Pith. sign in

REVIEW 3 major objections 4 minor 119 references

This paper shows that swapping image and question order changes vision-language predictions by 6-26 percentage points, and that a label-free test-time update aligning the weaker branch to the stronger one closes most of the gap.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 10:03 UTC pith:BAWL5HRB

load-bearing objection A real gap, a plausible repair, and a missing control: image-first reliability per instance is not shown, so the causal story needs one more table. the 3 major comments →

arxiv 2607.20351 v1 pith:BAWL5HRB submitted 2026-07-22 cs.CV cs.CL

Test-Time Training for Modality Order Consistency in Vision-Language Models

Aditi Gupta , Yossi Gandelsman This is my paper
classification cs.CV cs.CL
keywords vision-language modelsmodality orderprompt sensitivitytest-time trainingasymmetric consistencyactivation patchingvisual question answeringrobustness
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper is trying to establish that modality-order sensitivity in vision-language models is a structured, causally localizable failure, not random prompt noise, and that a simple label-free test-time procedure can repair it. If true, it means the order in which an image and a question are fed to a model is a meaningful robustness axis for evaluation, and that the stronger prediction branch can serve as a teacher for the weaker branch at test time without labels. The authors support this with consistent evidence across three models and three benchmarks, showing accuracy gaps of 6-26 percentage points, and by localizing the failure to a narrow mid-network layer band with activation patching.

Core claim

The paper's central claim is that modern vision-language models have a repeatable, semantically irrelevant failure: presenting the same image and question in image-first order consistently beats question-first order by 6-26 percentage points across three models and three benchmarks. This is neither diffuse prompt noise nor a shallow decoding artifact: activation patching shows that replacing question-first hidden states with image-first states in a narrow mid-network band sharply restores accuracy, while patching early layers does almost nothing. The proposed repair is an asymmetric, label-free test-time training objective: for each instance, the question-first output distribution is aligned

What carries the argument

The central object is the asymmetric, label-free test-time training objective: for each test instance, the model is run twice (image-first and question-first), the image-first output distribution is detached and used as a fixed-within-step teacher, and a KL-divergence loss aligns the question-first distribution to it via a few gradient steps on the language-model backbone. The supporting machinery is activation patching, which swaps question-first hidden states with image-first hidden states layer by layer to localize the failure to a narrow mid-network band, and a moving-versus-frozen teacher ablation showing that recomputing the teacher each step is what produces the bootstrapping effect.

Load-bearing premise

The method's safety rests on the assumption that whenever the two orderings disagree, the image-first answer is the one to trust; the paper demonstrates this only through aggregate accuracy and never reports the rate at which question-first is correct and image-first wrong.

What would settle it

Compute, on a labeled benchmark, the two disagreement rates separately: examples where image-first is correct and question-first wrong versus examples where question-first is correct and image-first wrong. If the second rate is non-negligible (say, more than a few percent of examples), then the asymmetric KL objective can push question-first predictions toward wrong image-first answers, and the reported repair is partly error transfer rather than genuine alignment.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Modality order should be treated as a standard robustness axis in VLM evaluation; the 6-26 point gap is systematic across model families and dataset types.
  • A label-free, per-instance asymmetric KL update can recover most of the order gap and slightly improve the stronger branch, so no labeled data or retraining is needed to repair the failure.
  • Because the failure localizes to a narrow mid-layer band, test-time adaptation targeted at those layers is more effective than updating early or late layers.
  • Symmetric consistency objectives are substantially less effective, so directionality—using the empirically stronger ordering as teacher—is load-bearing for the repair.
  • Recomputing the teacher each TTT step is critical: a frozen teacher recovers only about 9 points of question-first accuracy versus 26 with the moving teacher in the primary setting.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the image-first-is-right assumption holds more broadly, the same asymmetric recipe should transfer to other semantics-preserving prompt perturbations, such as option order or instruction phrasings, whenever one variant is consistently more accurate.
  • The activation-patching localization hints at an even cheaper inference-time fix: directly copying the stronger branch's mid-layer activations into the weaker branch, which would avoid per-instance gradients at the cost of a controlled approximation.
  • The paper's evidence is strongest for multiple-choice and one captioning setting; an immediate testable extension is whether the gap and repair appear in free-form generation and in cross-attention-style VLMs, where fusion happens in a different architectural location.
  • The central safety assumption is testable directly: on any labeled benchmark, measure the rate of question-first-correct/image-first-wrong disagreements. If that rate is non-trivial, the asymmetric update could be transferring errors, and a confidence-weighted or two-directional objective would be needed.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper documents a repeatable modality-order effect in VLMs: for the same image, question, and answer options, image-first prompts outperform question-first prompts by 6–26 percentage points across three models and three benchmarks. It proposes a label-free, per-instance asymmetric test-time training (TTT) objective, L_align = D_KL(stopgrad(p_IF) || p_QF), which uses the image-first branch as a detached teacher and adapts only the question-first branch, with a moving teacher and a layer-window restriction. The authors report that this procedure recovers most of the question-first gap across all nine model–dataset pairs and even slightly improves image-first accuracy in seven of nine settings. Activation patching localizes the ordering failure to a narrow mid-network band, and NIoU analysis indicates that TTT aligns late-layer MLP computation between the two orderings. The paper includes ablations of the loss, layer window, and teacher variant, plus a text-only control, a CHAIR hallucination evaluation, and a reasoning-mode control.

