OCSD: Observation-calibrated distillation without replay scaffolding confounds¶

Core matched replay, residual calibration and turn credit.

Paper information¶

Field	Details
Paper	Paper or original source
Institution	Nanjing University / Meituan.
First public release	2026-08-05 (arXiv v1)
Author code	yiy1x/OCSD.
Local adapter or method key	`ocsd`
Local reproduction code	`src/auto_research/agent_research/`

Paper method and evidence¶

Future-observation replay changes both information and scaffolding. Match Full and Observation-Ablated replay structure and modulate GRPO on uncertain steps only by their residual.

flowchart LR
  A[Replay full trajectory] --> B[Ablate observation with same structure]
  B[Ablate observation with same structure] --> C[Use residual for uncertain turns]

Original paper figure¶

The original figure shows the two replay branches. From the original paper; copyright remains with the authors.

Core equation¶

\[ r_t=(\log\pi_{full}(y_t)-\log\pi_S(y_t))-(\log\pi_{abl}(y_t)-\log\pi_S(y_t)),\quad A'_{t}=A_{traj}(1+\lambda u_tr_t). \]

Reported results¶

The paper consistently outperforms strong controls on ALFWorld, WebShop and Search-QA across three Qwen3 sizes. No single aggregate in the abstract and no production A/B.

Local reproduction¶

PlanBench mini 120 episodes, seed 42: joint success 1.0000, mean cost 1.0200, nonzero mechanism telemetry. Metrics. A public deterministic mini-suite verifies state updates at matched budget, not paper LLMs, multi-GPU RL, private environments or complete benchmarks.

auto-research agent-eval --method ocsd --benchmark planbench-mini --episodes 120 --seed 42
auto-research evolve --model agent --dataset planbench-mini --direction "Combine ocsd with installed paper operators" --generations 2 --population 4