Pith. sign in

REVIEW 2 major objections 5 minor 22 references

Visual Access Boundaries in Vision-Language Model Reasoning

T0 review · 2 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Chain-of-thought prompting does not improve the models' answers by prolonging direct attention to image tokens; the visual-access boundary barely moves despite ~50x longer generations, and gains are gated by perceptual readout.

desk verdict Novel causal intervention and a solid qualitative finding; the abstract's 'at most two layers' precision claim outruns the statistics. read the letter →

arxiv 2607.12815 v2 pith:VNYTV67U submitted 2026-07-14 cs.AI

Hiroto Osaka , Shohei Taniguchi , Gouki Minegishi , Kai Yamashita , Masahiro Suzuki , Yutaka Matsuo This is my paper
classification cs.AI
keywords chain-of-thoughtpromptingvision-languagemodelsvisualaccessboundaryattentionmaskinginterventionperceptualreadouttest-timescalingdecayattributecounting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Chain-of-thought prompting is a standard way to scale vision-language models at test time, but the paper asks what is actually being extended: direct re-reading of image tokens, or language-side computation over visual information that has already entered hidden states. To answer, it masks attention from generated tokens to image tokens along layer depth and generation time, defining a Visual Access Boundary as the minimal access window that preserves accuracy. Across six configurations of two model families, both direct answering and chain-of-thought show finite boundaries, and in the larger models the chain-of-thought boundary sits within two layers of the direct-answering boundary even though chain-of-thought generates roughly fifty times more tokens. A second analysis holds the reasoning operation fixed and varies the visual attribute, showing chain-of-thought gains track how reliably the model can transcribe the attribute; an oracle that supplies attributes as text restores the gains, placing the bottleneck at perceptual readout rather than at counting.

What carries the argument

Visual Access Sweep, an additive-mask intervention that sets attention logits to negative infinity so that generated-token queries cannot attend to image-token keys outside an allowed rectangle defined by layer cutoff and token cutoff. Sweeping these cutoffs traces the Visual Access Boundary (VAB), the minimal access region that preserves accuracy within tolerance epsilon = 0.05. Its load-bearing property is that it tests functional necessity by intervention: it removes a specific computational channel while leaving image-derived information already propagated through the residual stream intact, so a finite VAB means later direct visual access is not necessary for the behavior.

What would settle it

Recompute the Direct and CoT Visual Access Boundaries with at least 1,000 images per attribute and confidence intervals; if the boundaries separate by more than two layers, or if a task that truly requires iterative visual inspection shows the CoT boundary deepening with generation length, the paper's decoupling claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that chain-of-thought prompting does not improve answers by prolonging direct visual access over a longer trace. Visual Access Sweep masks attention from generated tokens to image-token keys inside a layer/token rectangle, and the Visual Access Boundary is the smallest such rectangle that keeps accuracy within a 5-percentage-point tolerance; across six model configurations, finite boundaries appear under both Direct and CoT prompting, with the CoT boundary in larger models within about two layers of the Direct boundary despite roughly 50x longer generations. Separate readout experiments locate the ultimate limit in perceptual readout rather than in counting.

Load-bearing premise

The central comparison assumes that a 5-percentage-point accuracy tolerance on samples of 150 counting images and 100 GQA questions, with no confidence intervals, can resolve one- or two-layer differences in the Visual Access Boundary; if boundary estimates are that noisy, the claimed CoT-versus-Direct closeness is not established at the stated resolution.