Significance. If the results hold, the paper identifies a new and practically relevant robustness axis for VLMs and demonstrates a simple, label-free repair that works across model families. The cross-model consistency, the asymmetric-objective ablation, the moving-teacher comparison, and the K=0 decoding control in Appendix E are genuine strengths, and the code link supports reproducibility. However, the central causal claim depends on the image-first branch being a reliable per-instance teacher, and the paper does not yet provide the disagreement-conditioned analysis needed to support that assumption. The reported image-first deltas are small enough to be within sampling noise, and the window-ablation evidence for 'mechanism-guided' repair is quantitatively weak. With those points addressed, the contribution would be significant.

major comments (3)
  1. [§3.2 and §5.1] The asymmetric objective treats p_IF as a per-instance gold target: L_align = D_KL(stopgrad(p_IF) || p_QF). This is safe only if, on examples where the two orderings disagree, image-first is correct more often than question-first. The paper justifies directionality using aggregate accuracy (Fig. 1) and analyzes only the IF-correct/QF-wrong 'fixable subset' in §5.1. It never reports the complementary cell (QF-correct/IF-wrong) or the conditional accuracy of the teacher on disagreement examples. If that cell is nontrivial, the KL update actively moves QF toward an incorrect distribution, and because parameters are shared, the IF branch can be corrupted. The moving-teacher variant (Table 4) amplifies this risk because errors can propagate across steps. Please report the full 2x2 disagreement contingency, or at least P(IF correct | disagreement) and the TTT outcome on the QF-correct/IF-wrong
  2. [Table 1, §4.2] All results are reported as point estimates with no confidence intervals, standard errors, or significance tests. This matters most for the claim that image-first accuracy improves by +0.0 to +1.2 points in seven of nine settings. On RealWorldQA (n=765), a +1.2-point delta is approximately 9 correct answers, which is within the binomial standard error of roughly 1.6 points; on several rows the deltas are clearly noise (e.g., +0.0, +0.3). The central question-first gains are large and likely robust, but the 'bootstrapping' / mutual-consistency claim rests on IF deltas that are not distinguishable from noise. Please provide confidence intervals, bootstrap resampling, or multiple independent evaluation runs, and temper the IF-improvement claim accordingly.
  3. [Table 3, §4.3 and §5] The layer-window ablation is presented as evidence that mid-layer adaptation is mechanism-guided, but the numbers do not support a strong localization claim. Across the six windows in Table 3, question-first recovery ranges only from +25.2 to +26.2 (a 1.0-point spread), and image-first gains range from +0.0 to +0.7. On AI2D (Table 8) the spread is similarly small: early 82.0, mid 83.0, late 80.1. Without significance testing or repeated runs, the statement that mid-layers 'produce the strongest question-first recovery' (and that early/late windows are less effective) is not supported. Either add statistical evidence or rewrite the mechanistic claim as a weak trend. This does not undermine the core TTT effect but does undercut the 'mechanism-guided repair' contribution.
minor comments (4)
  1. [§4.2] The text states that InternVL2-8B on RealWorldQA improves to 60.0%, 'exceeding the image-first baseline of 61.8%'; 60.0 is below 61.8. Please correct this arithmetic error.
  2. [Appendix D] The sentence 'consistent with a targeted repair at the causally identified failure site from Section ??' contains an unresolved cross-reference ('??').
  3. [§5.2] The NIoU analysis refers to 'top components' without specifying the number k of selected components per layer and per component type. This parameter affects both IO and the random baseline; please report it.
  4. [References] Chou et al. appear as both [2026a] and [2026b] with identical titles and venues; this appears to be a duplicate reference. If they are distinct papers, please disambiguate.

