Pith. sign in

REVIEW 2 major objections 5 minor 18 references

Signed visual evidence extracted during a single prefill pass ranks wrong closed multimodal answers better than confidence alone, with a mean error-AP gain of 0.134.

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 03:37 UTC pith:ZRNYIV3Q

load-bearing objection A strong empirical detector paper with an honest scope statement and a real but contained caveat about the local readout's sign fidelity. the 2 major comments →

arxiv 2607.27667 v1 pith:ZRNYIV3Q submitted 2026-07-30 cs.CV

Witness Evidence Portfolios: Single-Prefill Risk Detection for Closed Multimodal Answers

Fexiang Liu , Shiye Wang , Qiang Qiu , Zheng Wang This is my paper
classification cs.CV
keywords multimodal hallucination detectionselective predictionerror rankingvisual evidence attributionattention readoutsingle-prefill risk scoringquestion-bound witnessesevidence concentration
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.

This paper aims to show that internal visual evidence—which image tokens push the model toward or against its predicted answer—carries risk information that confidence scores miss. The proposed detector, WEP, extracts this signed evidence from the same prefill pass that produces the answer, requiring no extra image views, decoding changes, backward passes, or external verifiers. It summarizes the evidence along two interpretable axes: whether it lands near question-relevant witnesses (provenance) and whether it is spatially concentrated or diffuse (concentration). Across three multimodal large language models and four binary-answer benchmarks, WEP improves mean error average precision by 0.134 over the negative-margin confidence baseline, with positive gains on all 12 model–dataset pairs. The result matters because it offers deployers a cheap, inspectable ranking signal for deciding which confident answers should be reviewed or routed.

Core claim

The central claim is that internal signed decision evidence—an estimate, layer by layer, of which visual tokens increase or decrease the predicted candidate's margin—contains risk information beyond the candidate margin alone. Concretely, WEP projects layer-wise attention/value readouts onto a one-vs-competitors direction to get signed source-token contributions, transports text-mediated contributions back into image space through question-token binding maps, and then summarizes the signed evidence with two route families: provenance (how much evidence overlaps question-bound witnesses relative to a uniform visual reference) and concentration (how narrowly or diffusely supportive or contradi

What carries the argument

The engine is the signed source-token readout C_{\ell,y}(s) = \sum_h \langle \tilde d_y, W_O^{h}(A^{h}(t,s)V^{h}(s)) \rangle, which estimates, at layer \ell, how much source token s pushes the readout toward candidate y versus its competitors. The paper proves an exact first-order completeness identity: summing over source tokens recovers the bias-free attention update's projection along the candidate direction, though this identity decomposes only a local directional derivative, not the full network. To make this spatial, question tokens define visual witness maps P_Q, S, O (with S=O=P_Q as fallback), and text-mediated contributions are transported to image space via question-token binding.

Load-bearing premise

The whole method rests on the assumption that the per-layer estimate of which image tokens push for or against the predicted answer is a true reflection of why the model answered that way; if a layer's local readout diverges from the final decision evidence, the provenance and concentration signals lose their meaning and the AP gain could vanish on other models.

What would settle it

On a held-out transformer MLLM, compute the signed source-token readout at each layer, patch the top-supporting and top-contradicting visual tokens back into that same layer, run the remaining blocks, and compare the sign of the final-logit margin change against the readout sign. If the sign-agreement rate falls well below the 94.6–97.2% range the paper reports for its three backbones, or if the evidence-shuffle null (permuting readouts across images while fixing confidence and labels) still yields positive AP gains, the local readout is not tracking final decision evidence and WEP's central c

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

If this is right

  • Deployments can rank closed answers by error risk using the exact same forward pass that produced the answer, with no second model, image perturbation, or backward pass.
  • At fixed review budgets of 1%, 5%, and 10% of the queue, WEP concentrates errors earlier: average review precision improves by roughly +0.196, +0.206, and +0.179, with image-cluster intervals that remain positive.
  • Provenance and concentration encode complementary failure modes; nested family selection (+0.134) beats either family alone (+0.129 and +0.124) and a joint route pool (+0.128).
  • In the zero-target-label leave-dataset-out setting, fixing the provenance family still transfers (+0.095 mean AP over confidence), so the signal is not overfit to one benchmark's label statistics.
  • The final score remains inspectable: each selected route is a named layer/sign scalar, so a reviewer can see which evidence channel pushed an answer up or down the queue.

