Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes¶

L1 mechanism diagnostic. Spatially masked teacher KL, without scene generation or an MLLM.

Paper information¶

Field	Value
Paper	arXiv 2610.02117
Affiliations	Valeo.ai, first author's affiliation
First public release	2026-10-01 (arXiv v1)
Authors' code	Where-OPD
Adapter / method	`where-opd`
Local code	`src/auto_research/post_training/latest_20261002.py`

Original paper¶

Background and contribution¶

Procedural scenes supply object identities and coordinates. Give the teacher textual spatial privilege while the student sees only image and question. Distill teacher scores on the student's on-policy tokens.

flowchart LR
  G[Privileged spatial description] --> T[Teacher scores]
  I[Image and question] --> S[Student rollout]
  T --> K[Masked KL]
  S --> K

Key figure from the paper¶

Original paper diagram of spatially guided distillation. From the paper; copyright belongs to the authors. Not a local result.

Core mechanism¶

\(L=\sum_t m_t KL(\pi_T(\cdot|x,g),\pi_S(\cdot|x))/\sum_tm_t\).

Paper results¶

Training only on synthetic scenes reportedly gains 3.23 points on average across six real vision benchmarks.

Local reproduction¶

Compare disabled/default and enabled mechanisms. Three-seed diagnostics are diagnostic_only=true, checking invariant, gradient, or state behavior only.

Reproduction limits¶

No synthetic scenes were generated locally and no MLLM was loaded. Random logits are not visual results.