Generative Reward Models
read the original abstract
Reinforcement Learning from Human Feedback (RLHF) has greatly improved the performance of modern Large Language Models (LLMs). The RLHF process is resource-intensive and technically challenging, generally requiring a large collection of human preference labels over model-generated outputs. Reinforcement Learning from AI Feedback (RLAIF) addresses this data collection challenge by leveraging synthetic preferences generated by an LLM. However, recent work has shown that synthetic preferences labels may not align well with human preference judgments. To address this, we propose a hybrid approach that unifies RLHF and RLAIF methodologies. We introduce GenRM, an iterative algorithm that trains an LLM on self-generated reasoning traces, leading to synthetic preference labels matching human preference judgments. Empirically, we show that zero-shot LLM-based judgments under-perform compared to Bradley-Terry reward models on in-distribution tasks (between 9-36%). In contrast, GenRM achieves in-distribution accuracy comparable to Bradley-Terry models, while significantly outperforming them on out-of-distribution tasks (between 10-45%). Moreover, GenRM surpasses the performance of using LLMs as judges on both in-distribution (by 9-31%) and out-of-distribution tasks (by 2- 6%). Our results show that combining the strengths of RLHF and RLAIF offers a promising approach for improving the quality of synthetic preference labels.
This paper has not been read by Pith yet.
Forward citations
Cited by 22 Pith papers
-
The Hidden Signal of Verifier Strictness: Controlling and Improving Step-Wise Verification via Selective Latent Steering
VerifySteer selectively steers hidden states at paragraph boundaries using latent correctness signals to control verifier strictness and outperform baselines on ProcessBench and Hard2Verify with lower compute.
-
RMGAP: Benchmarking the Generalization of Reward Models across Diverse Preferences
RMGAP benchmark shows state-of-the-art reward models reach at most 49.27% Best-of-N accuracy when forced to select responses matching diverse preferences.
-
Bringing Value Models Back: Generative Critics for Value Modeling in LLM Reinforcement Learning
GenAC introduces generative critics with chain-of-thought reasoning and in-context conditioning to improve value approximation and downstream RL performance in LLMs compared to value-based and value-free baselines.
-
Beyond Verifiable Rewards: Rubric-Based GRM for Reinforced Fine-Tuning SWE Agents
A rubric-based generative reward model improves reinforced fine-tuning of SWE agents by supplying richer behavioral guidance than binary terminal rewards alone.
-
Autodata: An agentic data scientist to create high quality synthetic data
Autodata introduces an agentic method with meta-optimization to create higher-quality synthetic data, yielding performance gains over standard methods on CS, legal, and math tasks.
-
Autodata: An agentic data scientist to create high quality synthetic data
Autodata trains meta-optimized AI agents to generate superior synthetic datasets, yielding performance gains over classical methods on CS research, legal, and math reasoning tasks.
-
RUBRIC-ARROW: Alternating Pointwise Rubric Reward Modeling for LLM Post-training in Non-verifiable Domains
RUBRIC-ARROW is an alternating rubric generator and judge framework that uses probability-based scoring and pairwise preferences to improve pointwise reward modeling accuracy for LLM post-training in non-verifiable domains.
-
RationalRewards: Reasoning Rewards Scale Visual Generation Both Training and Test Time
RationalRewards recovers rationales from preference data via PARROT to create a critique-first reward model that improves visual generators at both training time through RL and test time through prompt refinement, mat...
-
Utilizing and Calibrating Hindsight Process Rewards via Reinforcement with Mutual Information Self-Evaluation
MISE proves that hindsight self-evaluation rewards equal minimizing mutual information plus KL divergence to a proxy policy, and experiments show 7B LLMs reaching GPT-4o-level results on validation tasks.
-
ReflectRM: Boosting Generative Reward Models via Self-Reflection within a Unified Judgment Framework
ReflectRM improves generative reward models by adding self-reflection on analysis quality within a unified training setup for response and analysis preferences, yielding accuracy gains and reduced positional bias on b...
-
UniCreative: Unifying Long-form Logic and Short-form Sparkle via Reference-Free Reinforcement Learning
UniCreative uses reference-free RL with an adaptive constraint-aware reward model to unify long-form coherence and short-form creativity in AI writing, producing an emergent ability to switch between planning and dire...
-
PaTaRM: Bridging Pairwise and Pointwise Signals via Preference-Aware Task-Adaptive Reward Modeling
PaTaRM converts pairwise preference data into pointwise reward signals via a novel PAR mechanism and task-adaptive rubrics, reporting 8.7% gains on RewardBench/RMBench and 13.6% relative RLHF improvement.
-
Learning to Refine: Self-Refinement of Parallel Reasoning in LLMs
GSR jointly trains LLMs to generate candidate solutions and refine a superior final answer from them, achieving state-of-the-art performance on five mathematical benchmarks while transferring across model scales.
-
RewardBench 2: Advancing Reward Model Evaluation
RewardBench 2 is a new benchmark that supplies challenging fresh human prompts for reward model evaluation, yielding lower average scores but higher correlation with downstream best-of-N sampling and RLHF training per...
-
REAR: Test-time Preference Realignment through Reward Decomposition
REAR decomposes the reward into question and preference components, rescales their balance, and expresses the result as a linear combination of token log-probabilities for efficient integration with best-of-N and tree search.
-
When RL Fails after SFT: Rejuvenating Model Plasticity for Robust SFT-to-RL Handoff
Excessive SFT reduces LLM plasticity for RL; Rejuvenation restores it via base-anchored fusion and targeted neuron resets, yielding better RL performance and OOD generalization.
-
Trust Region On-Policy Distillation
TrOPD stabilizes on-policy distillation for LLMs with trust-region learning, outlier estimation, and off-policy guidance, outperforming prior OPD methods on reasoning and code benchmarks.
-
CARE-RL: Capability-Aware Reinforcement Learning for Mitigating Cross-Domain Conflicts
CARE-RL combines PA-GRM for task-adaptive rewards on open-ended tasks and DACSP for modulating RL updates using historical capability directions, reporting higher total average scores than baselines on Qwen models.
-
Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges
The paper introduces the Proxy Compression Hypothesis as a unifying framework explaining reward hacking in RLHF as an emergent result of compressing high-dimensional human objectives into proxy reward signals under op...
-
ConsistRM: Improving Generative Reward Models via Consistency-Aware Self-Training
ConsistRM improves generative reward models via consistency-aware self-training, outperforming vanilla RFT by 1.5% on average across five benchmarks and four base models.
-
Scaling Test-Time Compute to Achieve IOI Gold Medal with Open-Weight Models
GenCluster scales test-time compute via large-scale generation, behavioral clustering, ranking, and round-robin submission to achieve IOI gold medal performance with the open-weight gpt-oss-120b model.
-
Seed1.5-VL Technical Report
Seed1.5-VL is a compact multimodal model that sets new records on dozens of vision-language benchmarks and outperforms prior systems on agent-style tasks.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.