Editorial extensions

If this is right

  • The well-documented decay of attention to image tokens during generation is not by itself evidence of failure; in the tested regimes, later direct access is not functionally necessary.
  • CoT gains on visual counting can be predicted by a cheap multi-object readout check, and routing CoT only when the attribute is readable preserves most of the benefit at roughly half the output-token cost.
  • Active-perception methods that re-acquire visual evidence based on the reasoning state target a different channel than ordinary CoT; the paper frames them as complementary, not redundant.
  • The finite VAB under both prompting modes suggests late-layer direct visual access is redundant in these architectures, which could inform where to prune or freeze visual-token pathways.
  • Because the bottleneck is readout rather than counting, improving the model's ability to verbalize visual attributes should raise the ceiling for CoT gains more than adding reasoning tokens alone.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: The VAB could serve as a task-level diagnostic for visual grounding efficiency; models that front-load visual extraction should show shallower boundaries, and a task that deepens the boundary would indicate a genuine need for iterative re-reading.
  • Editorial extension: A training implication not tested in the paper is that supervising attribute transcription (asking the model to name each object's attributes before answering) should increase CoT headroom on hard attributes like angle, size, and location.
  • Editorial extension: The small boundary shift despite 50x longer generations hints at a saturation pattern; testing VAB on tasks that deliberately require sequential visual search would bound how general this decoupling is.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper introduces Visual Access Sweep, a causal intervention that masks attention from generated-token queries to image-token keys along two axes—layer depth and generation time—and defines the Visual Access Boundary (VAB) as the minimal access region preserving task accuracy within a tolerance ε=0.05. Across six model configurations from Qwen2.5-VL and InternVL3 on controlled attribute-counting tasks, and on a GQA-derived real-image yes/no task, the authors report finite VABs under both direct answering and chain-of-thought (CoT) prompting. In the larger models (Qwen2.5-VL-32B, InternVL3-14B/38B), the CoT VAB layer is claimed to differ from the Direct VAB by at most two layers when evaluated against the no-CoT full-access target, despite roughly 50× longer generations. The paper further argues that CoT gains are limited by perceptual readout: CoT helps when the queried visual attribute is reliably readable, not when readout is unreliable. A symbolic-attribute oracle restores CoT gains when ground-truth attributes are given as text, and a single-object probe-vs-decode check shows that hard attributes can be linearly recoverable from hidden states yet difficult for the model to output. The paper interprets these results as evidence that CoT in VLMs primarily extends language-side computation over image-derived hidden-state information rather than prolonging direct image-token access.

Significance. If the main claims hold, this is a valuable mechanistic contribution. The causal intervention goes beyond observational attention-decay studies, and the distinction between direct image-token access and image-derived hidden-state information is well motivated. The paper ships a clean experimental design with controlled synthetic tasks, two intervention controls (null-sink and query-text block), an ε-sensitivity check, a cross-family/cross-scale comparison, and a real-image extension. The oracle-bypass and probe-vs-decode diagnostics usefully separate latent encoding, symbolic readout, and counting ability. The qualitative finding—that CoT does not require proportionally extended direct visual access—is supported by the sweep geometry and controls. However, the headline quantitative claim of 'at most two layers' lacks uncertainty quantification; the precision of the boundary estimates is not established at the sample sizes used. The readout-Co T-gain correlation, while suggestive, rests on five attributes per scale and is not statistically quantified. These issues are fixable and do not undermine the core qualitative contribution, but they need to be addressed before the paper can b