Where Pith is reading between the lines

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

  • We infer that the same signed-evidence construction can extend to open-ended generation if free-form answers are first segmented into verifiable claims, each assigned a finite candidate set; the paper only tests closed finite candidates.
  • We infer that the concentration family could serve as a general-purpose 'evidence diffusion' diagnostic for other structured inputs, such as video or audio tokens, where spatial or temporal spread of contradicting evidence might flag hallucinations.
  • We infer that WEP's routes could be used as a training signal for grounding objectives: penalizing answers whose signed evidence provenance misses question-bound witnesses may reduce hallucination at the source.
  • We infer that the method's success hinges on the local readout approximation holding across architectures; if future backbones show lower closure fidelity, the provenance route may need layer-specific recalibration.

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

2 major / 5 minor

Summary. This paper proposes Witness Evidence Portfolios (WEP), a single-prefill white-box risk detector for closed multimodal answers. From one instrumented forward pass, WEP computes per-layer signed contributions of source tokens to the predicted candidate margin (Eq. 4), transports these into visual space, and builds two route families: question-bound provenance (global overlap and question-residual) and signed concentration. A nested image-grouped validation selects one family, a sparse top-k route set, and a single fusion weight, all fitted on a calibration slice and evaluated out-of-fold. On three MLLMs and four binary-answer benchmarks, WEP improves mean error AP by 0.134 over candidate confidence, with positive gains on all 12 model–dataset pairs and strictly positive image-cluster bootstrap intervals on 10 pairs. The paper includes extensive controls: 50-seed algorithmic stability, leave-dataset-out transfer, evidence shuffles, label-permutation nulls, local closure audits, end-to-end patching, and comparisons with matched-budget learned selectors.

Significance. If the reported gains hold, WEP offers a practical and unusually well-validated signal for auditing, abstaining, or routing closed multimodal answers without perturbing the input or changing decoding. The evaluation is a model of care: predictions are evaluated image-grouped out-of-fold; uncertainty is quantified by image-cluster bootstrap and 50-seed algorithmic reselection; the evidence shuffle and label-permutation nulls check for artifacts; and transfer is tested without target labels. The ablations isolate the contribution of signs, transport, question residuals, and family selection. The main risk is interpretive: the method's headline narrative—signed visual evidence that supports or contradicts the answer—depends on the local readout C_{ℓ,y}(s), and the paper's own patching audit shows only partial and depth-dependent sign fidelity to final logits. This does not undermine the ranking result (the signed-vs-unsigned gap is large), but it requires either more evidence or a more cautious framing.

