Pith. sign in

REVIEW 3 major objections 5 minor 60 references

TRAPSBench: Vision-Language Models Encode but Fail to Express Epistemic Restraint

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read VLMs internally know when video evidence is insufficient but generate confident guesses anyway.

desk verdict Genuinely new video unanswerability benchmark with a plausible representation–output gap story; the steering claim needs null-direction controls before the 'expression, not perception' framing fully lands. read the letter →

arxiv 2608.13167 v1 pith:EEVSKW7K submitted 2026-08-13 cs.CV cs.AIcs.CLcs.LG

Fnu Pramono , John Cai , Sourabh Kulkarni This is my paper · ORCID
classification cs.CVcs.AIcs.CLcs.LG
keywords vision-languagemodelsepistemicrestraintselectiveabstentionphysicalreasoningvideounderstandingactivationsteeringlinearprobinguncertaintycalibration
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

Vision-language models that watch short physics videos would rather guess than admit they cannot know: across sixteen models and five families, no model exceeds a PECS of 0.292 under ordinary prompting, even though PECS is defined to reward exactly the combination of correct answering when evidence suffices and abstention when it does not. The paper's central claim is that this failure is expressive, not perceptual. Linear probes decode the answerable/unanswerable distinction from hidden states and transfer it across physics domains at up to 0.91 AUROC, and adding a single-layer “void direction” to hidden states causally induces abstention on answerable videos and suppresses it on unanswerable ones. The same signature appears in three open-weight families with no shared training pipeline, so the bottleneck sits between an internal epistemic signal and the autoregressive output that fails to express it.

What carries the argument

The load-bearing object is the void direction: for a given layer $\ell$, $v_\ell = (\bar{h}^{\mathrm{void}}_\ell - \bar{h}^{\mathrm{control}}_\ell)/\|\bar{h}^{\mathrm{void}}_\ell - \bar{h}^{\mathrm{control}}_\ell\|$, the normalized difference between mean hidden states on unanswerable and answerable videos. The paper uses this vector in two ways: a linear probe tests whether the direction separates void from control across datasets, and activation steering adds $\pm \alpha v_\ell$ during generation to show causation. Around it sit TRAPSBench, a MuJoCo-generated set of 1,404 matched answerable/unanswerable video pairs across occlusion, chaotic sensitivity, and ill-posed questions, and PECS = $\mathrm{Acc} \times \max(0, \mathrm{AbsRec} - \mathrm{FalseAbs})$, a conjunction metric that zeros both always-answer and always-abstain policies.

What would settle it

Create control videos that match the void videos’ low-level appearance (same motion energy, same occlusion footprint, same rendered object sizes) but remain deterministically answerable; retrain the probe and the steering direction on this matched set. If cross-dataset AUROC drops to chance and steering no longer changes abstention, the decoded direction was a visual artifact rather than an epistemic signal.

Watch

Extended reading notes

Core claim

On TRAPSBench—1,404 matched MuJoCo video pairs in which a control clip has a deterministic outcome and a near-identical void clip is unanswerable via occlusion, chaotic sensitivity, or an ill-posed question—the paper establishes that models' spontaneous restraint is poor across all sixteen VLMs tested. The new conjunction metric PECS = Acc × max(0, AbsRec − FalseAbs) forces selectivity: six degenerate strategies including never-abstain and always-abstain score zero, and the best unguided score is 0.292. Yet the void/control distinction is linearly decodable from frozen hidden states and transfers across domains, across visual and textual void types, and even to the specific void samples the model confabulates on (AUROC far above chance). Steering one layer's hidden states along the normalized void-minus-control direction turns abstention on for control videos and off for void videos, replicating across three open-weight families. The authors conclude that VLMs encode answerability but fail to express it, that visual evidence gaps are about four times harder to detect than textual impossibility, and that closing this gap will require output-stage interventions.

Load-bearing premise

The story stands on the assumption that the computed void direction—the normalized difference between hidden states on void and control videos—is an epistemic signal about answerability rather than a low-level perceptual or question-text artifact; if that premise fails, the conclusion that models internally encode answerability independently of output behavior is not established.