major comments (2)
  1. [§5.4, Appendix C.6] The headline claim that the CoT VAB layer differs from the Direct VAB by 'at most two layers' (abstract, §5.2) is not statistically grounded. Each accuracy estimate in the sweep uses 150 counting images per attribute (100 for GQA; Appendix A.1, C.6), and ε=0.05 is the accuracy-drop tolerance. For a proportion near 0.8 with n=150, the standard error is ~3.2 percentage points; the difference between full-access and cutoff accuracy has standard error ~4.6 percentage points, nearly the entire ε budget. The boundary is therefore a threshold crossing of a noisy difference, yet no confidence intervals or bootstrap estimates are provided. Appendix B.2 itself shows that ℓ*_D for color moves from 44 to 42 as ε varies from 0.03 to 0.10, indicating boundary resolution is coarse. With boundary estimates of this precision, the reported Δℓ*_DA values (0, 0, −2, +2, +1, +1 in Table 8) are indistinguisha
  2. [§5.4, Appendix C.6] The claim that CoT gains are 'predicted by' perceptual readout relies on correlations computed over only five attributes per scale. The Pearson/Spearman values (r=+0.92, +0.88, +0.51; ρ=+1.00, +1.00, +0.50) are not accompanied by p-values, confidence intervals, or permutation tests. With n=5, a Spearman ρ of +0.50 at 32B is not significant, and the high values at 3B/7B are driven by a single point (angle) at the low end. The paper appropriately calls these 'descriptive summaries' in Appendix D.2, but the main-text phrasing ('predicts', 'relationship is positive at every scale') overstates the strength. Because the readout bottleneck is a central claim, I ask for a permutation or bootstrap test over the 150-image sample, or a more conservative statement limiting the conclusion to the qualitative direction.
minor comments (5)
  1. [§4.2, Eq. (2)] Potential typo: 'The three drops we compare against ε are' is followed by equations but no explicit labels; naming the three quantities (Δ_D, Δ_CoT, Δ_DA) in the sentence would improve readability.
  2. [Appendix B.2, Table 6] The text refers to 'shaded columns' but the table as rendered has no shading; use bold or another visual cue consistent with the description.
  3. [Appendix C.6, Table 10] The notation 'Direct full accuracy /ℓ ∗ 0.84/40(N ∗ = 24)' is unclear; define N* or remove it. Also, report the number of questions used for the boundary estimate alongside the accuracies.
  4. [§5.2] The sentence 'across the 12 (family, scale, task) settings' is not fully consistent with the tables (Tables 7–8 list 6 model configurations × 2 tasks = 12 settings, but the GQA and cross-attention pilot add more). Clarify the count or specify the subset.
  5. [Appendix D.2, Table 14] Minor formatting: 'Pearsonr= +0.92' lacks a space; also consider adding a table note that the correlations are over five attributes and should not be read as a scaling law.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the VAB is a measured causal quantity, and the supporting diagnostics are independent of the boundary definition.

full rationale

The paper's central contribution is an empirical intervention, not a derivation from fitted parameters. The Visual Access Boundary is defined operationally in Section 4.2 as the minimal access region whose measured accuracy drop is within epsilon=0.05, using separately measured full-access baselines. The direct-anchored comparison (Eq. 2: ΔDA = A^D_full − A^CoT(ℓ,τ)) fixes a target accuracy, but the layer at which CoT crosses that target is an empirical observation: Table 8 reports non-uniform shifts (0, −2, +2, +1, +1) across settings, and settings where the target is never reached are marked N/A. Thus the headline 'at most two layers' is not forced by definition. The readout analysis is also not circular: multi-object readout accuracy is measured on a separate fill-in task, the oracle experiment removes image input entirely and supplies ground-truth attributes as text, and the linear probe is a separate 500-image single-object diagnostic. None of these quantities is derived from the VAB or from the CoT-gain values being predicted. The epsilon choice is arbitrary but is explicitly tested in Appendix B.2 for sensitivity, and the qualitative finite-boundary finding is supported by the sweep geometry and by control interventions (null-sink and late-layer query-text block). The only self-citations are standard references to prior CoT work (e.g., Kojima et al. 2022) and are not load-bearing. Concerns about sample size, confidence intervals, and tolerance sensitivity are statistical-correctness issues, not circularity, and are excluded from this score by the review rules.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central causal claims rest on the intervention's validity, the tolerance choice, and the readout measure. None of these is machine-checked or shipped; the paper supplies controls but they are qualitative. The oracle and probe are independent supporting diagnostics.

free parameters (3)
  • VAB tolerance epsilon = 0.05
    Defines the minimal access region in Section 4.2; sensitivity is checked on one model/task pair (Appendix B.2). Layer boundaries shift with this choice.
  • Adaptive CoT routing threshold = 0.5
    Appendix E uses readout accuracy > 0.5 to route to CoT; explicitly illustrative, not tuned. Not central to the main claim but a free choice.
  • Linear probe regularization lambda = 1.0
    Appendix D.1, Table 13: ell2-regularized multinomial logistic regression with lambda=1.0. Affects the probe-vs-decode gap magnitudes.
assumptions (4)
  • domain assumption Decoder-only prefix-fusion VLM: visual features enter as an image-token prefix, and direct visual access is self-attention from generated-token queries to image-token keys.
    Section 3.1. The intervention operationalizes 'visual access' this way; cross-attention models are treated only as a pilot (Appendix C.5).
  • domain assumption Masking attention with -inf removes the targeted channel without otherwise altering computation; hidden-state information persists.
    Eq. (1), Section 4.1-4.2. Controls (Appendix B.1) test redistribution and query-mediated re-reading but are qualitative.
  • domain assumption Accuracy within 5 percentage points defines functional preservation; task accuracy is a valid proxy for the target behavior.
    Section 4.2. Depends on sampling noise and small sample sizes (150 counting images, 100 GQA questions).
  • domain assumption The multi-object readout task measures whether an attribute is available as a usable symbol for the counting operation.
    Section 6.1. This bridges readout accuracy and CoT gain; if readout fails for format or scaffolding reasons, the correlation may be spurious.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Visual Access Boundaries in Vision-Language Model Reasoning." pith.science (2026). https://pith.science/paper/VNYTV67U