Circularity Check

0 steps flagged

No significant circularity: central TTT-accuracy claims are measured against external labels; only minor non-load-bearing self-citation for the NIoU metric.

full rationale

The paper's main derivation chain is not circular. The modality-order gap is measured as accuracy on external benchmark labels (Fig. 1, Table 1), and the TTT objective's success is also evaluated against those labels, not against the KL objective itself. Thus the claimed QF gain and IF preservation are not equivalent to the loss by construction. No fitted parameter is renamed as a prediction: K, learning rate, and adaptation window are fixed/selected on a held-out calibration split and reported at fixed step counts. The activation-patching localization is causal and falsifiable: patching IF hidden states into QF states helps only in a mid-layer band, which is not guaranteed by the definition of the fixable subset. The only self-citation is in Sec. 5.2: 'Following Nikankin et al. [2025a], we measure overlap...' for the NIoU metric; the co-author is Yossi Gandelsman, but this metric is not load-bearing for the central accuracy claims, which stand on Tables 1-4 and Fig. 3. The paper's limitations (mechanistic analysis strongest for InternVL2-8B; window ablation on MMStar only) and the missing QF-correct/IF-wrong disagreement cell are correctness or generality risks, not circularity: the asymmetric KL teacher could reinforce wrong IF answers, but whether it does is an empirical question that the paper does not assume away. Overall, the derivation is self-contained against external benchmarks; the score reflects only the minor non-load-bearing self-citation.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The load-bearing inputs are: the aggregate asymmetry used to choose the teacher direction, per-model hyperparameters tuned on calibration splits, the causal-reading assumptions for activation patching and NIoU, and the unstated per-example reliability of the image-first branch. No new physical or architectural entities are introduced.

free parameters (4)
  • TTT step count K = K=5 InternVL2-8B, K=4 Qwen2.5-VL-7B, K=3 Qwen3-VL-8B
    Selected on a held-out calibration split (§4.1); main results and IF gains depend on this choice.
  • learning rate eta = 5e-4 for InternVL2-8B, 5e-5 for Qwen2.5-VL and Qwen3-VL
    Per-model manual choice (§4.1); scales with model family and affects convergence.
  • adaptation layer window = Default layers 18-22 (mid-narrow)
    Selected from Table 3 ablation on MMStar as the best IF/QF trade-off; results vary across windows.
  • top-k component count for NIoU = not reported
    §5.2 defines normalized IoU over 'top components' but does not state how many components are selected; Figure 4 depends on this unspecified rule.
axioms (5)
  • domain assumption Image-first branch is the reliably correct teacher on examples where the two orderings disagree
    §3.2 uses stopgrad p_IF as the detached teacher. The paper justifies this from aggregate accuracy (Fig. 1) and analyzes mainly the IF-correct/QF-wrong 'fixable subset'; if QF is correct on many IF-wrong disagreements, the objective can reinforce errors.
  • domain assumption The two prompt orders differ only in the relative position of the image token block
    §3.1 and §4.1 claim both prompts use the model's standard chat template with the same image, question, and options; any hidden template or formatting difference would weaken the 'modality order' attribution.
  • domain assumption Per-instance TTT with parameter reset gives independent, generalizable benefits
    §3.2 resets to the pretrained parameters after each example; the method assumes a few gradient updates on one instance transfer to other examples and do not overfit the teacher's mistakes.
  • domain assumption Replacing the final-token hidden state at layer L is a causally faithful intervention
    §5.1 uses h_QF^(L) <- h_IF^(L); in residual-stream transformers a single-vector patch can propagate non-locally through attention/MLP, so the sharp localization conclusion depends on this faithfulness assumption.
  • domain assumption NIoU of selected forward-pass components measures functional alignment
    §5.2 assumes that overlap of top-used attention-head/MLP components reflects causal circuit similarity, even though the measure is correlational and threshold-dependent.

