acceptodds
Withdrawn from ICLR 2027

CircuitProbe: Tracing Visual Temporal Evidence Flow in Video Language Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Autoregressive large vision–language models (LVLMs) process video by projecting video features into the LLM's embedding space as continuous visual token embeddings. Where temporal evidence is represented along this pathway, and how it causally shapes decoding, remains unclear. We present **CircuitProbe**, a circuit-level analysis framework that dissects the end-to-end video-language pathway in two stages. *Visual Auditing* localizes object semantics within the projected video-token sequence and establishes their causal necessity through targeted ablations and controlled substitutions. *Semantic Tracing* applies logit-lens probing to track the layer-wise emergence of object and temporal concepts, and uses temporal frame interventions to measure sensitivity to temporal structure. The analysis locates a consolidation interval in which visual content becomes language-aligned, together with a set of temporally specialized attention heads. Amplifying these heads in a single window fixed inside the consolidation interval (Layers 25–30), without any parameter update, improves all seven evaluated benchmarks, with significant held-out gains of +4.0% absolute on EgoSchema, +2.6% on VideoMME, and +2.4% on NExTQA, and smaller gains on the development benchmark TempCompass. These results provide interventional evidence that the identified layer interval and temporal-routing heads are functionally relevant to LVLM performance on temporal video benchmarks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Mechanistic Interpretability, Causal Intervention, Video language models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_gK9TOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a study of video language models (VLMs), focusing on action-object binding and chronological dependencies in the temporal understanding of three 7B VLM backbones. Towards this aim, the authors first carefully curate a subset of the STAR benchmark (Wu et al., 2021), implementing certain heuristics to make sure that visual reasoning is clearly isolated from language priors, and employing both open-ended and close-ended probes. The intervention experiments, defined based on the existing bounding box annotations of objects, demonstrate that task-critical semantics is highly localized with object-aligned token positions. Next, the paper explores the layer-wise emergence of language-aligned object and action concepts, and shows that they are concentrated to mid-to-late LLM layers. Based on these observations, the authors propose some measures to identify the so-called temporal heads, which route strongly to video tokens and distribute attention across multiple temporal slices. Then, they test the hypothesis that amplifying these temporal heads will carry cross-frame evidence in a more effective way, where they validate this on seven tested benchmarks, providing statistically-significant consistent gains.

Critical strengths
  • The story is coherent throughout the paper, and the experiments are consistent with the story. Each analysis is well chosen and creates a building block for the next section of the paper.
  • The proposed temporal head amplification is shown to provide consistent gains in seven tested benchmarks, without any finetuning.
Critical weaknesses
  • The paper needs to report a more detailed discussion on the conceptual similarities and differences between the findings of the proposed approach and the one in (Kim et al., ICLR 2026) to better assess the contributions of the work.
  • The rules to identify temporal heads depend on four different thresholds. It is not clear how the values for these thresholds are selected.
  • The name “circuitprobe”, and calling the framework “a circuit-based tracing framework” gives a false impression about the approach. As the authors say in the paper, it is a layer-wise analysis but using the word “circuit” implies a neuron-level detailed computational analysis.
Questions for authors
  1. Provide more details on the selected thresholds, or a sensitivity analysis would be much better. Do all the experiments in different benchmarks use the same thresholds?
  2. How does amplifying the temporal heads qualitatively affect the attention maps? What extra evidence can the authors provide to understand the contribution of the amplification?
Reviewer_6C1yOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents a mechanistic analysis of how visual evidence flows in Video-LLMs. The main findings are: (1) visual token ablation shows that object semantics are concentrated at object-aligned token positions rather than at other positions; (2) logit-lens analysis shows that object and action concepts become language-aligned mainly in middle-to-late layers; and (3) "temporal heads" are identified by thresholding statistics of their attention from the answer position to video tokens (i.e., the video attention ratio, the number of attended frames, and the temporal span). Amplifying the outputs of these heads in late layers (Layers 25–30) without training improves EgoSchema (+4.0), VideoMME (+2.6), and NExTQA (+2.4) on Qwen2.5-VL without degrading the other evaluated video benchmarks. Conversely, scaling down or removing these heads lowers performance on temporal benchmarks (e.g., −6.3 on TempCompass when fully ablated).