Editorial extensions

If this is right

  • Reliability work should shift toward output-stage interventions—steering, decoding constraints, or explicitly permitted abstention—because the epistemic representation already exists internally.
  • Simply adding an “I don’t know” option to the system prompt is a cheap, reproducible fix that raises abstention recall a median 1.9× across main-family video-native models without hurting control accuracy.
  • Chain-of-thought reasoning is not a reliable calibration aid: in at least one family, thinking overrides the model’s own expressed doubt and increases confabulation, so training rewards must penalize speculation.
  • Evaluations of VLM reliability need visual-evidence voids, not just textual tricks: unanswerability from missing visual evidence is detected roughly four times less readily than question-side impossibility.
  • Internal monitoring is a viable safety signal: because the void/control distinction is decodable even on confabulated outputs, a probe reading hidden states could flag should-abstain cases that the text gets wrong.

Reading between the lines

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

  • If the paper’s claim holds, an untested consequence is that a decoding-side probe could serve as an abstention head for models whose weights are not open, reading hidden states to veto confident answers—the paper only demonstrates direct steering in open-weight models.
  • The geometric result—occlusion directions transfer across domains while chaotic directions are near-orthogonal—predicts that benchmarks built on visible occluders will overstate model restraint relative to benchmarks with implicit evidence gaps.
  • The visual/textual asymmetry is consistent with training-data statistics: text corpora include many explicit “cannot determine” statements, whereas video corpora rarely label missing evidence. If so, finetuning on void-labeled video pairs should narrow the gap—a test the paper does not run.
  • The success of single-layer steering in three unrelated open-weight families suggests the mechanism is architectural rather than pipeline-specific, but it also means the claim cannot be checked in closed models with current API access.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces TRAPSBench, a procedurally generated video benchmark of matched physics pairs in which a single modification (occlusion, chaotic truncation, or an ill-posed question) makes the outcome undeterminable, together with a new metric, Penalized Epistemic Calibration Score (PECS), that rewards correct answering on answerable controls and selective abstention on unanswerable voids. Across 16 VLMs, spontaneous abstention is poor (best PECS 0.292). The central claim is that the bottleneck is expressive rather than perceptual: linear probes decode the void/control distinction from hidden states with cross-dataset AUROC up to 0.91, and activation steering with a single-layer void direction causally induces or suppresses abstention. These mechanistic results are replicated on three open-weight families (Qwen3-VL-8B, Gemma 4 E4B, LLaVA-NeXT-Video-7B). The paper also reports a visual-vs-textual asymmetry and reasoning-dependent effects on calibration.

Significance. If the central claim holds, the paper makes a valuable contribution: TRAPSBench is a well-motivated, procedurally generated benchmark; PECS is a sensible conjunction metric whose degenerate-strategy behavior is explicitly verified; the Cf-to-Cf restriction is a strong control for behavioral confounds in probing; and the cross-family replication on three open-weight families is commendable. The transparent evaluation-provenance appendix is also a positive feature. However, the headline mechanistic claims rest on two load-bearing methodological choices—target-label-based best-layer selection in probing, and a steering experiment without null-direction controls—that currently leave the strength of the 'encode but fail to express' conclusion uncertain. The benchmark and metric contributions are likely to be useful to the community regardless, but the central mechanistic claim needs additional controls before it can be accepted as stated.