pith-pipeline@v1.3.0-alltime-deepseek · 13292 in / 15916 out tokens · 131606 ms · 2026-08-01T10:03:41.227103+00:00 · methodology

0 comments
read the original abstract

We find that vision-language models are sensitive to a specific semantically irrelevant change: the order in which the image and question are presented. Across three models and three benchmarks, image first prompting consistently outperforms question-first prompting, revealing a repeatable modality order failure. We use this gap to design an order-consistent test-time training method. Our method substantially closes the modality-order gap across all evaluated settings. Surprisingly, it also yields consistent improvements in the stronger image-first branch over the baseline, hence bootstrapping both orderings toward mutual consistency. Activation patching localizes the ordering failure to a narrow mid-network region where representations diverge sharply between prompt orders. We find that the test-time training method repairs this misalignment across layers. Together, our results identify modality-order sensitivity as a circuit-level failure in VLMs and demonstrate that simple, asymmetric test-time adaptation can effectively mitigate it and even improve performance over the baseline.

Figures

Figures reproduced from arXiv: 2607.20351 by Aditi Gupta, Yossi Gandelsman.

Figure 1
Figure 1. Figure 1: Accuracy under image-first and question-first prompting. Image-first prompting consistently outperforms question-first prompting across datasets and model families. Prompt Sensitivity. LLMs are sensitive to prompt formatting, including option order [Zhao et al., 2021], instruction phrasing [Mizrahi et al., 2024], and the ordering of in-context examples [Lu et al., 2022]. In VLMs, sensitivity to prompt form… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative TTT trajectories. Each panel shows image-first (IF) and question-first (QF) predictions across TTT steps; the correct answer is highlighted in green. In these examples, IF reaches the correct answer earlier, while QF catches up over the adaptation trajectory. Prompt orders. For each instance, we evaluate two prompts that contain the same image, question, and answer candidates. The image-first p… view at source ↗
Figure 3
Figure 3. Figure 3: Activation patching localizes the ordering failure. We replace the question-first final￾token hidden state at layer L with the corresponding image-first hidden state and continue the forward pass normally. Patching becomes effective only in a narrow mid-layer band, where patched question￾first accuracy increases and IF-correct/QF-wrong examples are recovered at high rate. If the two prompt orders diverged … view at source ↗
Figure 4
Figure 4. Figure 4: TTT aligns late-layers MLP computation across prompt orders. We measure normalized intersection-over-union (NIoU) between the top late-layer components used by image-first and question-first forward passes, separately for attention heads and MLP sublayers. Negative NIoU indicates lower overlap than a size-matched random baseline. Before TTT, late-layer MLP sublayers show strongly below-random IF/QF overlap… view at source ↗
Figure 5
Figure 5. Figure 5: Activation patching results - Qwen3. Early adaptation (layers 0 4) Middle adaptation (layers 18 22) Late adaptation (layers 25 31) Layers finetuned 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 A g r e e m e n t r a t e P r(y IF = y QF ) 0.432 0.459 0.385 0.973 0.919 0.923 Before TTT After TTT (a) Output-level consistency. 0 5 10 15 20 25 30 Transformer layer 0.0 0.1 0.2 0.3 0.4 Cosine drift Early adaptation… view at source ↗
Figure 6
Figure 6. Figure 6: TTT restores order consistency with window-dependent internal drift. TTT increases agreement between image-first and question-first predictions for all tested adaptation windows. The corresponding layerwise cosine drift differs by window: early adaptation produces broad changes across the network, middle adaptation concentrates drift near the localized failure band, and late adaptation changes mainly the f… view at source ↗
Figure 7
Figure 7. Figure 7: TTT trajectories and image-first preservation across datasets. Top row: question-first accuracy across T=10 TTT steps for InternVL2-8B (K=5, blue) and Qwen2.5-VL-7B (K=4, red). Dashed lines indicate zero-shot baselines. Bottom row: image-first accuracy before and after TTT at the reported step. Image-first accuracy is maintained or slightly improved across most settings. H Broader Impact Modality order sen… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

119 extracted references · 20 linked inside Pith

  1. [1]

    Advances in neural information processing systems , volume=

    Attention is all you need , author=. Advances in neural information processing systems , volume=

  2. [2]

