Pith. sign in

REVIEW 2 major objections 4 minor 45 references

This paper claims naturalistic presentation edits—lighting, background, staging—can more than double the chance a vision-language model selects an image, across product, hiring, housing, and hotel tasks.

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 ↗

Naturalistic edits to background, lighting, and staging systematically shift the choices of frontier vision-language models, and a new optimization-plus-interpretation framework (CVPO) discovers and explains the visual themes that drive those shifts.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Large, replicated effect of image edits on VLM choices, but the 'presentation-only' claim rests on an unenforced identity constraint; the paper needs a post-hoc identity check and a relabeled conclusion. the 2 major comments →

arxiv 2602.15278 v2 pith:JCUMHOG5 submitted 2026-02-17 cs.CV cs.AI

Visual Persuasion: What Influences Decisions of Vision-Language Models?

Manuel Cherep , Pranav M R , Pattie Maes , Nikhil Singh This is my paper
classification cs.CV cs.AI
keywords vision-language modelsvisual prompt optimizationrevealed preferencevisual utilityagentic choiceimage editsVLM auditingbehavioral evaluation
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 reading

This paper tries to establish that vision-language models used as decision-making agents have systematic visual preferences that can be measured and manipulated by naturalistic image edits. The authors place VLMs in pairwise choice tasks across product, hiring, housing, and hotel domains, then iteratively modify images with an editing model guided by model feedback. They report that even zero-shot edits move choice probabilities far from chance, and further optimization increases selection probability by roughly 0.1–0.3; recurring themes such as pools, twilight lighting, and business attire emerge. The work matters because the same pipeline can both expose and exploit vulnerabilities in image-based AI agents before they are used at scale.

Core claim

On the paper's own terms, the central discovery is that visual presentation alone—holding the underlying object or scene fixed—has a large, consistent effect on VLM agentic choice. In controlled head-to-head comparisons, zero-shot edits can more than double the probability that an image is selected, and optimization adds further gains; a novel competitive optimization method (CVPO) wins most comparisons against two adapted baselines. An automatic interpretability pipeline identifies stable themes (pools, fire pits, luxury furniture, twilight lighting, formal attire) that drive these decisions, and human participants show similar preferences in most tasks. The authors also report that asking

What carries the argument

The central object is visual prompt optimization: instead of perturbing pixels, the authors optimize an editable text prompt parameterizing an image-editing operator, treating the VLM's decision function as a latent utility landscape inferred from pairwise comparisons. The load-bearing device is the identity-maintenance constraint—edits must depict the same entity or scene up to nuisance variation—intended to separate presentation from substance, but implemented only as a verbal instruction to the editing model. Their new competitive method, CVPO, stages head-to-head contests among candidate edits judged by a panel of VLMs with consistency checks, generates challengers from the losing image,

Load-bearing premise

The whole result depends on the claim that edits change only presentation, not substance—but the paper never verifies this; it only tells the editing model to keep the object, house, hotel, or person the same, and some of its own discovered themes (pools, fire pits, suits) look like substantive changes.

What would settle it

Have human raters classify each optimized edit as presentation-only (lighting, angle, background, framing) versus substance-changing (added pool, outdoor kitchen, furniture, business attire, changed amenity), then re-run the head-to-head comparisons using only the presentation-only subset, with edits that violate a strict identity check removed. If the large choice-probability shifts disappear or shrink to noise, the claim that presentation alone drives VLM choices is falsified.

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

If this is right

  • If the central claim holds, standard accuracy benchmarks miss a major behavioral risk: a model can be 'correct' about what an image shows while still being steered by irrelevant presentation details.
  • Optimization procedures like CVPO give auditors a tool to discover hidden visual preferences proactively rather than waiting for real-world manipulation.
  • The recurring themes imply concrete, transferable edit recipes that could be applied zero-shot to new images, amortizing the cost of optimization.
  • Because normalization only partially removes the effect, deployment of VLM agents in high-stakes visual decisions (hiring, real estate, purchasing) should include explicit presentation-robustness checks.
  • The human results suggest the discovered preferences are not purely an artifact of one model family, so the framework could inform the study of visual persuasion more broadly.

