OPD-Aha: From Linguistic Momentum to Visual Reflection in Multimodal On-Policy Distillation¶

L1 core mechanism diagnostic. The target reconstruction formula runs locally. No multimodal checkpoint was loaded.

Paper information¶

Field	Details
Paper	OPD-Aha: From Linguistic Momentum to Visual Reflection in Multimodal On-Policy Distillation
Institutions	Arizona State University / University of Virginia / Stevens Institute of Technology (first author affiliations)
First public release	2026-09-15 (arXiv v1)
Original source code	Yes: Echochef/OPD-Aha
Adapter/method	`opd-aha`
Local reproduction code	`src/auto_research/post_training/latest_20260916.py`

Original paper summary¶

Background and main changes¶

Reconstruct a visually preferred target from the difference between the teacher's distributions with actual and empty visual input. This counters momentum from an incorrect linguistic prefix.

The local reference kernel retains the defining gates, weights or state transitions and emits auditable intermediates for comparison under a shared budget.

flowchart LR
  I[Public inputs and state] --> M[OPD-Aha core mechanism]
  M --> A[Auditable intermediates]
  A --> O[Output or policy update]

Key figure from the paper¶

Figure 2: Core architecture, components and connections. From the paper; copyright belongs to the authors. Click for the source.

Core mechanism¶

The implementation preserves the defining state transition and exposes relevant weight, gate, depth, stage or cache statistics. Fixed budgets and three seeds check mechanism invariants, not paper-scale training.

Offline and online results¶

Capability, speed and cost claims are attributed to the paper only. Local metrics are not directly compared or extrapolated to online gains.

Local reproduction¶

See metrics/mechanism-seeds42-44.json. The unified receipt marks diagnostic_only=true, so the public dashboard does not treat this as a complete capability gain.

Scope¶

The target reconstruction formula runs locally; no multimodal checkpoint was loaded.