    Proceedings of the IEEE International Conference on Computer Vision , pages =

    Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks , author =. Proceedings of the IEEE International Conference on Computer Vision , pages =

  3. [3]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Hpsv3: Towards wide-spectrum human preference score , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  4. [4]

    Advances in Neural Information Processing Systems , editor=

    Test-Time Training with Masked Autoencoders , author=. Advances in Neural Information Processing Systems , editor=. 2022 , url=

    2022

  5. [5]

    arXiv preprint arXiv:1707.06347 , year=

    Proximal policy optimization algorithms , author=. arXiv preprint arXiv:1707.06347 , year=

    Pith/arXiv arXiv

  6. [6]

    arXiv preprint arXiv:2001.08361 , year=

    Scaling laws for neural language models , author=. arXiv preprint arXiv:2001.08361 , year=

    Pith/arXiv arXiv 2001

  7. [7]

    arXiv preprint arXiv:2203.15556 , volume=

    Training compute-optimal large language models , author=. arXiv preprint arXiv:2203.15556 , volume=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:1712.00409 , year=

    Deep learning scaling is predictable, empirically , author=. arXiv preprint arXiv:1712.00409 , year=

    Pith/arXiv arXiv

  9. [9]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Scaling vision transformers , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  10. [10]

    Proceedings of the National Academy of Sciences , volume=

    Explaining neural scaling laws , author=. Proceedings of the National Academy of Sciences , volume=. 2024 , publisher=

    2024

  11. [11]

    Journal of Machine Learning Research , volume=

    Scaling laws from the data manifold dimension , author=. Journal of Machine Learning Research , volume=

  12. [12]

    Advances in Neural Information Processing Systems , volume=

    Beyond neural scaling laws: beating power law scaling via data pruning , author=. Advances in Neural Information Processing Systems , volume=

  13. [13]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Superposition Yields Robust Neural Scaling , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  14. [14]

    arXiv preprint arXiv:2602.07488 , year=

    Deriving Neural Scaling Laws from the statistics of natural language , author=. arXiv preprint arXiv:2602.07488 , year=

    Pith/arXiv arXiv

  15. [15]

    Advances in Neural Information Processing Systems , volume=

    The quantization model of neural scaling , author=. Advances in Neural Information Processing Systems , volume=

  16. [16]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Scaling properties of diffusion models for perceptual tasks , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  17. [17]

    The Eleventh International Conference on Learning Representations , year=

    Progress measures for grokking via mechanistic interpretability , author=. The Eleventh International Conference on Learning Representations , year=

  18. [18]

    Proceedings of the International Conference on Learning Represenatations 2014 , year=

    Exact solutions to the nonlinear dynamics of learning in deep linear neural networks , author=. Proceedings of the International Conference on Learning Represenatations 2014 , year=

    2014

  19. [19]

    Proceedings of the National Academy of Sciences , volume=

    A mathematical theory of semantic development in deep neural networks , author=. Proceedings of the National Academy of Sciences , volume=. 2019 , publisher=

    2019

  20. [20]

    arXiv preprint arXiv:2201.02177 , year=

    Grokking: Generalization beyond overfitting on small algorithmic datasets , author=. arXiv preprint arXiv:2201.02177 , year=

    Pith/arXiv arXiv

  21. [21]

    Transactions on Machine Learning Research , year=

    Emergent Abilities of Large Language Models , author=. Transactions on Machine Learning Research , year=

  22. [22]

    Advances in neural information processing systems , volume=

    Neural tangent kernel: Convergence and generalization in neural networks , author=. Advances in neural information processing systems , volume=

  23. [23]

    Proceedings of the 41st International Conference on Machine Learning , pages=

    The linear representation hypothesis and the geometry of large language models , author=. Proceedings of the 41st International Conference on Machine Learning , pages=

  24. [24]

    Advances in Neural Information Processing Systems , volume=

    Refusal in language models is mediated by a single direction , author=. Advances in Neural Information Processing Systems , volume=

  25. [25]

    Advances in neural information processing systems , volume=

    Distributed representations of words and phrases and their compositionality , author=. Advances in neural information processing systems , volume=

  26. [26]

    arXiv preprint arXiv:2602.15029 , year=

    Symmetry in language statistics shapes the geometry of model representations , author=. arXiv preprint arXiv:2602.15029 , year=

