Map the Flow: Revealing Hidden Pathways of Information in VideoLLMs¶

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Bohyung Han (SNU); ICLR 2026; mechanistic analysis of information flow in VideoLLMs

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

This paper applies mechanistic interpretability to Video Large Language Models (VideoLLMs), revealing a four-stage internal information flow during temporal reasoning: cross-frame interaction → video-language integration at temporal keywords → answer generation. Using Attention Knockout, the authors show that 58% of attention edges in LLaVA-NeXT-7B-Video-FT can be suppressed while retaining VideoQA performance.

Problem¶

VideoLLMs are typically studied from the outside (dataset scaling, key-frame selection, token compression), leaving their internal mechanisms opaque. It was unknown how spatiotemporal information is extracted from flattened video token sequences, how it merges with textual reasoning, and at what layer answer generation becomes reliable. Prior mechanistic work on image-based MLLMs does not straightforwardly transfer to the video setting.

Method¶

The authors apply Attention Knockout (Geva et al., 2023) — selectively setting attention mask entries to \(-\infty\) to sever specific token-to-token attention paths — and measure the relative change in answer probability \(\Delta p = \frac{p_\text{knockout} - p_\text{base}}{p_\text{base}} \times 100\). Complementing this, Logit Lens projects intermediate hidden states through the language model head to trace which semantic concepts (spatial vs. temporal) emerge in video tokens across layers.

Analysis focuses on LLaVA-NeXT-7B fine-tuned on VideoChat2-IT (LLaVA-NeXT-7B-Video-FT) using 8-frame sampling with 144 tokens per frame, evaluated on five multiple-choice tasks from TVBench. The question prompt is decomposed into fine-grained components (non-option question, true option, false option) to isolate which tokens serve as critical routing checkpoints for video-to-language information flow.

Key Contributions¶

First systematic blueprint of VideoLLM temporal reasoning via mechanistic interpretability, identifying four sequential processing stages across transformer layers.
Demonstration that video instruction fine-tuning explicitly induces cross-frame attention in early-to-middle layers absent in base ImageLLMs.
Logit Lens analysis showing spatial concepts emerge in early layers and temporal concepts emerge in middle layers with non-overlapping token positions.
Evidence that temporal keyword tokens (especially answer option tokens) act as information integration checkpoints through which video content flows before reaching the final generation token.
Practical pathway pruning: suppressing 58% of attention edges in LLaVA-NeXT-7B-Video-FT while retaining VideoQA accuracy.

Results¶

Blocking cross-frame attention in layers 1–16 causes accuracy drops of at least 18% on all five TVBench tasks for the video-fine-tuned model, while the base ImageLLM is largely unaffected:
Action Antonym: −24.1%
Action Sequence: −20.2%
Scene Transition: −18.0%
Moving Direction: −44.8%
Object Count: −60.8%
Layer-wise answer probability rises immediately after middle-layer video-language integration, confirming the model is answer-ready before final layers.
The effective pathway set (cross-frame, video→keyword, keyword→last token edges) suffices for VideoQA performance, with 58% of attention edges in LLaVA-NeXT-7B-Video-FT suppressed without significant degradation (Table 3 results partially cut off in source).
Results replicated across LLaVA-NeXT-13B-Video-FT, Mini-InternVL-4B-Video-FT, and VideoLLaMA3-7B (reported in appendix).

Limitations¶

Analysis is restricted to correct predictions only, excluding instances where the model fails, which may introduce selection bias toward idealized behavior.
Primary analysis uses multiple-choice VideoQA; open-ended and long-form video extensions are deferred to the appendix and may exhibit different flow patterns.
8-frame sampling is a strong compression of temporal content; findings may not generalize to higher frame-rate or very long video settings.
Attention Knockout and Logit Lens are correlational tools; they identify necessary pathways but cannot fully attribute causal mechanisms within MLP layers.
Study is limited to autoregressive decoder-only VideoLLMs; encoder-decoder or diffusion-based architectures are not addressed.

Relevance to Vision-Language Models¶

This work advances VLM interpretability beyond image-text understanding by characterizing how temporal reasoning emerges as a distinct capability acquired through video instruction tuning — not present in base ImageLLMs. The identification of temporal keyword tokens as structured information routing checkpoints provides a concrete target for architectural improvements such as attention sparsification or targeted fine-tuning, which is directly actionable for VLM efficiency and generalization research. The finding that 58% of attention edges are functionally redundant opens a clear path toward structured pruning of VideoLLMs. More broadly, the staged processing blueprint (spatial-first, temporal-second, generation-last) offers a mechanistic grounding for understanding why current VideoLLMs succeed or fail on temporal reasoning benchmarks.