencies in human preferences. As a solution, we propose Weight Averaged Reward Models (WARM), first fine-tuning multiple RMs, then averaging them in the weight space. This strategy follows the observation that fine-tuned weights remain linearly mode connected when sharing the same pre-training. By averaging weights, WARM improves efficiency compared to the traditional ensembling of predictions, while improving reliability under distribution shifts and robustness to preference inconsistencies. Our experiments on summarization tasks, using best-of-N and RL methods, shows that WARM improves the overall quality and alignment of LLM predictions; for example, a policy RL fine-tuned with WARM has a 79.4% win rate against a policy RL fine-tuned with a single RM.

Discussion (0). Sign in to comment.

Forward citations

Cited by 7 Pith papers

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

  1. Don't Mix Rewards, Mix Policies: Policy Decomposition and Optimization for Multi-Reward RL

    cs.AI 2026-07 conditional novelty 6.0 of 10

    PRISM trains one positive policy per reward plus one global negative policy and merges their token logits, improving multi-reward RL for LLMs with inference-time controllability.

  2. Multi-Turn On-Policy Distillation with Prefix Replay

    cs.LG 2026-07 conditional novelty 6.0 of 10

    ReOPD offline-distills multi-turn agentic LLMs via teacher-prefix replay plus step-decay sampling, matching online OPD accuracy at ≥4× speed with zero tool calls.

  3. Off-Policy Corrected Reward Modeling for Reinforcement Learning from Human Feedback

    cs.LG 2025-07 conditional novelty 6.0 of 10

    Applying importance weighting to reward model training to correct for policy distribution shift in RLHF improves final policy quality without new labels.

  4. Tiny Reward Models

    cs.CL 2025-07 conditional novelty 6.0 of 10

    TinyRM shows that 400M-parameter bidirectional masked language models, tuned with FLAN-style prompting, DoRA, and layer freezing, outperform a 70B reward model on RewardBench reasoning and come close on safety.

  5. Bradley-Terry and Multi-Objective Reward Modeling Are Complementary

    cs.LG 2025-07 conditional novelty 6.0 of 10

    Jointly training a Bradley-Terry preference head and a multi-attribute regression head on a shared embedding improves reward-model robustness to reward hacking and boosts multi-objective scoring performance.

  6. Learning a Pessimistic Reward Model in RLHF

    cs.LG 2025-05 reject novelty 6.0 of 10

    Pessimistic fine-tuning of reward models against rejection-sampling policies lets RLHF agents optimize greedily without KL regularization and still avoid reward hacking.

  7. Towards Reliable, Uncertainty-Aware Alignment

    cs.LG 2025-07 reject novelty 4.0 of 10

    Variance-aware RLHF adds a variance-weighted KL penalty to PPO and reduces reward variance and the risk of underperforming the reference policy in the paper's experiments.

Pith tools