major comments (2)
  1. [Eqs. (4)–(5), 'Estimated Decision Evidence'; Table 20] Central claim that WEP identifies tokens supporting/contradicting the prediction rests on sign of C_{ℓ,y}(s). Eq. (5) is exact only for the bias-free attention update; the text concedes it is a local directional derivative, not the full network. Table 17 establishes local fidelity, but Table 20's end-to-end patching shows final-margin effects in only 13/18 fixed depth tests, with sign reversals at some depths (InternVL AMBER-D: L24 −0.048, L30 −0.030, L35 +0.010). Thus per-token, per-layer sign fidelity to final logits is not established, while route names, concentration semantics, and Fig. 6 audits depend on it. The ranking result and signed-vs-unsigned ablation (+0.134 vs +0.014) survive, but the mechanism claim overreaches. Please add a per-token ablation-versus-final-margin check for selected layers or explicitly scope claims to 'locally signed readouts.'
  2. [Supplementary Table 12; 'Calibration and Provenance-Route Stability'] WEP's stated advantage includes 'sparse, named routes' that make the risk score inspectable. However, Table 12 shows the selected provenance route sets have low pairwise Jaccard overlap across held-out folds (0.424 for Qwen, 0.567 overall; 25.7 unique routes on average), and Table 13 shows substantial variation in family choice and residual slots under 50 seeds. The final score is stable, but the specific named routes are not. If auditability is a contribution, the paper should either report route-level stability (e.g., frequency of each route in the selected portfolios) or soften the inspectability claims. Without this, the reader cannot tell whether a reported 'dominant route' for an example is a reproducible signal or a fold-specific artifact.
minor comments (5)
  1. [Abstract; Table 1] The abstract's 'all 12 gains positive' refers to the full-label, nested-selection setting. Low-label (+0.069/+0.107 at 100/250 labels) and zero-target transfer (+0.095) are smaller. Consider adding a qualifier to the abstract.
  2. [Table 8] The rows 'No subject/object split' and 'Question-only' are numerically identical in many cells; clarify whether they are the same ablation under different names or whether the difference is too small to be visible.
  3. [Table 20] Specify what fraction of total layers the early/middle/late choices correspond to for each backbone; the layer indices (L21–L35) are not directly comparable across models.
  4. [Related Work] Consider replacing the LessWrong reference 'Riggs, Nanda et al. 2024' with a peer-reviewed publication on logit-lens decompositions, if one exists.
  5. [Supplement, 'Exact operational definitions'] The fixed hyperparameters λ_sink=1.0 and the evidence-attenuation floor 0.05 are not varied; a small sensitivity analysis would help rule out tuning artifacts.

Circularity Check

0 steps flagged

No significant circularity: the risk score is built from forward-pass evidence features and evaluated out-of-fold; no equation reduces the predicted risk to the fitted labels.

full rationale

WEP's derivation chain is self-contained. The evidence estimate C_{l,y}(s) in Eq. (4) is defined from attention weights, value vectors, and an unembedding direction; Eq. (5) is an exact algebraic identity for the local attention update and is explicitly qualified by the paper: 'This identity decomposes a local directional derivative, not the complete nonlinear network.' The final risk score in Eq. (12) is a standardized fusion of confidence with a mean of selected route scalars; selection of the route family, portfolio size k, and fusion weight beta is performed on calibration folds, and all reported AP values are out-of-fold. Thus the predicted error ranking is not the fitting label by construction. The paper also provides evidence-shuffle, label-permutation, and leave-dataset-out transfer controls that separate the AP gain from score-scale artifacts or route-search freedom. The skeptical concern that C may be only a proxy for final decision evidence is a validity/robustness question, not a circularity in the derivation; the paper separately tests this with end-to-end patching and reports mixed but nonzero effects. No load-bearing self-citation chain or imported uniqueness theorem is used to force the method's choice.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 2 invented entities

The method is a supervised calibration pipeline built on label-free white-box readouts. Most free parameters are selected on a labeled calibration split and evaluated out-of-fold; the paper's controls show the gains are not purely selection artifacts. The axioms are interpretability assumptions about attention/value readouts, not standard mathematical axioms, and are only partially addressed by intervention tests.

