Video-OPSD: Privileged self-distillation with evidence frames¶

Core mechanism mini-suite. Implements an evidence-view self-teacher and evidence-guided token weights.

Paper information¶

Field	Details
Paper	arXiv 2608.27065
Institution	Nanyang Technological University (first listed affiliation of the first author)
First public release	2026-08-27 (arXiv v1)
Author code	No public author code found on 2026-08-31.
Local adapter/method	`video-opsd`
Local reproduction code	`src/auto_research/post_training/latest_20260831.py`

Original paper summary¶

Background and main changes¶

The student reads the full video while the training-time self-teacher sees only human-labeled evidence frames. Distillation is weighted by each token's dependence on that evidence. The paper reports performance close to GRPO with approximately 60% less training time.

flowchart LR
  V[Full video] --> S[Student rollout]
  F[Evidence frames] --> T[Privileged self-teacher]
  T --> W[Evidence-dependent token weights]
  S --> W
  W --> U[Policy update]

Key figure from the paper¶

Figure 1: Overall design and components. From the paper; copyright belongs to the authors. Click for the source.

Core equation¶

\[ \mathcal L_{EGO}=\sum_t w_t^{evidence}D_{KL}(p_T^t\Vert p_\theta^t). \]

Local reproduction¶

Arithmetic-smoke over 100 steps reached 0.1953 → 0.6562 accuracy, recording evidence-token weights and the teacher gap. See metrics.

Scope¶

The public smoke data has no frames. A separate reward dimension simulates evidence dependence. No video LLM was trained, and this concept check is not the paper's measured result.

Actual video checkpoint path¶

auto-research video-opsd-eval compares full videos against privileged views from evidence_frame_indices on public Video-MME-v2-compatible inputs. Both arms share the pinned SmolVLM2 checkpoint and decoding budget. It reports three-seed accuracy, parse rate and answer agreement.

auto-research video-opsd-eval --annotations video-mme-v2-evidence.jsonl \
  --video-root videos --seeds 42,43,44

Per-question evidence annotations are mandatory; missing annotations cause failure. This audits the public checkpoint's evidence views, not the paper's full OPSD training with 6,500 examples and 8×H100. A100 acceptance used official demo clips from a pinned Video-MME-v2 revision. Repository frame labels are additional protocol annotations, not the authors' OPSD annotations. See Video-OPSD A100 receipt.