@misc{pith2026260712815,
  author       = {Pith},
  title        = {Pith review of: Visual Access Boundaries in Vision-Language Model Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VNYTV67U}},
  note         = {Machine review of arXiv:2607.12815}
}
read the original abstract

Chain-of-Thought (CoT) prompting is widely used as a test-time scaling strategy for Vision-Language Models (VLMs), but it remains unclear what is extended when VLMs generate longer reasoning traces. We ask whether CoT requires continued access to image tokens, or whether it mainly operates over visual information already made available earlier in the forward pass. We introduce Visual Access Sweep, a causal intervention that masks attention from generated-token queries to image-token keys along layer depth and generation time, and define the Visual Access Boundary (VAB) as the minimal access region that preserves task accuracy. Across six model configurations from Qwen2.5-VL and InternVL3, both no-CoT direct answering and CoT prompting exhibit finite VABs. In Qwen2.5-VL-32B and InternVL3 at 14B and 38B scales, when CoT is evaluated against the no-CoT full-access target, its VAB layer differs from the no-CoT boundary by at most two layers, despite substantially longer generations. This suggests that CoT does not primarily improve performance by prolonging direct image-token access throughout the reasoning trace, but by extending language-side computation over image-derived hidden-state information. We further show that CoT gains are constrained by perceptual readout. CoT helps when the queried visual attribute can be reliably read out by the model, but not when that readout is unreliable. A symbolic-attribute oracle shows that CoT can improve counting once ground-truth attributes are supplied as text, while a single-object probe-vs-decode check shows that hard attributes can be linearly recoverable from hidden states yet difficult for the model itself to output. Together, these analyses place the bottleneck at readout rather than counting.

Figures

Figures reproduced from arXiv: 2607.12815 by the authors.