free parameters (6)
  • fusion weight β = positive; selected from grid {-3.00,...,3.00} step 0.25; all 60 final selections positive
    Chosen on the outer calibration fold; one scalar controls confidence-evidence fusion in r_wep.
  • portfolio size k = selected from {1,2,3,5,8,13,21,36}, exact per-fold values not reported
    Number of evidence routes averaged; selected by calibration AP and affects score variance.
  • route family F* = provenance 2094/3000 and concentration 906/3000 reseeded outer folds
    Discrete model-selection choice using inner grouped validation; when target labels are unavailable, provenance fallback is used.
  • top-k route set P_k = mean 15.6 selected routes for provenance; residual slots 47.1% under 50 seeds
    Route identities chosen by per-route error AP on calibration; the final risk is their standardized mean.
  • entity sink subtraction weight λ_sink = 1.0
    Hand-set constant in entity witness map construction; no sensitivity analysis reported.
  • evidence attenuation floor = 0.05
    Hand-set lower bound when attenuating evidence by control-token binding; not tuned.
axioms (6)
  • domain assumption Additive attention/value readout faithfully decomposes candidate-margin decision evidence.
    Eqs. (4)-(5) define C_{ℓ,y}(s); the paper states it is a local directional derivative, not the full nonlinear network; patch tests partially justify.
  • domain assumption Question-token attention/value binding marks where the question is visually referenced.
    Witness maps P_Q, S, O are built from B_j(v); validated only on VSR boxes, not on the other three benchmarks.
  • domain assumption Source-to-visual transport maps text-token evidence into image space without distorting sign.
    Eq. (6) transports question-token contributions through binding distributions and nulls other text mass; it is a construction, not a derived result.
  • domain assumption Benchmark labels in the calibration slice are correct and the closed yes/no set is the decision space.
    Evaluation labels e(x)=1[ŷ≠y*] drive route selection and AP; noisy labels would bias all results.
  • domain assumption Histogram intersection of evidence with witness maps measures support/contradiction.
    Eqs. (7)-(8) use min-overlap as support/contradiction scores; this is a heuristic similarity, not a theory-derived measure.
  • domain assumption Uniform visual reference is an appropriate baseline for generic evidence overlap.
    U=1/|V| and the residual r_Δ = r_Q - r_G assume that subtractive normalization isolates question-specific alignment.
invented entities (2)
  • Signed visual evidence maps E_{ℓ,y}(v) independent evidence
    purpose: Represent per-layer, per-candidate visual support or contradiction of the predicted margin.
    Validated by closure error (<0.1%), rank correlation >0.975, and value-patching final-margin effects.
  • Question-bound witness maps P_Q, S, O independent evidence
    purpose: Localize question- or entity-bound visual regions for provenance scoring.
    Validated on Grounded-VSR boxes: point accuracy 0.538–0.776 vs. random 0.355–0.393.

pith-pipeline@v1.3.0-daily-deepseek · 27894 in / 17051 out tokens · 176042 ms · 2026-08-01T03:37:25.406232+00:00 · methodology

0 comments
read the original abstract

Reliable deployment of multimodal large language models (MLLMs) requires deciding whether a confident visual answer should be trusted, reviewed, or routed to a stronger system. Confidence scores capture candidate margins, but not where the estimated signed visual readouts associated with those margins come from or how they are distributed. We study inference-time risk detection for closed visual answers using the same white-box prefill path that produces the answer. Witness Evidence Portfolios (WEP) first estimates, layer by layer, which visual contributions support or contradict the predicted candidate. It summarizes these contributions through two interpretable route families: question-related evidence provenance and signed evidence concentration. Nested grouped validation chooses the more reliable family and a sparse top-k route portfolio, which is fused with candidate confidence. WEP needs no image perturbation, decoding change, backward pass, or external verifier. Across three MLLMs and four binary-answer benchmarks, WEP improves mean error AP by 0.134. All 12 model--dataset gains are positive, and image-cluster bootstrap intervals are strictly positive on 10 pairs. WEP targets white-box closed-answer systems and uses a labeled calibration slice.

Figures

Figures reproduced from arXiv: 2607.27667 by Fexiang Liu, Qiang Qiu, Shiye Wang, Zheng Wang.

Figure 1
Figure 1. Figure 1: Overview of Witness Evidence Portfolios (WEP). One instrumented prefill path supplies candidate logits and sparse [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Error-ranking and review-budget gains. Left: AP gains over confidence with 95% image-cluster bootstrap intervals. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Calibration-label efficiency. (a) Mean AP gain over confidence as the number of calibration labels increases. (b) [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Zero-target-label leave-dataset-out transfer. Each [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Provenance-family robustness and evidence-sanity diagnostics. (a) Seed sensitivity over 50 image-grouped fold seeds; [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative evidence and route contributions. Stored provenance cases show the input, witness, signed evidence, overlap, [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] 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

18 extracted references · 6 linked inside Pith

  1. [4]

