Clue-OPSD: privileged-view distillation from long-video clues

简体中文

Core-mechanism mini-suite. Executes the defining control flow and objective. Deterministic task success does not reproduce the paper's benchmarks.

Paper information

Field	Details
Paper	arXiv 2608.25356
Organization	University of Maryland, College Park / Johns Hopkins University (first author's first affiliation in the source)
First public release	2026-08-26 (arXiv v1)
Authors' code	No public author code was found on 2026-08-28.
Adapter / method	clue-opsd
Local reproduction code	src/auto_research/post_training/latest_20260827.py

Original paper

Motivation and method

The student sees the full long video. During training a frozen teacher sees only the question-relevant clue interval. On-policy self-distillation uses the student's own rollouts; inference needs no clues, labels, or external teacher.

flowchart LR
  I[Task and current state] --> M[Privileged-clue teacher, on-policy trajectory, full-context student]
  M --> O[Policy action or training objective]
  O --> E[Auditable metrics and feedback]
  E --> M

Key figure from the paper

Figure 2: Overall workflow, stages, and data flow. From the original paper; copyright remains with the authors.

Core equation

\[ \mathcal L_{OPSD}=\mathbb E_{y\sim\pi_\theta(\cdot|v,q)}\sum_t D_{KL}(\pi_T(\cdot|c,q,y_{<t})\Vert\pi_\theta(\cdot|v,q,y_{<t})). \]

Paper results

Qwen3.5-2B improves 6.94/3.45/9.81/3.41 points on MLVU/LVBench/LongVideoBench/MMVU, respectively. No industrial online A/B is reported.

Local reproduction

Local comparison: Arithmetic-smoke at 100 steps raises accuracy 0.1250 → 0.5938 (+375.00%); a process-reward proxy substitutes for clues.

See the metrics and batch index.

auto-research post-train --algorithm clue-opsd --dataset arithmetic-smoke --steps 100 --seed 42 --offline

Reproduction boundaries

The public deterministic candidate/plan tasks isolate the mechanism. The paper checkpoint, private training data, external services, and full benchmarks are not reproduced.