    Pith/arXiv arXiv

  27. [27]

    The Twelfth International Conference on Learning Representations , year =

    Language Models Represent Space and Time , author=. The Twelfth International Conference on Learning Representations , year =

  28. [28]

    arXiv preprint arXiv:2209.11895 , year=

    In-context learning and induction heads , author=. arXiv preprint arXiv:2209.11895 , year=

    Pith/arXiv arXiv

  29. [29]

    arXiv preprint arXiv:2209.10652 , year=

    Toy models of superposition , author=. arXiv preprint arXiv:2209.10652 , year=

    Pith/arXiv arXiv

  30. [30]

    International Conference on Learning Representations , year=

    Function Vectors in Large Language Models , author=. International Conference on Learning Representations , year=

  31. [31]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    In-context learning creates task vectors , author=. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    2023

  32. [32]

    International Conference on Machine Learning , pages=

    Scaling Collapse Reveals Universal Dynamics in Compute-Optimally Trained Neural Networks , author=. International Conference on Machine Learning , pages=. 2025 , organization=

    2025

  33. [33]

    International Conference on Learning Representations , year =

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. International Conference on Learning Representations , year =

  34. [34]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Best-buddies similarity for robust template matching , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  35. [35]

    International Conference on Learning Representations , year=

    GAN Dissection: Visualizing and Understanding Generative Adversarial Networks , author=. International Conference on Learning Representations , year=

  36. [36]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Network dissection: Quantifying interpretability of deep visual representations , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  37. [37]

    Distill , volume=

    Multimodal neurons in artificial neural networks , author=. Distill , volume=

  38. [38]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    Transformer feed-forward layers are key-value memories , author=. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages=

    2021

  39. [39]

    Transactions on Machine Learning Research , year=

    Universal Neurons in GPT2 Language Models , author=. Transactions on Machine Learning Research , year=

  40. [40]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Rosetta neurons: Mining the common units in a model zoo , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  41. [41]

    arXiv preprint arXiv:1704.01444 , year=

    Learning to generate reviews and discovering sentiment , author=. arXiv preprint arXiv:1704.01444 , year=

    Pith/arXiv arXiv

  42. [42]

    2023 , howpublished =

    Language models can explain neurons in language models , author=. 2023 , howpublished =

    2023

  43. [43]

    International Conference on Learning Representations , year=

    Natural language descriptions of deep visual features , author=. International Conference on Learning Representations , year=

  44. [44]

    Forty-first International Conference on Machine Learning , year=

    A multimodal automated interpretability agent , author=. Forty-first International Conference on Machine Learning , year=

  45. [45]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year =

    Vision Transformers Don't Need Trained Registers , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year =

  46. [46]

    The Thirteenth International Conference on Learning Representations , year=

    Interpreting the Second-Order Effects of Neurons in CLIP , author=. The Thirteenth International Conference on Learning Representations , year=

  47. [47]

    arXiv preprint arXiv:2601.22594 , year=

    Language Model Circuits Are Sparse in the Neuron Basis , author=. arXiv preprint arXiv:2601.22594 , year=

    Pith/arXiv arXiv

  48. [48]

    Advances in neural information processing systems , volume=

    Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability , author=. Advances in neural information processing systems , volume=

  49. [49]

    Proceedings of the National Academy of Sciences , volume=

    Neural representational geometry underlies few-shot concept learning , author=. Proceedings of the National Academy of Sciences , volume=. 2022 , publisher=

    2022

  50. [50]

    International conference on machine learning , pages=

    Similarity of neural network representations revisited , author=. International conference on machine learning , pages=. 2019 , organization=

    2019

  51. [51]

    Advances in neural information processing systems , volume=

    Revisiting model stitching to compare neural representations , author=. Advances in neural information processing systems , volume=

  52. [52]

    Hopcroft , title=

    Yixuan Li and Jason Yosinski and Jeff Clune and Hod Lipson and John E. Hopcroft , title=. ICLR , crossref=. 2016 , cdate=

    2016

  53. [53]

    Forty-first International Conference on Machine Learning , year=

    Position: The platonic representation hypothesis , author=. Forty-first International Conference on Machine Learning , year=

  54. [54]

    Advances in neural information processing systems , volume=

    Insights on representational similarity in neural networks with canonical correlation , author=. Advances in neural information processing systems , volume=

