ResRL decouples shared semantics between positive and negative responses in LLM reinforcement learning via SVD-based projection residuals, outperforming baselines including NSR by up to 9.4% on math reasoning benchmarks.
InThe Twelfth Inter- national Conference on Learning Representations
3 Pith papers cite this work. Polarity classification is still indexing.
3
Pith papers citing it
citation-role summary
background 1
citation-polarity summary
years
2026 3roles
background 1polarities
unclear 1representative citing papers
Rewarding an LLM against mid-level skill-prototype rubrics during RL improves AIME25 math accuracy from 43.3% to 63.3% on a 4B model and boosts multi-turn tool-use success.
RLAVR uses the Corrective Advantage Gap metric and CARE policy to actively acquire ground-truth labels for key samples, stabilizing RLVR training and boosting performance with limited annotation budgets.
citing papers explorer
-
ResRL: Boosting LLM Reasoning via Negative Sample Projection Residual Reinforcement Learning
ResRL decouples shared semantics between positive and negative responses in LLM reinforcement learning via SVD-based projection residuals, outperforming baselines including NSR by up to 9.4% on math reasoning benchmarks.
-
SCRIBE: Structured Mid-Level Supervision for Tool-Using Language Models
Rewarding an LLM against mid-level skill-prototype rubrics during RL improves AIME25 math accuracy from 43.3% to 63.3% on a 4B model and boosts multi-turn tool-use success.
-
When Self-Belief Misleads: Active Label Acquisition for Reinforcement Learning with Verifiable Rewards
RLAVR uses the Corrective Advantage Gap metric and CARE policy to actively acquire ground-truth labels for key samples, stabilizing RLVR training and boosting performance with limited annotation budgets.