major comments (3)
  1. [Section 5.4, Table 3] The reported probe AUROC values are per-pair maxima over layers, with the best layer selected using target-dataset labels. As the authors acknowledge in the text, 'best-layer selection does' see target labels. This is a form of test-set selection that can inflate AUROC, especially across 37 layers, and it directly affects the abstract's 'up to 0.91 AUROC' claim. Please report transfer AUROC with a layer chosen without target labels (e.g., a fixed layer, a layer chosen by source-side validation, or the average over layers), and apply the same protocol to the Cf-to-Cf restriction. The current numbers are not unbiased estimates of transferable decodability.
  2. [Section 5.5, Appendix I.5] The steering experiment compares +alpha and -alpha only against the alpha=0 baseline; it does not include any null-direction control. Because h_void and h_control differ in low-level properties (presence of an occluder, truncated video length, question text), the void direction v_l could encode a generic 'different input' or 'hedging' axis rather than an epistemic signal. Adding such a vector at every token position with alpha up to 10 could push activations off-manifold and produce generic refusal that a text judge labels as abstention. The paper's construct caveat in Appendix I.5 is appropriate but the cited converging evidence does not include the one experiment that would rule out a non-epistemic direction. Please add norm-matched random directions, control-vs-control difference directions, and ideally an unrelated task direction, and show that abstention induction is specific to the void direction and is not simply a degradation artifact at high alpha.
  3. [Appendix P] The paper mixes two evaluation epochs in its headline results: the PECS leaderboard (Table 1) and per-dataset tables use the corrected occlusion/chaotic splits but reuse original-evaluation chaotic ill-posed measurements, while the visual-vs-textual asymmetry (Table 2, Figure 3b) and the activation-steering analyses are reported entirely on the original evaluation. The appendix discloses this clearly, but the main text presents the abstract's 4x asymmetry claim and the PECS numbers without this provenance distinction, and the steering results were not recomputed on the corrected benchmark. Because the benchmark correction changed questions and ground truths, results from the two epochs are not directly comparable. Please either recompute all headline numbers on a single benchmark version or clearly separate, in the main text, which claims rest on which evaluation.
minor comments (5)
  1. [Abstract] The abstract states '1,404 matched physics pairs', but Appendix P indicates that the released and evaluated set comprises 1,354 pairs per model per regime (202 occlusion, 500 chaotic, 202 occlusion ill-posed, 450 chaotic ill-posed). Please reconcile the pair count.
  2. [Table 6 and Appendix F] Table 6 lists Chaotic Ill-Posed as N=500, while Appendix F reports analyses on N=450 and the text says 50 restored Plinko items have no ill-posed measurements. Please clarify the released versus evaluated number.
  3. [Table 1 and throughout] Several table entries omit the leading zero (e.g., '.292', '.568'), which is inconsistent with the rest of the tables that use '0.292'. Please standardize numeric formatting.
  4. [Appendix P, steering] The statement that the steering recompute 'was blocked by a build-environment incompatibility' should be moved into the main-text limitations section, since activation steering is a central causal claim and the paper should not rely on an appendix-only disclosure for this point.
  5. [Equation (1)] Please define Acc explicitly as control-accuracy in the equation caption or surrounding text, since the metric is applied only to answerable control videos.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the benchmark, metric, probes, and steering are externally grounded and do not reduce to their inputs.

full rationale

I walked the claimed derivation chain and found no step where a prediction or first-principles result is equivalent to its inputs by construction. PECS is a defined metric (Eq. 1) rather than a fitted quantity, and its robustness is checked against degenerate strategies in Appendix M; it is transparently a product of accuracy and Youden's J, so its behavior follows from its definition rather than from any circular fit. The probing results train a linear probe on one dataset's hidden states with source labels and evaluate on a different dataset whose labels were not used for training; the Cf->Cf restriction further removes the behavioral confound by limiting both train and test sets to samples the model confabulated on. The steering experiments compute a void direction from hidden-state means of one dataset and apply it to different datasets, with the causal claim resting on the intervention changing abstention behavior on held-out pairs. No self-citations are load-bearing, and no uniqueness or ansatz is imported from the authors' prior work. The per-pair best-layer selection disclosed in Section 5.4 is a model-selection issue that can inflate AUROC, not circularity. The construct caveat in Appendix I.5, acknowledging that the direction may capture a generic hedging-style axis and that the ill-posed conditions exclude visual confounds only for those directions, is a validity limitation rather than a circular derivation. Cross-architecture replication on Gemma and LLaVA with no shared training pipeline provides independent support for the central representation-output gap claim. The remaining weaknesses, such as the absence of matched-norm null-direction steering controls, are concerns about experimental completeness and alternative explanations, not about the derivation reducing to its own inputs. Therefore the appropriate circularity score is 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The benchmark and metric make standard domain assumptions about simulator ground truth and judge validity. The mechanistic claims add assumptions about hidden-state directions: linear probes operationalize internal knowledge, and mean-difference steering vectors are interpreted as epistemic rather than perceptual. The two free parameters that most affect reported numbers are the steering strength and the per-pair best-layer probe selection.