Where Pith is reading between the lines

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

  • Inference: If identity were enforced by a real post-hoc check rather than a verbal instruction, the measured effects would likely shrink; separating lighting/background-only edits from amenity edits (pools, furniture, clothing) is a direct test the paper did not run.
  • Inference: The convergence of themes across optimization methods may partly reflect properties of the specific edit and judge models used in the pipeline, so the discovered 'preferences' should be replicated across model families before treating them as general VLM truths.
  • Inference: The same loop could be used adversarially by marketplace actors to differentially advantage listings without changing substantive quality, giving a measurable 'machine fluency' advantage to those who know the tricks—an extension the paper notes only as a risk.
  • Inference: A cheap practical extension is to use the distilled theme prompts as a red-teaming checklist for any image-fed agent, since they require no per-image optimization.
Share X Bluesky LinkedIn Reddit HN

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 / 4 minor

Summary. The paper introduces a framework for studying the visual preferences of vision-language models (VLMs) by iteratively editing images with a text-to-image model and measuring changes in pairwise choice probabilities. The authors propose three optimization methods (VTG, VFD, and the novel CVPO), evaluate them on nine frontier VLMs across four realistic agentic tasks (product purchasing, hiring, house search, hotel booking), and find that zero-shot edits and optimized edits substantially increase the probability of an image being chosen. They also run human experiments showing similar effects, develop an automatic interpretability pipeline that identifies recurring visual themes, and test an image-normalization mitigation strategy that partially reduces the effect. The central claim is that visual presentation alone—while keeping semantic content constant—can systematically steer VLM decisions.

Significance. If the central claim were fully established, this would be a valuable contribution: a scalable, interpretable methodology for auditing and red-teaming image-based AI agents, plus a new optimization algorithm (CVPO) that outperforms adapted baselines. The paper's strengths are its large scale (1.8M+ API requests, 9 VLMs, 4 tasks), its careful pairwise-choice design (order counterbalancing, filtering of order-inconsistent judgments, cluster-robust standard errors, Benjamini-Hochberg corrections), and the inclusion of human validation showing that the edited images also shift human choices. The zero-shot effects are particularly clean because they do not involve an optimization loop. However, the headline interpretation that effects occur 'even when primary semantic content is held constant' is currently undercut by the paper's own admitted limitations about the identity-preserving constraint.

major comments (2)
  1. [§3.2, Definition 3.1; §5 Limitations] The central claim that 'visual presentation alone has a large effect' (§4.1) and 'even when primary semantic content is held constant' (§6) rests entirely on the identity-preserving constraint C(x0) in Eq. (6). The paper states that the identity predicate I can be 'approximated via simple verbal instructions' and that 'simple instructions suffice,' but no post-hoc check, similarity threshold, or verification of I is reported. The auto-interpretability results (Appendix C) show that the optimizer discovers edits such as adding pools, outdoor kitchens, and business suits; the limitations section concedes that 'the boundary between presentation and substance can be fuzzy' and that amenity changes 'may offer some increased utility.' Therefore the empirical finding that edited images shift choices is robust, but the interpretation that this occurs 'even when primary semantic content is held c
  2. [§3.4, §4.1, Table 1] The optimization loop uses Gemini 3 Flash as the judge model for all three methods (stated in §3.4: 'we use Gemini 3 Flash as the judge model in all optimization pipelines'), and Gemini 3 Flash is also one of the nine evaluator VLMs in the main results. Consequently, for this model, the optimized images are selected to satisfy the same model that later evaluates them; the gains from optimization are partly in-sample. The zero-shot results are unaffected, but the claim that optimized images shift choices 'across all evaluators' is stronger than what is demonstrated. I recommend reporting all main contrasts with the optimizer-judge model excluded, or at least a sensitivity analysis; given Table 1, the aggregate conclusion is likely robust, but the point estimate for Gemini 3 Flash and the method-comparison (CVPO vs. VFD) should be labeled as in-sample for that model.
minor comments (4)
  1. [Appendix E.3] Typo: 'oberservations' should be 'observations.'
  2. [§4.2] The efficiency metric 'Budget %' measures iterations only, but CVPO generates more images per iteration. Consider reporting API cost or total images generated to make the efficiency comparison fair across methods.
  3. [§3.2] The notation for the number of judges k, challengers K, and mitigation passes κ is easy to confuse. Consider renaming to avoid k/K/κ collision.
  4. [§5] The limitations paragraph on identity maintenance is appropriately candid but should be moved earlier or cross-referenced in §4.1 where the 'presentation alone' claim is first made, so readers can immediately calibrate the strength of the claim.

