LEGO-OPD: Factorized Teacher Composition for Multimodal On-Policy Distillation¶
L1 core-mechanism diagnostic. Executes factorized teacher composition without loading LLM or VLM teachers.
Paper information¶
| Field | Details |
|---|---|
| Paper | LEGO-OPD: Factorized Teacher Composition for Multimodal On-Policy Distillation |
| Organization | First-author affiliation: KAIST |
| First public release | 2026-09-29 (arXiv v1) |
| Authors' code | No public author implementation was found as of 2026-10-03. |
| Adapter / method | lego-opd |
| Local reproduction code | src/auto_research/post_training/latest_20261003.py |
Original paper¶
Motivation and method¶
Use a language expert as the token prior and a grounding expert only as the visual likelihood. Their product of experts forms the OPD teacher without also distilling the VLM's language bias.
flowchart LR
I[Public input and current state] --> M[lego-opd mechanism]
M --> A[Auditable intermediates]
A --> O[Output or state update]Key figure from the paper¶
Figure 1: Overview of the paper's method and its key components. From the original paper; copyright remains with the authors. Click the image for the source.
Core equation¶
\(p_T(y|x,v)\propto p_L(y|x)p_G(y|x,v)^\alpha\).
Paper results¶
The paper reports better grounding on multimodal benchmarks while preserving language reasoning better than using a VLM teacher directly.
Local reproduction¶
Local comparison: The baseline disables the mechanism or uses default state; the treatment executes the defining operator. L1 reports no formal relative improvement, and relative percentages are not applicable.
- Three-seed diagnostic
diagnostic_only=true, ineligible for formal capability ranking.
Reproduction boundaries¶
Factorized teacher composition executes locally. No LLM or VLM teacher was loaded.