free parameters (2)
  • alpha (steering strength) = swept 0, 2, 5, 10; headline at 10
    Activation steering causal claims use this hand-chosen perturbation scale. The dose-response is monotone for occlusion directions and flat for chaotic directions, so the choice of endpoint affects reported abstention rates.
  • probe best layer per transfer pair = varies L1-L42, per-pair maxima
    LR probe AUROC is reported at the best layer chosen using target labels. This selection inflates reported transfer AUROC and is a model-selection choice rather than a fixed protocol.
assumptions (5)
  • domain assumption MuJoCo rigid-body simulations provide correct deterministic ground truth for control videos.
    TRAPSBench labels come from the simulator; if simulator physics diverge from intended scenarios, accuracy and PECS are mis-measured.
  • domain assumption Minimal control and void pairs differ only in the epistemic availability of the answer.
    The design claims a single targeted change renders the outcome undeterminable; any additional visual or semantic difference could confound abstention comparisons.
  • domain assumption The text-only judge panel correctly identifies abstention and answer correctness.
    All benchmark metrics pass through judge labels; although inter-judge agreement is high, judges never see the videos and may mislabel evasive answers.
  • domain assumption Linear separability of hidden states under an LR probe operationalizes 'internally distinguishes'.
    The paper equates decodability with internal knowledge; nonlinear structure or behavioral confounds could weaken this equation.
  • ad hoc to paper The mean-difference void direction v_l is an epistemic direction whose addition or removal at one layer is a faithful causal intervention.
    Steering conclusions rest on this direction. The paper itself lists construct, manifold, and mediation caveats in Appendix I.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TRAPSBench: Vision-Language Models Encode but Fail to Express Epistemic Restraint." pith.science (2026). https://pith.science/paper/EEVSKW7K

@misc{pith2026260813167,
  author       = {Pith},
  title        = {Pith review of: TRAPSBench: Vision-Language Models Encode but Fail to Express Epistemic Restraint},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EEVSKW7K}},
  note         = {Machine review of arXiv:2608.13167}
}
read the original abstract

When visual evidence is occluded or chaotic, models should abstain. In this paper, we show that Vision-Language Models (VLMs) can internally distinguish when abstention is required, but fail to express it anyway. We introduce TRAPSBench, a procedurally generated video benchmark of 1,404 matched physics pairs in which a single targeted change renders the outcome undeterminable from the visual evidence. Furthermore, we introduce Penalized Epistemic Calibration Score (PECS), a new robust metric that requires models to both answer correctly when the outcome is knowable, and abstain when the outcome is not. Across 16 VLMs spanning five families, spontaneous restraint is poor: the best PECS is 0.292. The bottleneck is expression, not perception: linear probes decode answerability from hidden states at up to 0.91 AUROC across physics domains; steering a single-layer void direction causally induces or suppresses abstention. Our results replicate across three open-weight families (Qwen, Gemma, LLaVA). The failure is also more pronounced in visual than textual uncertainty: models detect textual impossibility about 4x more readily than missing visual evidence. Closing this representation--output gap likely requires output-stage interventions.

Figures

Figures reproduced from arXiv: 2608.13167 by the authors.