Circularity Check

2 steps flagged

Partial circularity: Gemini 3 Flash is both the optimization judge and an evaluated model, so its CVPO win is in-sample; the 'presentation-only' interpretation rests on an unverified verbal identity constraint rather than a measured predicate. Cross-model zero-shot results remain independent.

specific steps
  1. fitted input called prediction [§3.3 (Specific Optimization Methods) and Table 1; see also Algorithm 1]
    ""Note: we use Gemini 3 Flash as the judge model in all optimization pipelines." Table 1 row: "Gemini 3 Flash ... CVPO 0.761"."

    All three optimizers (CVPO, VFD, VTG) use Gemini 3 Flash's pairwise judgments as the acceptance/selection signal (Algorithm 1: "k judges vote via pairwise comparisons with consistency checks"; VFD: evaluator J; VTG: LLM critic). The final edited images for Gemini 3 Flash are therefore selected to maximize that model's own preference. Reporting Gemini 3 Flash's 0.761 preference for CVPO's final images as evidence that optimization shifts VLM choices is an in-sample measure of the fitted objective, not an independent prediction. The non-circular evidence is the zero-shot effect on all models and the optimized effects on the eight evaluator models never used as judge or editor.

  2. self definitional [§3.2 (Definition 3.1, identity constraints) + §5 Limitations + Appendix C]
    ""The predicate I can be implemented by post-hoc checks (e.g. prompting a VLM to verify invariants and reject violating proposals), similarity thresholds, or approximated via simple verbal instructions in the editing instructions. In practice, with a sufficiently controllable editing model Edit, we find that simple instructions suffice." §5: "the boundary between presentation and substance can be fuzzy in some instances"."

    The paper's conclusion that edits shift choices "even when primary semantic content is held constant" (Conclusion) treats the editing instruction "Keep the product itself exactly unchanged" as if it implemented the identity predicate I(x,x0). But I is never computed: there is no post-hoc check or similarity threshold, only a verbal prompt. Appendix C shows the optimizer added pools, fire pits, outdoor kitchens, business suits, and "Product internal content exposure"—plausibly substantive changes. Thus the "presentation-only" interpretation is an assumption embedded in the prompt text, not a measured property of the images; the headline interpretive claim reduces to the unverified identity instruction rather than to an enforced constraint.

full rationale

The central empirical result—zero-shot edits shift choices on all nine VLMs, and optimization generalizes to the eight models not used as judge—is not a fitted value: those are out-of-sample measurements, and the human studies (N=154) are independent external evidence. However, two load-bearing pieces are circular. First, the Gemini 3 Flash row in Table 1 measures the exact objective that all three optimizers were maximizing (Gemini judged every proposal), so that row's CVPO advantage and optimized-vs-original shift are in-sample by construction. Second, the interpretation that effects are due to "visual presentation alone" depends on the identity predicate I(x,x0), which is never computed; the paper substitutes a verbal instruction to the editor for the constraint and then concludes "even when primary semantic content is held constant." Appendix C themes (pools, outdoor kitchens, business suits, revealing product contents) show the edits plausibly cross into substantive changes, and §5 concedes the boundary is fuzzy and that amenity changes "may offer some increased utility." Thus the presentation-only interpretation reduces to the unverified prompt text. This is partial circularity rather than total: the aggregate zero-shot/cross-model evidence and the human results do not depend on the circular steps, so the score is 6, not 8 or 10.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 1 invented entities

No physics-style free parameters or invented entities appear. The load-bearing weight sits on the identity constraint (an axiom, not a parameter) and on the judge-model choice. The empirical quantities are statistical estimates from 1.8M+ API calls, so the honest ledger is dominated by modeling assumptions rather than fitted constants.

