SEED: Self-evolving on-policy distillation¶

Core state and update mechanism on a deterministic mini-suite, not paper-scale training.

Paper information¶

Field	Details
Paper	SEED (arXiv 2607.14777)
Institution	Tsinghua University / Zhejiang University / CUHK / NTU / Tongji University.
First public release	2026-07-16
Author code	jinyangwu/SEED.
Local adapter or method key	`seed`
Local reproduction code	`src/auto_research/agent_research/`

Paper method and evidence¶

Reflect on completed trajectories to extract reusable hindsight skills. Compare action log probabilities with and without skill context to supply dense on-policy distillation credit.

flowchart LR
  T[Completed trajectories] --> S[Hindsight skills]
  S --> C[Contrast skill-conditioned action probabilities]
  C --> D[Dense on-policy credit]

Original paper figure¶

The original figure shows extracting and distilling hindsight skills. From the original paper; copyright remains with the authors.

Core equation¶

\[ r_t^{\mathrm{skill}}=\log\pi_\theta(a_t\mid s_t,z)-\log\pi_\theta(a_t\mid s_t),\quad \mathcal L=\mathcal L_{\mathrm{RL}}+\lambda\mathcal L_{\mathrm{OPD}}. \]

Reported results¶

The paper reports consistent capability and sample-efficiency gains in text and visual agent tasks and evaluates transfer to unseen settings. No production A/B.

Local reproduction¶

Compress success/failure traces into structured skills in PlanBench mini and record reuse across episodes. Joint success 1.0000, mean cost 0.9800, hindsight skills / dense credit updates 12 / 360.

auto-research agent-eval --method seed --episodes 120 --seed 42

Fixed metrics. No visual model or large-scale policy gradient training.