Figure 1
Figure 1. Overview of Visual Access Sweep and the Visual Access Boundary. (a) Visual Access Sweep. At each sweep point (tcutoff, ℓcutoff), we allow direct image-token access only for generated￾token queries satisfying t ≤ tcutoff and layers satisfying ℓ ≤ ℓcutoff. Outside this allowed rectangle, attention from generated-token queries to image-token keys is masked. The Visual Access Boundary (VAB) is the minimal access pair (ℓ… view at source ↗
Figure 2
Figure 2. Controlled attribute-counting task. The reasoning operation (COUNT) is fixed while the queried attribute varies across {color, shape, location, angle, size}, isolating the ef￾fect of perceptual difficulty on a single, fixed reasoning program. Controlled attribute-counting tasks. The con￾trolled tasks vary the queried visual attribute while holding the reasoning operation fixed (counting). Each scene is procedurally … view at source ↗
Figure 3
Figure 3. Visual Access Boundary on Qwen2.5-VL-32B. (a) Direct inference shows a sharp layer￾wise transition: accuracy is preserved while a broad upper-layer region is blocked, but drops once the intervention reaches earlier layers. (b) Under CoT, the high-accuracy region remains concentrated in a similar depth range despite much longer generations. Stars mark the empirical VAB under ϵ = 0.05. Exact boundary values are report… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Real-image extension: VAB on GQA yes/no (Qwen2.5-VL-32B). (a) Direct inference reproduces the qualitative pattern of the controlled tasks. A broad upper-layer region can be blocked with little accuracy loss before a sharp transition. (b) Under CoT, the high-accuracy re…
Figure 5
Figure 5. Figure 5: Perceptual readout predicts CoT gain across Qwen2.5-VL scales. Each point is one queried attribute (color, shape, location, size, angle). The x-axis is multi-object perceptual readout accuracy: the model is asked to fill in the queried attribute for every object in a m…
Figure 6
Figure 6. Figure 6: Oracle bypass with the Qwen2.5-VL-3B checkpoint without image input. Ground-truth at￾tributes are serialized as text. (a) Minimal oracle pro￾vides only the queried attribute for each object. (b) Full oracle provides all object attributes [PITH_FULL_IMAGE:figures/full_…
Figure 7
Figure 7. Figure 7: Example controlled counting instance. Left: example image. Right: per-object attributes [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Late-layer query-text block control (Color Counting, Qwen2.5-VL-32B). In addition to blocking generated-token → image-token attention, this control blocks query-text → image￾token attention from the same cutoff layer upward. Early-layer query-text access is left intact…
Figure 9
Figure 9. Figure 9: Null-sink control (Color Counting, Qwen2.5-VL-32B). Removed image-token attention mass is redirected to a dummy non-visual sink instead of being redistributed over real non-image tokens. The qualitative VAB pattern is preserved, weakening the attention-redistribution a…
Figure 10
Figure 10. Figure 10: Qwen2.5-VL family sweeps. Rows correspond to model scale; left panels show Direct [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]
Figure 11
Figure 11. Figure 11: InternVL3 family sweeps. Rows correspond to model scale; left panels show Direct [PITH_FULL_IMAGE:figures/full_fig_p024_11.png]
Figure 12
Figure 12. Figure 12: CoT-side Visual Access Sweep on a cross-attention fusion VLM (Llama-3.2-11B-Vision [PITH_FULL_IMAGE:figures/full_fig_p025_12.png]
Figure 13
Figure 13. Figure 13: Angle counting as a limiting case (Qwen2.5-VL-32B). Full-access performance is near [PITH_FULL_IMAGE:figures/full_fig_p026_13.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 18 linked inside Pith

  1. [5]

    URLhttps://arxiv.org/abs/2506.08008. Y . Gandelsman, A. A. Efros, and J. Steinhardt. Interpreting CLIP’s image representation via text-based decomposition. InInternational Conference on Learning Representations (ICLR),

    arXiv

  2. [8]

    URLhttps://arxiv.org/abs/2407.21783. D. A. Hudson and C. D. Manning. GQA: A new dataset for real-world visual reasoning and compositional question answering. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR),

    arXiv

  3. [11]

    URL https://arxiv.org/abs/2307.13702. Y . Lin, A. Wang, M. Chen, J. Liu, H. Liu, J. Su, and X. Xiao. Investigating inference-time scaling for chain of multi-modal thought: A preliminary study,

    arXiv

  4. [12]

    URL https://arxiv.org/abs/ 2502.11514. 11 H. Liu, C. Li, Q. Wu, and Y . J. Lee. Visual instruction tuning. InAdvances in Neural Information Processing Systems (NeurIPS),

    arXiv

  5. [13]

    H. Liu, C. Li, Y . Li, and Y . J. Lee. Improved baselines with visual instruction tuning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024a. R. Liu, S. Bai, N. Saunshi, and S. Kakade. Chain-of-thought can reduce performance on tasks where thinking makes humans worse. InInternational Conference on Learning Re...

    arXiv 2025

  6. [15]

    URL https://arxiv.org/ abs/2305.04388. NeurIPS

    arXiv

  7. [16]

    URL https://proceedings.neurips.cc/paper_ files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf. J. Vig, S. Gehrmann, Y . Belinkov, S. Qian, D. Nevo, Y . Singer, and S. Shieber. Investigating gender bias in language models using causal mediation analysis. InAdvances in Neural Information Processing Systems (NeurIPS), volume 33, pages 12388–12401,

    2017

  8. [17]

    URLhttps://arxiv.org/abs/2503.12799. G. Xu, P. Jin, Z. Wu, H. Li, Y . Song, L. Sun, and L. Yuan. LLaV A-CoT: Let vision language models reason step-by-step. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2087–2098,

    arXiv

Show all 22 references
  1. [18]

