Pith. sign in

REVIEW 19 cited by

Reward Model Ensembles Help Mitigate Overoptimization

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 2310.02743 v2 pith:WCYOYPRQ submitted 2023-10-04 cs.LG

Reward Model Ensembles Help Mitigate Overoptimization

classification cs.LG
keywords optimizationrewardoveroptimizationmodelconservativeensemble-basedhumanmodels
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
read the original abstract

Reinforcement learning from human feedback (RLHF) is a standard approach for fine-tuning large language models to follow instructions. As part of this process, learned reward models are used to approximately model human preferences. However, as imperfect representations of the "true" reward, these learned reward models are susceptible to overoptimization. Gao et al. (2023) studied this phenomenon in a synthetic human feedback setup with a significantly larger "gold" reward model acting as the true reward (instead of humans) and showed that overoptimization remains a persistent problem regardless of the size of the proxy reward model and training data used. Using a similar setup, we conduct a systematic study to evaluate the efficacy of using ensemble-based conservative optimization objectives, specifically worst-case optimization (WCO) and uncertainty-weighted optimization (UWO), for mitigating reward model overoptimization when using two optimization methods: (a) best-of-n sampling (BoN) (b) proximal policy optimization (PPO). We additionally extend the setup of Gao et al. (2023) to include 25% label noise to better mirror real-world conditions. Both with and without label noise, we find that conservative optimization practically eliminates overoptimization and improves performance by up to 70% for BoN sampling. For PPO, ensemble-based conservative optimization always reduces overoptimization and outperforms single reward model optimization. Moreover, combining it with a small KL penalty successfully prevents overoptimization at no performance cost. Overall, our results demonstrate that ensemble-based conservative optimization can effectively counter overoptimization.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 19 Pith papers

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

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

    cs.LG 2026-07 accept novelty 7.0

    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. Variance-aware Reward Modeling with Anchor Guidance

    stat.ML 2026-05 unverdicted novelty 7.0

    Anchor-guided variance-aware reward modeling uses two response-level anchors to resolve non-identifiability in Gaussian models of pluralistic preferences, yielding provable identification, a joint training objective, ...

  3. TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment

    cs.LG 2026-05 unverdicted novelty 7.0

    TMPO uses Softmax Trajectory Balance to match policy probabilities over multiple trajectories to a Boltzmann reward distribution, improving diversity by 9.1% in diffusion alignment tasks.

  4. TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment

    cs.LG 2026-05 unverdicted novelty 7.0

    TMPO replaces scalar reward maximization with trajectory-level matching to a Boltzmann distribution via Softmax-TB, improving generative diversity by 9.1% while keeping competitive reward performance.

  5. Theoretical Limits of Language Model Alignment

    cs.LG 2026-05 unverdicted novelty 7.0

    The maximum reward gain under KL-regularized LM alignment is a Jeffreys divergence term, estimable as covariance from base samples, with best-of-N approaching the theoretical limit.

  6. Reinforcement Learning via Value Gradient Flow

    cs.LG 2026-04 unverdicted novelty 7.0

    VGF solves behavior-regularized RL by transporting particles from a reference distribution to the value-induced optimal policy via discrete value-guided gradient flow.

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

    cs.LG 2026-07 conditional novelty 6.0

    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.

  8. STEAM: Self-Supervised Temporal Ensemble Advantage Modeling for Real-World Robot Learning

    cs.RO 2026-06 unverdicted novelty 6.0

    STEAM learns advantages from expert trajectories via self-supervised temporal ensemble modeling to improve policy learning on real robot tasks like bimanual folding and pick-and-place.

  9. Uncertainty-Aware Reward Modeling for Stable RLHF

    cs.LG 2026-06 unverdicted novelty 6.0

    UARM equips reward models with quantile-based conformal prediction uncertainty and reweights GRPO advantages via heteroscedastic variance decomposition to improve calibration and reduce reward hacking in RLHF.

  10. Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs

    cs.AI 2026-05 unverdicted novelty 6.0

    MOOD benchmark shows guard models fail to generalize to OOD alignment failures in LLMs, but combining them with Mahalanobis and perplexity OOD detectors improves recall from 39% to 45% with better scaling than larger ...

  11. Benchmarking and Improving Monitors for Out-Of-Distribution Alignment Failure in LLMs

    cs.AI 2026-05 conditional novelty 6.0

    Introduces MOOD benchmark for OOD LLM alignment failures and shows guard models plus Mahalanobis and perplexity OOD detectors improve recall from 39% to 45% with positive scaling.

  12. FUSE: Ensembling Verifiers with Zero Labeled Data

    stat.ML 2026-04 unverdicted novelty 6.0

    FUSE ensembles verifiers unsupervisedly by controlling their conditional dependencies to improve spectral ensembling algorithms, matching or exceeding semi-supervised baselines on benchmarks including GPQA Diamond and...

  13. Provably avoiding over-optimization in Direct Preference Optimization without knowing the data distribution

    cs.LG 2026-02 unverdicted novelty 6.0

    PEPO uses pessimistic ensembling of DPO policies on data subsets to achieve single-policy concentrability sample bounds and avoid over-optimization in tabular settings.

  14. Factored Causal Representation Learning for Robust Reward Modeling in RLHF

    cs.LG 2026-01 unverdicted novelty 6.0

    A factored causal representation learning method improves robustness of reward models in RLHF by isolating causal factors from biases like length and sycophancy using adversarial gradient reversal.

  15. Designing Reward Signals for Portable Query Generation: A Case Study in Industrial Semantic Job Search

    cs.LG 2026-06 unverdicted novelty 5.0

    Empirical study of RLAIF for portable query generation finds reward shaping controls performance more than optimizer choice and a rule-based reward floor yields +0.147 quality gain.

  16. Quantifying the Utility of User Simulators for Building Collaborative LLM Assistants

    cs.CL 2026-05 unverdicted novelty 5.0

    Fine-tuned simulators grounded in real human data produce LLM assistants that win more often against real users than those trained against role-playing simulators.

  17. Provably avoiding over-optimization in Direct Preference Optimization without knowing the data distribution

    cs.LG 2026-02 unverdicted novelty 5.0

    PEPO is a single-step pessimistic ensemble algorithm for direct preference optimization that provably avoids over-optimization by depending only on single-policy concentrability without knowing the data distribution o...

  18. Proxy Reward Internalization and Mechanistic Exploitation: A Learned Precursor to Reward Hacking and Its Generalization

    cs.AI 2026-06 unverdicted novelty 4.0

    Proxy RL produces a staged proxy-internalization capability that emerges before and predicts reward hacking in coding environments.

  19. Against Proxy Optimization

    cs.AI 2026-06 unverdicted novelty 2.0

    Discusses conditions under which maximizing proxy utilities is harmful and suggests problems for decision theory applications.