    Transformer Circuits Thread

    AMathematicalFrameworkforTransformerCircuits. Transformer Circuits Thread. Favero, A.; Zancato, L.; Trager, M.; Choudhary, S.; Perera, P.;Achille,A.;Swaminathan,A.;andSoatto,S.2024. Multi- ModalHallucinationControlbyVisualInformationGround- ing. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. Gal, Y.; and Ghahramani, Z

    2024

  2. [5]

    Point” is the fraction of examples whose maximum-witness tokencenterliesinsidethetargetbox,comparedwiththebox-arearandomexpectation.“Mass

    These counts should not bereadastherawsizesoftheupstreambenchmarkreleases. Expanded Main Results Table5expandsthecompactmain-papersummarywithper- cell review precision alongside the final nested AP results. Metric Confidence WEP∆ AURC↓0.214 0.179 -0.035 E-AURC↓0.147 0.112 -0.035 Acc.@90% coverage↑0.707 0.727 +0.020 Error recall@5% review↑0.096 0.135 +0.03...

    2026

  3. [6]

    InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    HallusionBench: An Advanced Di- agnostic Suite for Entangled Language Hallucination and Visual Illusion in Large Vision-Language Models. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Guo,C.;Pleiss,G.;Sun,Y.;andWeinberger,K.Q.2017. On Calibration of Modern Neural Networks. InProceedings of the International Conferen...

    2017

  4. [7]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive De- coding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Li, Y.; Du, Y.; Zhou, K.; Wang, J.; Zhao, W. X.; and Wen, J.-R.2023. EvaluatingObjectHallucinationinLargeVision- Language Models. InProceedings of the 2023 Conference o...

    2023

  5. [9]

    InFindings of the As- sociation for Computational Linguistics: ACL 2025,14949– 14965.Vienna,Austria:AssociationforComputationalLin- guistics

    VADE: Visual Attention Guided Hal- lucinationDetectionandElimination. InFindings of the As- sociation for Computational Linguistics: ACL 2025,14949– 14965.Vienna,Austria:AssociationforComputationalLin- guistics. Riggs, L.; Nanda, N.; et al

    2025

  6. [10]

    LessWrong

    Logit Prisms: Decom- posingTransformerOutputsforMechanisticInterpretability. LessWrong. Rohrbach, A.; Hendricks, L. A.; Burns, K.; Darrell, T.; and Saenko,K.2018. ObjectHallucinationinImageCaptioning. InProceedings of the Conference on Empirical Methods in Natural Language Processing. Sarkar,A.;Li,Y.;Cheng,J.;Mishra,S.K.;andVasconcelos, N

    2018

  7. [12]

    In Findings of the Association for Computational Linguistics: ACL 2024, 14379–14391

    Unsupervised Real-Time Hallucination Detection based on the Internal States of Large Language Models. In Findings of the Association for Computational Linguistics: ACL 2024, 14379–14391. Sundararajan, M.; Taly, A.; and Yan, Q

    2024

  8. [13]

    Wang, P.; Xiao, T.; He, T.; Han, Z.; Zhang, Z.; and Shou, M

    AMBER: An LLM-Free Multi- Dimensional Benchmark for MLLMs Hallucination Eval- uation.arXiv preprint arXiv:2311.07397. Wang, P.; Xiao, T.; He, T.; Han, Z.; Zhang, Z.; and Shou, M. Z

    Pith/arXiv arXiv

  9. [14]

    Wang,X.;Pan,J.;Ding,L.;andBiemann,C.2024

    A Survey of Multimodal Hallucination Evalua- tion and Detection.arXiv preprint arXiv:2507.19024. Wang,X.;Pan,J.;Ding,L.;andBiemann,C.2024. Mitigat- ing Hallucinations in Large Vision-Language Models with Instruction Contrastive Decoding. InFindings of the Asso- ciation for Computational Linguistics: ACL 2024, 15840– 15853. Wang, Z.; Aniri; Chen, T.; Zhang...

    arXiv 2024

  10. [16]

    arXiv preprint arXiv:2310.16045

    Woodpecker: Halluci- nation Correction for Multimodal Large Language Models. arXiv preprint arXiv:2310.16045. Zhang, C.; Sun, C.; Jiang, X.; Li, W.; and Tian, X. 2026a. Prefill-Time Intervention for Mitigating Hallucination in Large Vision-Language Models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 25293–25303. Zh...