  55. [55]

    Nature , volume=

    Emergence of simple-cell receptive field properties by learning a sparse code for natural images , author=. Nature , volume=. 1996 , publisher=

    1996

  56. [56]

    International Conference on Machine Learning , pages=

    On linear identifiability of learned representations , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

  57. [57]

    Proceedings of the national academy of sciences , volume=

    Performance-optimized hierarchical models predict neural responses in higher visual cortex , author=. Proceedings of the national academy of sciences , volume=. 2014 , publisher=

    2014

  58. [58]

    Frontiers in systems neuroscience , volume=

    Representational similarity analysis-connecting the branches of systems neuroscience , author=. Frontiers in systems neuroscience , volume=. 2008 , publisher=

    2008

  59. [59]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Understanding image representations by measuring their equivariance and equivalence , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  60. [60]

    Behavioral and brain sciences , volume=

    Representation is representation of similarities , author=. Behavioral and brain sciences , volume=. 1998 , publisher=

    1998

  61. [61]

    Neuron , volume=

    Matching categorical object representations in inferior temporal cortex of man and monkey , author=. Neuron , volume=. 2008 , publisher=

    2008

  62. [62]

    Science , volume=

    Distributed and overlapping representations of faces and objects in ventral temporal cortex , author=. Science , volume=. 2001 , publisher=

    2001

  63. [63]

    Journal of Neuroscience , volume=

    The representation of biological classes in the human brain , author=. Journal of Neuroscience , volume=. 2012 , publisher=

    2012

  64. [64]

    The Eleventh International Conference on Learning Representations , year=

    Relative representations enable zero-shot latent space communication , author=. The Eleventh International Conference on Learning Representations , year=

  65. [65]

    Journal of Neuroscience , volume=

    Representational similarity analysis reveals commonalities and differences in the semantic processing of words and objects , author=. Journal of Neuroscience , volume=. 2013 , publisher=

    2013

  66. [66]

    Forty-second International Conference on Machine Learning , year=

    Universal sparse autoencoders: Interpretable cross-model concept alignment , author=. Forty-second International Conference on Machine Learning , year=

  67. [67]

    The Thirteenth International Conference on Learning Representations , year =

    Scaling and evaluating sparse autoencoders , author=. The Thirteenth International Conference on Learning Representations , year =

  68. [68]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    Pith/arXiv arXiv

  69. [69]

    The Twelfth International Conference on Learning Representations , year=

    Sparse autoencoders find highly interpretable features in language models , author=. The Twelfth International Conference on Learning Representations , year=

  70. [70]

    2023 , journal=

    Towards Monosemanticity: Decomposing Language Models With Dictionary Learning , author=. 2023 , journal=

    2023

  71. [71]

    Transactions of the Association for Computational Linguistics , volume=

    Linear algebraic structure of word senses, with applications to polysemy , author=. Transactions of the Association for Computational Linguistics , volume=. 2018 , publisher=

    2018

  72. [72]

    arXiv preprint arXiv:2101.00027 , year=

    The pile: An 800gb dataset of diverse text for language modeling , author=. arXiv preprint arXiv:2101.00027 , year=

    Pith/arXiv arXiv

  73. [73]

    2009 IEEE conference on computer vision and pattern recognition , pages=

    Imagenet: A large-scale hierarchical image database , author=. 2009 IEEE conference on computer vision and pattern recognition , pages=. 2009 , organization=

    2009

  74. [74]

    OpenAI blog , volume=

    Language models are unsupervised multitask learners , author=. OpenAI blog , volume=

  75. [75]

    International conference on machine learning , pages=

    Pythia: A suite for analyzing large language models across training and scaling , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  76. [76]

    arXiv preprint arXiv:2205.01068 , year=

    Opt: Open pre-trained transformer language models , author=. arXiv preprint arXiv:2205.01068 , year=

    Pith/arXiv arXiv

  77. [77]

    2025 , eprint=

    Qwen2.5 Technical Report , author=. 2025 , eprint=

    2025

  78. [78]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Reproducible scaling laws for contrastive language-image learning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  79. [79]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  80. [80]

    Transactions on Machine Learning Research , issn=

    Maxime Oquab and Timoth. Transactions on Machine Learning Research , issn=. 2024 , url=

    2024

Showing first 80 references.