Critical strengths

Simple head identification validated by interventions. The paper proposes a simple method to identify attention heads that aggregate visual evidence across frames from statistics of the model's internal attention on a small calibration set, and validates it by directly intervening on the identified heads: amplifying them improves performance, while scaling them down or removing them lowers it, and the same interventions on randomly chosen heads do not yield such effects. The amplification requires no training or additional forward passes, and yields gains of 2.0–2.8 points on VideoMME and EgoSchema with Qwen2.5-VL.

Critical weaknesses
  1. Unclear whether the findings are specific to temporal information or reflect general visual recognition. Prior work defines temporal information through cross-frame interaction or binding (e.g., temporal IDs bound to textual representations [1], or temporal cues formed through inter-frame attention [2, 3]) and designs experiments accordingly. In contrast, this paper adopts setups from image-based VLM interpretability [4] without such a definition, so it is unclear whether each finding is specific to temporal information or simply reflects general visual recognition.

    • Visual auditing uses questions centered on object identity (e.g., "Which object did the person throw?"; Appendix A.1), whose answers are objects rather than temporal relations. Under this setup, the finding that ablating object-aligned tokens hurts performance is largely expected and mirrors [4] in the image domain.
    • Semantic tracing follows the logit-lens analysis of [4] and traces object and action names together, rather than temporal relations (Sec. 3.3.4).
    • Temporal heads are selected by attention from the answer position to visual tokens, which is closely related to how retrieval heads are identified in LLMs [5] (and concurrently in VLMs [6, 7]), but retrieval heads are not discussed in the paper. Finding heads that aggregate evidence across frames is useful in itself, but since their ablation affects order-independent Interaction questions as much as Sequence questions (Appendix D.7), it would be valuable to compare them with retrieval heads, e.g., by measuring their overlap with heads selected by retrieval scores [5–7], to clarify what is specifically temporal about them.
  2. Lack of clarity on the main take-aways.

    • The main text presents re-examinations of prior analyses (Circuits ① and ②) with the same weight as the paper's own contribution (the head-level analysis), making it difficult to identify the unique findings.
    • Meanwhile, experiments that are essential for interpreting the claims are only in the appendix, e.g., the order-perturbation analysis of the selected heads (Appendix D.7), the threshold tuning on TempCompass (Appendix D.1), and the much smaller effect without the counterfactual filter (Appendix A.3).
    • In addition, the question format and data subset change across experiments and are not always clearly stated (e.g., the format of the STAR probe set used for head selection). The terminology also differs from that commonly used in the community and is not consistent: the same head-output scaling is called "Surgical Recovery" (Fig. 1), "CircuitProbe-Derived Surgery" (Sec. 4), and "Temporal Head Amplification" (Sec. 4.2).

[1] R. Kang et al. Linear Mechanisms for Spatiotemporal Reasoning in Vision Language Models. ICLR, 2026.

[2] Y. Shi et al. Causality Matters: How Temporal Information Emerges in Video Language Models. AAAI, 2026.

[3] M. Kim et al. Map the Flow: Revealing Hidden Pathways of Information in VideoLLMs. ICLR, 2026.

[4] C. Neo et al. Towards Interpreting Visual Information Processing in Vision-Language Models. ICLR, 2025.

[5] W. Wu et al. Retrieval Head Mechanistically Explains Long-Context Factuality. ICLR, 2025.

[6] R. Sun et al. Mechanistic Insights into Functional Sparsity in Multimodal LLMs via CoRe Heads. arXiv:2606.05843, 2026.

[7] A. Li et al. Can Retrieval Heads See Images? Multimodal Retrieval Heads in Long-Context Vision-Language Models. arXiv:2605.27243, 2026.