free parameters (2)
  • Optimization hyperparameters (k judges, challenger count K, patience, iteration bounds T_min/T_max, equilibrium threshol = k=3; patience=3; T_min=10; T_max=30; epsilon symbolic; kappa in {0,1,3}
    Chosen by hand and disclosed in §3.3-3.6 and Algorithm 1. They tune the search, but no central number in the paper's claims depends on a fitted value of these; they are conventional algorithm constants.
  • Hand-authored base-prior prompts p0, judge prompts, and optimizer instructions per task = per-task text, Appendix H
    These define 'appealing' and therefore shape which themes are discovered (e.g., 'luxury,' 'golden hour'). Not fitted to data, but the authors' choice of wording is a degree of freedom that influences the results.
axioms (4)
  • domain assumption Pairwise choices reveal a scalar latent utility via sigma(U_tau(x_a) - U_tau(x_b)) (Eq. 3).
    Bradley-Terry/Luce model in §3.2; standard in choice modeling but assumes transitivity and scalar-valued preferences behind noisy VLM judgments.
  • ad hoc to paper Verbal instruction to the editor suffices to keep edits inside the identity-preserving set C(x0).
    §3.2 'Identity constraints': 'we use Nano Banana, which generally obeys this constraint expressed via prompting... we find that simple instructions suffice.' The identity predicate I(.,.) is never actually computed. The paper's presentation-vs-substance interpretation depends on this.
  • domain assumption Order-inconsistent evaluator judgments are noise and may be discarded.
    §3.4: reversed-order disagreement is marked 'inconsistent' and excluded. If inconsistency carries information (weak preference, indecision), excluding it inflates measured effect sizes.
  • domain assumption Gemini 3 Flash judge preferences proxy for the preferences of other frontier VLMs.
    All optimization uses Gemini 3 Flash as the judge (§3.3). Whether discovered themes generalize is then an empirical claim; it holds in aggregate on the 8 other models, but per-model CVPO rankings partially inherit judge-specific tastes.
invented entities (1)
  • Latent visual utility landscape U_tau over images independent evidence
    purpose: Modeling device for the agent's decision function in §3.2; the optimization objective is stated as max U_tau(x(p)).
    A standard revealed-preference construct whose only observable consequences are the choice probabilities the paper measures; falsifiable in principle, but it adds no mechanistic content beyond the Bradley-Terry link assumed in Eq. 3.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Visual Persuasion: What Influences Decisions of Vision-Language Models?." pith.science (2026). https://pith.science/paper/JCUMHOG5

@misc{pith2026260215278,
  author       = {Pith},
  title        = {Pith review of: Visual Persuasion: What Influences Decisions of Vision-Language Models?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JCUMHOG5}},
  note         = {Machine review of arXiv:2602.15278}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The web is littered with images, once created for human consumption and now increasingly interpreted by agents using vision-language models (VLMs). These agents make visual decisions at scale, deciding what to click, recommend, or buy. Yet, we know little about the structure of their visual preferences. We introduce a framework for studying this by placing VLMs in controlled image-based choice tasks and systematically perturbing their inputs. Our key idea is to treat the agent's decision function as a latent visual utility that can be inferred through revealed preference: choices between systematically edited images. Starting from common images, such as product photos, we propose methods for visual prompt optimization, adapting text optimization methods to iteratively propose and apply visually plausible modifications using an image generation model (such as in composition, lighting, or background). We then evaluate which edits increase selection probability. Through large-scale experiments on frontier VLMs, we demonstrate that optimized edits significantly shift choice probabilities in head-to-head comparisons. We develop an automatic interpretability pipeline to explain these preferences, identifying consistent visual themes that drive selection. We argue that this approach offers a practical and efficient way to surface visual vulnerabilities, safety concerns that might otherwise be discovered implicitly in the wild, supporting more proactive auditing and governance of image-based AI agents.

Figures

Figures reproduced from arXiv: 2602.15278 by Manuel Cherep, Nikhil Singh, Pattie Maes, Pranav M R.

