technique

Hybrid Reward Advantage Splitting

techniqueactiveprovisionalhybrid-reward-advantage-splitting-c849668c·1 events·first seen 2d ago

Aliases: Hybrid Reward Advantage Splitting

Co-occurring entities

CORA GRPO (Group Relative Policy Optimization)

More like this (12)

RREDCoT: Segment-Level Reward Redistribution for Reasoning Models reward model reward hacking cluster voting reward CapReward Gradient-Guided Reward Optimization In-Context Reward Adaptation hybrid reasoning Scaling Laws for Reward Model Overoptimization Process Reward Model Rule-Based Rewards Using Reward Uncertainty to Induce Diverse Behaviour in Reinforcement Learning

Recent events (1)

5arXiv · cs.CL·2d ago·source ↗

CORA: Consistency-Oriented Reasoning Alignment addresses thinking-answer gap in multimodal RLVR

Researchers identify and analyze a systematic inconsistency between reasoning traces and final answers in RLVR-trained large vision-language models, showing the problem persists throughout GRPO training and inference. They propose CORA, which introduces a lightweight plug-and-play consistency reward model and a Hybrid Reward Advantage Splitting (HRAS) mechanism to coordinate task and consistency optimization. Experiments across multimodal reasoning benchmarks show CORA improves both task performance and reasoning faithfulness.

Evaluation and Benchmarking Alignment and RLHF CORA Hybrid Reward Advantage Splitting GRPO (Group Relative Policy Optimization)+1 more