Questions for authors
  1. Could the authors clarify the contributions and novel findings of this work beyond prior interpretability studies of VLMs and VideoLLMs [1–4] and work on retrieval heads [5–7]?
  2. How much do the selected heads overlap with retrieval heads [5–7], and are they more sensitive to frame-order than to spatial perturbations?
  3. Does the amplification generalize beyond Qwen2.5-VL, given that LLaVA-OV-V stays within ±1.0 in Table 3? If not, what would be the main reason (e.g., architecture, positional encoding, or training data)?
Reviewer_JjjoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This submission proposes CIRCUITPROBE, a circuit-level interpretability framework for autoregressive video LVLMs, organized as two probes. Visual auditing localizes object semantics within the projected video-token sequence and tests their causal necessity through token ablation and controlled substitution; the authors report 54–82 point accuracy drops when object-aligned tokens are removed, against 1–11 points for matched random tokens. Semantic tracing applies logit-lens probing to track the layer-wise emergence of object and action concepts and uses temporal frame interventions to measure sensitivity to temporal order, yielding a "consolidation interval" in mid-to-late layers together with a set of attention heads that route evidence across frames. The authors then amplify these heads (λ=1.75) inside a fixed five-layer window (Layers 25–30) with no parameter updates, reporting improvements on all seven benchmarks—held-out gains of +4.0 on EgoSchema, +2.6 on VideoMME, +2.4 on NExTQA—and dose-dependent degradation under suppression. The paper is accordingly framed as interventional evidence that the identified interval and heads are functionally relevant to temporal video reasoning, not as a new training recipe. Evaluation spans three 7B backbones, with the head-level effect established mainly on Qwen2.5-VL.

Critical strengths
  1. The mechanism claim is made falsifiable, and the controls are the right ones. The paper does not stop at profiling heads; it runs a bidirectional, count-matched intervention. Amplification yields Holm-significant held-out gains (+4.0 EgoSchema, +2.6 VideoMME, +2.4 NExTQA); suppression yields dose-dependent losses concentrated on temporal benchmarks while image benchmarks stay within roughly one point; and every count-matched non-temporal head set is flat. This matters because the video-LLM interpretability literature is still dominated by descriptive layer- and head-level readouts (cf. Kim et al., 2026). Here the claim can be wrong, and it survives the obvious alternative explanations. It sets an evidential bar for follow-up work and simultaneously hands practitioners a training-free diagnostic.

  2. The video- vs. image-trained contrast. Qualifying head counts differ by 2.2–2.6× (106–124 vs. 48), and amplifying the image-trained variant's same-layer heads does not help. This is design-relevant: it suggests video post-training installs cross-frame routing rather than refining existing heads — a testable claim about the field's current main lever (data- and training-side scaling), and one that holds even if the absolute gains prove modest.

Critical weaknesses
  1. The intervention's design signal is empty, and it is the paper's core deliverable. Every head-selection threshold (τ_m=0.1, τ_eff=2.0, τ_σ=0.8, τ_occ=0.6) and λ=1.75 is calibrated on TempCompass (D.1), yet TempCompass itself moves only +0.4 (n.s.) in the headline window — the development benchmark used for selection shows no effect. No raw λ sweep with error bars is reported, so λ=1.75 cannot be distinguished from a fitted peak, and the onset rules were, by the authors' own admission, "formalized afterwards." Combined with reporting the largest gain across three windows (+3.0, +3.6) as sensitivity rather than as selection, the "consolidation interval as a priori intervention target" — the paper's only substantive increment over Kim et al. (2026) — rests on post-hoc choices.

  2. The title and abstract promise temporal causality; the evidence supports multi-frame integration. The authors concede that separating temporal-order computation from generic cross-frame aggregation requires same-video controls (D.6), which are not run. VidHalluc — the largest gain (+3.6) — is largely a hallucination/static-consistency benchmark, and significant gains persist in Layers 15–20, before the stated onset. Hence "these heads route cross-frame evidence," not "temporal evidence flow."

Questions for authors

None

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.