On-policy distillation papers report reasoning and agent gains

Late September 2026 papers on OASIS, AdviSD, Neighborhood OPSD, PivotOPD, and multi-teacher distillation report reasoning and agent gains, with disagreements on some units.

Between 28 September and 1 October 2026, separate papers reported on-policy distillation methods for LLM reasoning and multi-turn agents. OASIS supervises mostly label-verified on-policy trajectories with only final-answer labels and reports average gains of 3.2 to 3.8 points over base Qwen3-1.7B, 4B, and 8B models on AIME 2024, AIME 2025, and HMMT 2025, beating standard OPSD by 3.05 points at 8B. Neighborhood OPSD reports Average@12 gains over standard OPSD of 2.75, 1.67, and 1.94 points on those sizes across AIME 2024, AIME 2025, and HMMT February 2025. AdviSD's Qwen3-8B advisors, which steer frozen Gemini and Claude executors, beat advisor-GRPO by 4.2 to 6.4 on BFCL-v3 and 3.9 to 5.1 on EnvScaler, but sources disagree on whether those are points, percentage points, or score points. PivotOPD, attributed to NVIDIA in one report, is the strongest average among 13 baselines for Qwen3-1.7B and Qwen3-8B; sources agree more than half of failed Qwen3 rollouts from 8B to 235B contain an early pivotal mistake, but disagree on whether the 1.7B ALFWorld gain is 5.5 points or 5.5 percent and whether the Nemotron-3.5 SWE-Bench Verified gain is 3.2 points or 3.2 percent, with one account calling the ALFWorld figure a gain over the best baseline. A multi-teacher study of Qwen3-1.7B with four same-initialization RL teachers reports Adam update similarity of 0.96 between averaging rules versus 0.83 between teachers, about 97 percent of FP32 weights changed versus 7 to 11 percent in BF16, and math accuracy 2.6 points higher or 2.1 points lower by averaging rule, while one source's summary credits loss averaging for favoring longer responses and its bullet says token averaging.