Figure 1
Figure 1. TRAPSBench: answer when evidence is sufficient, abstain when it is not (worked example, seesaw sorter). Top (control): the judge scores the response against the deterministic MuJoCo ground truth. Bottom (void): the video is truncated before the outcome resolves; correct behavior is abstention. A high PECS requires both [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Six released TRAPSBench pairs with their verbatim questions and ground truths; [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) Control accuracy vs. epistemic discrimination [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: (a) Reasoning’s effect on restraint (arrows: off [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Steering geometry (Qwen3-VL-8B). (a) Oc-family directions (blue) transfer strongly; [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Visual vs. textual AbsRec by model family (Standard). Textual unanswerability [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: “Knows but won’t say,” per transfer pair: Cf [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]
Figure 8
Figure 8. Figure 8: Void-direction cosine similarity for all three families at their steering layers, [PITH_FULL_IMAGE:figures/full_fig_p027_8.png]
Figure 9
Figure 9. Figure 9: Gemma 4 E4B within-domain dose-response at L35 (verification re-run; data of [PITH_FULL_IMAGE:figures/full_fig_p027_9.png]
Figure 10
Figure 10. Figure 10: Occlusion sub-scenario abstention recall (%) for 9 video-native models under [PITH_FULL_IMAGE:figures/full_fig_p030_10.png]
Figure 11
Figure 11. Figure 11: Per-model abstention recall, Standard (hollow) vs. Guided (filled): the per-model [PITH_FULL_IMAGE:figures/full_fig_p032_11.png]
Figure 12
Figure 12. Figure 12: Confabulation taxonomy prevalence by reasoning pair ( [PITH_FULL_IMAGE:figures/full_fig_p034_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 41 canonical work pages

  1. [1]

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

    The ``something something'' video database for learning and evaluating visual common sense , author=. Proceedings of the IEEE International Conference on Computer Vision , year=

    work page

  2. [2]

    International conference on learning representations , year=

    Clevrer: Collision events for video representation and reasoning , author=. International conference on learning representations , year=

    work page

  3. [3]

    arXiv preprint arXiv:1803.07616 , year=

    Riochet, Ronan and Ynocente Castro, Mario and Bernard, Mathieu and Lerer, Adam and Fergus, Rob and Izard, V. arXiv preprint arXiv:1803.07616 , year=

    arXiv

  4. [4]

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

    Clevr: A diagnostic dataset for compositional language and elementary visual reasoning , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=. 2017 , doi=

    work page 2017

  5. [5]

    Nature Human Behaviour , year=

    Intuitive physics learning in a deep-learning model inspired by developmental psychology , author=. Nature Human Behaviour , year=

    work page

  6. [6]

    Advances in neural information processing systems , pages=

    What uncertainties do we need in bayesian deep learning for computer vision? , author=. Advances in neural information processing systems , pages=

  7. [7]

    International conference on machine learning , pages=

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning , author=. International conference on machine learning , pages=

    work page

  8. [8]

    Advances in neural information processing systems , pages=

    Simple and scalable predictive uncertainty estimation using deep ensembles , author=. Advances in neural information processing systems , pages=

    work page

Show all 60 references
  1. [9]

    Proceedings of the IEEE international conference on computer vision , pages=

    Vqa: Visual question answering , author=. Proceedings of the IEEE international conference on computer vision , pages=

  2. [10]

    European conference on computer vision , pages=

    Microsoft coco: Common objects in context , author=. European conference on computer vision , pages=. 2014 , organization=

    2014

  3. [11]

    Know what you don't know: Unanswerable questions for

    Rajpurkar, Pranav and Jia, Robin and Liang, Percy , booktitle=. Know what you don't know: Unanswerable questions for. 2018 , doi=

    2018

  4. [12]

    arXiv preprint arXiv:2006.09454 , year=

    Selective question answering under domain shift , author=. arXiv preprint arXiv:2006.09454 , year=

    2006 arXiv

  5. [13]

    Tragoudaras, Antonios and Zhang, Chenyu and Cherniavskii, Daniil and Vozikis, Antonios and Nijdam, Thijmen and Prinzhorn, Derck W. E. and Bodracska, Mark and Sebe, Nicu and Zadaianchuk, Andrii and Gavves, Stratis , booktitle=. Evaluating

  6. [14]

    A Shortcut-aware

    Krojer, Benno and Komeili, Mojtaba and Ross, Candace and Garrido, Quentin and Sinha, Koustuv and Ballas, Nicolas and Assran, Mahmoud , journal=. A Shortcut-aware

  7. [15]

    arXiv preprint arXiv:2601.00138 , year=

    Explicit Abstention Knobs for Predictable Reliability in Video Question Answering , author=. arXiv preprint arXiv:2601.00138 , year=

  8. [16]

    bioRxiv , year=

    Paired evaluation of machine-learning models characterizes effects of confounders and outliers , author=. bioRxiv , year=

  9. [17]

