AIAny
Icon for item

VAD: Attributing Visual Evidence for Target Reconstruction in Multimodal On-Policy Distillation

Estimates the visually attributable portion of a privileged teacher’s next-token corrections and reconstructs student-anchored training targets for multimodal on-policy distillation. Uses counterfactual teacher queries and a signed proxy to raise supported tokens and suppress refuted ones, improving fine-grained visual knowledge transfer across model scales.

Introduction

Most privileged-view distillation mixes visual evidence with teacher priors, making it unclear which next-token corrections are truly supported by the image. VAD's core insight is to treat the change produced when adding or removing the visual evidence as a signed, intervention-derived proxy for the visual direction, and to reconstruct what to distill around the student distribution rather than naively matching the teacher.

Key Findings
  • VAD computes a signed proxy u_t from counterfactual teacher queries (teacher with evidence present vs removed) using centered log-probability shifts, then projects the teacher correction onto that proxy to separate an intervention-aligned component and a proxy-unexplained residual.
  • The method reconstructs a student-anchored target from the proxy-aligned component (raising supported candidates, suppressing refuted ones) while retaining a weak teacher-derived regularizer and splitting u_t into support/refutation branches for finer control.
  • Evaluated on six fine-grained visual benchmarks at 4B and 9B student scales, VAD consistently outperforms direct privileged-view distillation and visual-advantage weighting, achieving Avg_6 = 78.32 (4B) and 79.93 (9B) and leading scale-matched alternatives by ~2–3 points.
  • Gains are broad across tasks (detail localization, high-resolution perception, real-world recognition) and not driven by a single dataset; additional teacher views are required only during training, not at inference.
Who it's for & tradeoffs

Great fit if you train vision-language students with privileged-view teachers and need more precise visual-grounded corrections than source-mixed targets provide. It is most useful for fine-grained multimodal benchmarks and post-training distillation at scale. Look elsewhere if you cannot afford the extra training-time teacher queries (counterfactual views) or if simple advantage-weighting already meets your production constraints.

Where it fits

VAD is a post-training distillation technique positioned against Vision-OPD and decomposed OPD baselines; it prioritizes attributing which parts of a teacher correction are visually supported and reconstructing targets accordingly rather than only reweighting teacher signals.

Method detail (brief)
  1. For each student-generated prefix, query the fixed teacher with the relevant evidence present and removed; compute centered log-probability shifts to form u_t. 2) Project the original teacher correction onto the u_t proxy (regularized projection) to extract the intervention-aligned component and residual. 3) Reconstruct a target distribution anchored to the full-image student by shifting mass toward visually supported tokens and away from refuted ones; apply a light teacher regularizer. 4) Train the student on these reconstructed targets; at inference the student runs as usual with full images.

Information

  • Websitearxiv.org
  • AuthorsKangning Zhang, Yixing Li, Shuai Shao, Qingyao Li, Zhengxi Lu, Zhiyuan Yao, Jianghao Lin, Wenxiang Jiao, Yuan Lu, Weiwen Liu
  • Published date2026/07/30

More Items

Introduces WorldExam, a diagnostic benchmark that evaluates controllable video world models across four levels from visual quality to inherent world reactivity. Covers 1,474 cases across eight tasks and supports camera-, action-, and language-driven paradigms, measuring scene-conditioned reactions beyond explicit instructions.

Generates polygonal meshes from images using flow matching for fast, native mesh synthesis. Decodes vertices, edge connectivity, and face winding in one parallel pass, preserves artist-authored topology without vertex quantization or welding, supports a user-set vertex budget for face-count control, and completes image-to-mesh in ~6s median.

Converts text prompts into physically consistent videos by synthesizing executable Blender programs as a process-level chain-of-thought and using a dual-engine pipeline (deterministic simulation draft + draft-conditioned video editor). Ships with a VideoCoCo-3K draft–instruction–target dataset and shows substantial gains in physical-consistency benchmarks.