    J. Yang, C. Cui, Y . Zhou, Y . Chen, P. Xia, Y . Wei, T. Yu, Y . Huang, and L. Wang. IKOD: Mitigating visual attention degradation in large vision-language models, 2025a. URL https://arxiv.org/ abs/2508.03469. S. Yang, Y . Niu, Y . Liu, Y . Ye, B. Lin, and L. Yuan. Look-back: ...

    arXiv

  2. [19]

    URLhttps://arxiv.org/abs/2502.17425. Q. Yuan, Q. Zhang, Y . Liu, J. Chen, Y . Lu, H. Lin, J. Zheng, X. Han, and L. Sun. ShortV: Efficient multimodal large language models by freezing visual tokens in ineffective layers. InProceedings of the IEEE/CVF Conference on Computer Visi...

    arXiv

  3. [20]

    URL https://arxiv.org/abs/2504.00502. Z. Zhang, A. Zhang, M. Li, H. Zhao, G. Karypis, and A. Smola. Multimodal chain-of-thought reasoning in language models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),

  4. [21]

    URLhttps://arxiv.org/abs/2504.10479. X. Zou, Y . Wang, Y . Yan, Y . Lyu, K. Zheng, S. Huang, J. Chen, P. Jiang, J. Liu, C. Tang, and X. Hu. Look twice before you answer: Memory-space visual retracing for hallucination mitigation in multimodal large language models,

    arXiv

  5. [22]

    13 Appendix Contents A Experimental Details 15 A.1 Dataset Generation

    URLhttps://arxiv.org/abs/2410.03577. 13 Appendix Contents A Experimental Details 15 A.1 Dataset Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 A.2 Task Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 A.3 Pr...

    arXiv

  6. [23]

    We denote the corresponding CoT-side boundaries by ℓ∗ CoT and ℓ∗ DA, respectively, and report shifts relative to the Direct boundaryℓ∗ D following Eq

    The CoT-own-max criterion asks when CoT preserves its own full-access accuracy, whereas the direct-anchored criterion asks when CoT reaches Direct’s full-access accuracy. We denote the corresponding CoT-side boundaries by ℓ∗ CoT and ℓ∗ DA, respectively, and report shifts relat...

    2025

  7. [2016]

    URL https://arxiv.org/abs/1610.01644. J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, and J. Zhou. Qwen-VL: A versatile vision-language model for understanding, localization, text reading, and beyond,

    arXiv

  8. [2017]

    URLhttps://arxiv.org/abs/1612.06890. S. S. Kancheti, A. S. Kanade, V . N. Balasubramanian, and T. Ganu. Chain-of-thought degrades visual spatial reasoning capabilities of multimodal LLMs.arXiv preprint arXiv:2604.16060,

    arXiv

  9. [2021]

    M. Geva, J. Bastings, K. Filippova, and A. Globerson. Dissecting recall of factual associations in auto-regressive language models. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP),

    2023

  10. [2022]

    URL https://arxiv.org/abs/ 2202.05262. NeurIPS

    arXiv

  11. [2023]

    URL https://arxiv.org/abs/2303.08112. S. Fu, T. Bonnen, D. Guillory, and T. Darrell. Hidden in plain sight: VLMs overlook their visual representations,

    arXiv

  12. [2024]

    URL https://openreview.net/forum?id=qrGjFJVl3m. S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, H. Zhong, Y . Zhu, M. Yang, Z. Li, J. Wan, P. Wang, W. Ding, Z. Fu, Y . Xu, J. Ye, X. Zhang, T. Xie, Z. Cheng, H. Zhang, Z. Yang, H. Xu, and J....

    arXiv

  13. [2025]

    URL https://arxiv.org/abs/ 2502.13923

    doi: 10.48550/arXiv.2502.13923. URL https://arxiv.org/abs/ 2502.13923. N. Belrose, Z. Furman, L. Smith, D. Halawi, I. Ostrovsky, L. McKinney, S. Biderman, and J. Steinhardt. Eliciting latent predictions from transformers with the tuned lens,

    doi arXiv

  14. [2026]

    URL https://arxiv.org/abs/2604.16060. S. Kang, J. Kim, J. Kim, and S. J. Hwang. See what you are told: Visual attention sink in large multimodal models. InInternational Conference on Learning Representations (ICLR),

    arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.