Research map / Reinforcement learning / RL from feedback
SIRAS: Sibling-Relative Advantage Shaping for Reinforcement Learning from Verifiable Rewards
Zairun Yang, Jun Xu, Lei Liang, Huajun Chen, Qiang Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SIRAS, a sibling-relative advantage shaping method for RLVR that uses same-prompt rollouts to refine token-level advantages and extract signal from zero-advantage groups.
Problem
GRPO-style RLVR gives every token in a rollout the same response-level advantage and discards groups where all rollouts succeed or fail (zero group-relative advantage), underusing training signal.
Approach
Exploits the observation that rollouts sampled from the same prompt are not independent: sibling rollouts are used to shape advantages at the token level rather than assigning one response-level advantage, and to obtain learning signal from groups with zero group-relative advantage.
Experimental setup
Baselines: GRPO
Builds on: GRPO RLVR
Closest papers by meaning
- Conditional Advantage Estimation for Reinforcement Learning in Large Reasoning Models (ICLR 2026 · similarity 0.89)
- From Verifiable Dot to Reward Chain: Harnessing Verifiable Reference-based Rewards for Reinforcement Learning of Open-ended Generation (ICLR 2026 · similarity 0.89)
- BREAD: Branched Rollouts from Expert Anchors Bridge SFT & RL for Reasoning (NeurIPS 2025 · similarity 0.89)
- DUET: Optimize Token-Budget Allocation for Reinforcement Learning with Verifiable Rewards (NeurIPS 2026 · similarity 0.89)
- The Reasoning Boundary Paradox: How Reinforcement Learning Constrains Language Models (NeurIPS 2026 · similarity 0.88)
- Your Language Model is Its Own Critic: Reinforcement Learning with Value Estimation from Actor’s Internal States (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
