feat(daily): 2026-05-07 digest

[yayajjiang]

Papers added (2026-05-07)

• ResRL: Boosting LLM Reasoning via Negative Sample Projection Residual Reinforcement Learning (2605.00380) — SVD负样本投影解耦正负共享语义，12基准数学推理Avg@16提升9.4%、Pass@128提升7.0%，超越NSR基线。⭐ pick
• MoBiE: Efficient Inference of Mixture of Binary Experts under Post-Training Quantization (2604.06798) — 首个面向MoE LLM的二值化框架，联合SVD分解与梯度感知Hessian量化，压缩内存同时保持专家路由稳定性。
• Exploring Pass-Rate Reward in Reinforcement Learning for Code Generation (2605.02944) — 严格对照实验表明，无critic代码RL中pass-rate奖励并不稳定优于二元奖励，更密集奖励信号不能可靠引导梯度指向全通过解。

https://claude.ai/code/session_011xfrMRQbAqLkAB11AJMr5U

---

Generated by Claude Code