    Applied Sciences , volume=

    Impact on Classification Process Generated by Corrupted Features , author=. Applied Sciences , volume=

  10. [18]

    International Conference on Machine Learning , pages=

    Leveraging procedural generation to benchmark reinforcement learning , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    2020

  11. [19]

    arXiv preprint arXiv:2410.22584 , year=

    Benchmarking and Automated-Benchmarking with Large Language Models , author=. arXiv preprint arXiv:2410.22584 , year=

  12. [20]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Unsolvable Problem Detection: Robust Understanding Evaluation for Large Multimodal Models , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2025 , address=

    2025

  13. [21]

    doi:10.1109/TPAMI.2024.3437288 , year=

    Guo, Yangyang and Jiao, Fangkai and Shen, Zhiqi and Nie, Liqiang and Kankanhalli, Mohan , journal=. doi:10.1109/TPAMI.2024.3437288 , year=

    2024

  14. [22]

    Saadat, Asir and Aziz, Syem and Mahmud, Shahriar and Mahi, Abdullah Ibne Masud and Ahmed, Sabbir , journal=

  15. [23]

    Vardi, Ben and Nir, Oron and Shamir, Ariel , booktitle=

  16. [24]

    He, Xingwei and Zhang, Qianru and Jin, A-Long and Yuan, Yuan and Yiu, Siu-Ming , journal=

  17. [25]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    Selectively Answering Visual Questions , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=. 2024 , address=

    2024

  18. [26]

    International Conference on Learning Representations , year=

    Discovering latent knowledge in language models without supervision , author=. International Conference on Learning Representations , year=

  19. [27]

    Advances in Neural Information Processing Systems , volume=

    Inference-time intervention: Eliciting truthful answers from a language model , author=. Advances in Neural Information Processing Systems , volume=. 2023 , doi=

    2023

  20. [28]

    International Conference on Learning Representations , year=

    Certainly Uncertain: A Benchmark and Metric for Multimodal Epistemic and Aleatoric Awareness , author=. International Conference on Learning Representations , year=

  21. [29]

    Proceedings of the 34th International Conference on Machine Learning , pages=

    On Calibration of Modern Neural Networks , author=. Proceedings of the 34th International Conference on Machine Learning , pages=. 2017 , organization=

    2017

  22. [30]

    Information Fusion , volume=

    A Review of Uncertainty Quantification in Deep Learning: Techniques, Applications and Challenges , author=. Information Fusion , volume=. 2021 , publisher=

    2021

  23. [31]

    Artificial Intelligence Review , volume=

    A Survey of Uncertainty in Deep Neural Networks , author=. Artificial Intelligence Review , volume=. 2023 , publisher=

    2023

  24. [32]

    Advances in Neural Information Processing Systems , volume=

    Selective Classification for Deep Neural Networks , author=. Advances in Neural Information Processing Systems , volume=

  25. [33]

    2019 , organization=

    Geifman, Yonatan and El-Yaniv, Ran , booktitle=. 2019 , organization=

    2019

  26. [34]

    arXiv preprint arXiv:2207.05221 , year=

    Language Models (Mostly) Know What They Know , author=. arXiv preprint arXiv:2207.05221 , year=

    arXiv

  27. [35]

    Xiong, Miao and Hu, Zhiyuan and Lu, Xinyang and Li, Yifei and Fu, Jie and He, Junxian and Hooi, Bryan , booktitle=. Can

  28. [36]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , year=

    Just Ask for Calibration: Strategies for Eliciting Calibrated Confidence Scores from Language Models Fine-Tuned with Human Feedback , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , year=

    2023

  29. [37]

    Transactions on Machine Learning Research , year=

    Teaching Models to Express Their Uncertainty in Words , author=. Transactions on Machine Learning Research , year=

  30. [38]

    First Conference on Language Modeling (COLM) , year=

    The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets , author=. First Conference on Language Modeling (COLM) , year=

  31. [39]

    arXiv preprint arXiv:2310.02207 , year=

    Language Models Represent Space and Time , author=. arXiv preprint arXiv:2310.02207 , year=

    arXiv

  32. [40]