    Pith/arXiv arXiv

  11. [17]

    InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition Findings, 8897–8906

    Attention Never Lie: Visual Attention Defocus Reveals and Rectifies Hallucina- tions in MLLMs. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition Findings, 8897–8906. Zhao,J.;Zhang,F.;Sun,X.;Feng,C.;andTan,Z.2025. Tell ModelWheretoLook:VisualGroundingAttentionImproves VLMs’ Perception.arXiv preprint arXiv:2511.20032. Sup...

    Pith/arXiv arXiv 2025

  12. [2016]

    InProceedings of the International Conference on Machine Learning

    Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning. InProceedings of the International Conference on Machine Learning. Gautam,S.;Riegler,M.A.;andHalvorsen,P.2025.HEDGE: Hallucination Estimation via Dense Geometric Entropy for VQA with Vision-Language Models.arXiv preprint arXiv:2511.12693. Geifman,Y.;andEl-Yaniv,R.2017. S...

    arXiv 2025

  13. [2019]

    InProceedings of the Conference on Empirical Methods in Natural Language Processing

    Attention is not not Explanation. InProceedings of the Conference on Empirical Methods in Natural Language Processing. Xu,Z.;Wang,Z.;Wu,J.;Lu,J.;andWang,X.2025. Causal- HalBench:UncoveringLargeVision-LanguageModels’Ob- ject Hallucinations Through Causal Intervention.arXiv preprint arXiv:2511.10268. Yin,S.;Fu,C.;Zhao,S.;Xu,T.;Wang,H.;Sui,D.;Shen,Y.; Li, K....

    arXiv 2025

  14. [2021]

    Bai, S.; et al

    A Gentle Intro- duction to Conformal Prediction and Distribution-Free Un- certainty Quantification.arXiv preprint arXiv:2107.07511. Bai, S.; et al

    Pith/arXiv arXiv

  15. [2023]

    InTransactions of the Association for Computa- tional Linguistics

    Visual Spatial Reasoning. InTransactions of the Association for Computa- tional Linguistics. Liu, H.; Li, C.; Li, Y.; and Lee, Y. J. 2024a. Improved Baselines with Visual Instruction Tuning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Liu,H.;Li,C.;Wu,Q.;andLee,Y.J.2023.VisualInstruction Tuning.Advances in Neural In...

    2023

  16. [2024]

    InInternational Conference on Learning Representations

    INSIDE: LLMs’ Internal States Re- tain the Power of Hallucination Detection. InInternational Conference on Learning Representations. Chen,Z.;etal.2025. InternVL3.5:AdvancingOpen-Source MultimodalModelsinVersatility,Reasoning,andEfficiency. arXiv preprint arXiv:2508.18265. Dancette, C.; Whitehead, S.; Maheshwary, R.; Vedantam, R.;Scherer,S.;Chen,X.;Cord,M....

    Pith/arXiv arXiv 2025

  17. [2025]

    Belrose,N.;Furman,Z.;Smith,L.;Halawi,D.;Ostrovsky,I.; McKinney, L.; Biderman, S.; and Steinhardt, J

    Qwen3-VL Technical Report.arXiv preprint arXiv:2511.21631. Belrose,N.;Furman,Z.;Smith,L.;Halawi,D.;Ostrovsky,I.; McKinney, L.; Biderman, S.; and Steinhardt, J

    Pith/arXiv arXiv

  18. [2026]

    InInternational Confer- ence on Learning Representations

    Leveraging Data to Say No: Memory Augmented Plug-and-PlaySelectivePrediction. InInternational Confer- ence on Learning Representations. Selvaraju, R. R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh,D.;andBatra,D.2017. Grad-CAM:VisualExplana- tionsfromDeepNetworksviaGradient-BasedLocalization. InProceedings of the IEEE International Conference on Computer ...

    2017