    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=. 2024 , doi=

    2024

  33. [41]

    Proceedings of the 33rd International Conference on Machine Learning , pages=

    Learning Physical Intuition of Block Towers by Example , author=. Proceedings of the 33rd International Conference on Machine Learning , pages=. 2016 , organization=

    2016

  34. [42]

    Advances in Neural Information Processing Systems , volume=

    Interaction Networks for Learning about Objects, Relations and Physics , author=. Advances in Neural Information Processing Systems , volume=

  35. [43]

    Chow, Wei and Mao, Jiageng and Li, Boyi and Seita, Daniel and Guizilini, Vitor and Wang, Yue , booktitle=

  36. [44]

    Fu, Chaoyou and Dai, Yuhan and Luo, Yondong and Li, Lei and Ren, Shuhuai and Zhang, Renrui and Wang, Zihan and Zhou, Chenyu and Shen, Yunhang and Zhang, Mengdan and others , booktitle=

  37. [45]

    arXiv preprint arXiv:2404.18930 , year=

    Hallucination of Multimodal Large Language Models: A Survey , author=. arXiv preprint arXiv:2404.18930 , year=

    arXiv

  38. [46]

    Bell , booktitle=

    Polina Kirichenko and Mark Ibrahim and Kamalika Chaudhuri and Samuel J. Bell , booktitle=

  39. [47]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Ego4d: Around the world in 3,000 hours of egocentric video , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=. 2022 , doi=

    2022

  40. [48]

    Rescaling Egocentric Vision: Collection, Pipeline and Challenges for

    Damen, Dima and Doughty, Hazel and Farinella, Giovanni Maria and Furnari, Antonino and Kazakos, Evangelos and Ma, Jian and Moltisanti, Davide and Munro, Jonathan and Perrett, Toby and Price, Will and Wray, Michael , journal=. Rescaling Egocentric Vision: Collection, Pipeline a...

    2022

  41. [49]

    OpenAI Technical Report , note=

    Video generation models as world simulators , author=. OpenAI Technical Report , note=

  42. [50]

    Polyak, Adam and Zohar, Amit and Brown, Andrew and Tjandra, Andros and Sinha, Animesh and Lee, Ann and Vyas, Apoorv and Shi, Bowen and Ma, Chih-Yao and Chuang, Ching-Yao and others , journal=

  43. [51]

    Bansal, Hritik and Lin, Zongyu and Xie, Tianyi and Zong, Zeshun and Yarom, Michal and Bitton, Yonatan and Jiang, Chenfanfu and Sun, Yizhou and Chang, Kai-Wei and Grover, Aditya , journal=

  44. [52]

    Bakhtin, Anton and van der Maaten, Laurens and Johnson, Justin and Gustafson, Laura and Girshick, Ross , booktitle=

  45. [53]

    Baradel, Fabien and Neverova, Natalia and Mille, Julien and Mori, Greg and Wolf, Christian , booktitle=

  46. [54]

    arXiv preprint arXiv:2308.10248 , year=

    Steering Language Models With Activation Engineering , author=. arXiv preprint arXiv:2308.10248 , year=

    arXiv

  47. [55]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Steering llama 2 via contrastive activation addition , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2024 , doi=

    2024

  48. [56]

    Bordes, Florian and Garrido, Quentin and Kao, Justine T and Williams, Adina and Rabbat, Michael and Dupoux, Emmanuel , journal=

  49. [57]

    2012 , organization=

    Todorov, Emanuel and Erez, Tom and Tassa, Yuval , booktitle=. 2012 , organization=

    2012

  50. [58]

    Cancer , volume=

    Index for rating diagnostic tests , author=. Cancer , volume=. 1950 , doi=

    1950

  51. [59]

    Psychological Bulletin , volume=

    Measuring nominal scale agreement among many raters , author=. Psychological Bulletin , volume=. 1971 , doi=

    1971

  52. [60]

    Concrete problems in

    Amodei, Dario and Olah, Chris and Steinhardt, Jacob and Christiano, Paul and Schulman, John and Man. Concrete problems in. arXiv preprint arXiv:1606.06565 , year=

    arXiv

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.