Figure 1
Figure 1. Figure 1: Simplified overview of the iterative visual optimization process through feedback-driven prompt refinement. An original image is progressively improved over K rounds. Each iteration, judges provide feedback with possible improvements, and an LLM uses the feedback to generate editing instructions. These instructions are applied with an image generation model to produce the candidate for the next round. The … view at source ↗
Figure 2
Figure 2. Figure 2: Estimated marginal mean probability of choice by task (columns) × optimization method (rows) and optimization stage (X-axis; original image, zero-shot modified, and final after opti￾mization). Results are averaged across all VLMs. creases selection probability by roughly 0.2–0.4 relative to the original images, with all zero-shot versus original contrasts statistically significant. These shifts are substan… view at source ↗
Figure 3
Figure 3. Figure 3: Estimated marginal mean probability of choice for final optimized images produced by different optimization methods in head-to-head comparisons. Results are averaged across all VLMs with error bars showing 95% confidence intervals. VLM preferences most often favor CVPO In direct com￾parisons between the final outputs of each optimization method, CVPO wins most often on average across mod￾els, though only s… view at source ↗
Figure 4
Figure 4. Figure 4: Effect of image normalization for κ passes on est. proba￾bility of choosing the original vs. final variants [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of image normalization on human choices, com￾pared with original vs. final trials. After 3 passes, the probability of choosing the final optimized image decreases. lack the statistical power to detect small effect-size differ￾ences in head-to-head comparisons. Generalization to other agentic visual decision contexts with long-horizon temporal sequences would require substantial further development. … view at source ↗
Figure 7
Figure 7. Figure 7: shows examples of the visual prompt optimization with CVPO at different steps, and for all our datasets. T=0 T=1 T=3 T=5 T=12 Houses T=0 T=1 T=3 T=6 T=11 People T=0 T=1 T=3 T=9 T=14 Products T=0 T=1 T=2 T=5 T=8 Hotels [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: shows an analysis of how the mitigation procedure affects the perceptual similarity of the images to their own original states and to each other in a comparison pair (where the mitigating model is explicitly instructed to align visual features of the two). We provide several metrics for robustness: cosine similarity of CLIP (Radford et al., 2021) embeddings with and without backgrounds (matted using a U2 -… view at source ↗
Figure 9
Figure 9. Figure 9: Head-to-head experiment results disaggregated by model. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Head-to-head experiment results disaggregated by task. ************ ************ ********** ************ ************ ********** ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ******** ************ ************ ********** ************ ************ ******** * Claude Haiku 4… view at source ↗
Figure 11
Figure 11. Figure 11: Hotel experiment results disaggregated by model and strategy. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Hotel experiment results disaggregated by class and strategy. ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ************ ********… view at source ↗
Figure 13
Figure 13. Figure 13: Houses experiment results disaggregated by model and strategy. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: People experiment results disaggregated by model and strategy. ****** ********** ******** ******** * ************ ********** ************ ************ ******** *********** ************ ******** ******** * ************ ******** ******** ********** ******** ** * ******** ******** *********** ******** * ********** ************ ********** Claude Haiku 4.5 Claude Sonnet 4.5 Gemini 3 Flash Gemini 3 Pro GPT−4o G… view at source ↗
Figure 15
Figure 15. Figure 15: Products experiment results disaggregated by model and strategy. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Products experiment results disaggregated by class and strategy [PITH_FULL_IMAGE:figures/full_fig_p028_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Human mitigation experiment results disaggregated by task (dataset). 28 [PITH_FULL_IMAGE:figures/full_fig_p028_17.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

45 extracted references · 17 linked inside Pith

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    A., Tan, S., Soylu, D., Ziems, N., Khare, R., Opsahl-Ong, K., Singhvi, A., Shandilya, H., Ryan, M

    Agrawal, L. A., Tan, S., Soylu, D., Ziems, N., Khare, R., Opsahl-Ong, K., Singhvi, A., Shandilya, H., Ryan, M. J., Jiang, M., et al. Gepa: Reflective prompt evolution can outperform reinforcement learning. arXiv preprint arXiv:2507.19457, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    and Moustafa, M

    Ahmed, E. and Moustafa, M. House price estimation from visual and textual features. arXiv preprint arXiv:1609.08399, 2016

    Pith/arXiv arXiv 2016

  4. [4]

    Language models can explain neurons in language models, 2023

    Bills, S., Cammarata, N., Mossing, D., Tillman, H., Gao, L., Goh, G., Sutskever, I., Leike, J., Wu, J., and Saunders, W. Language models can explain neurons in language models, 2023. URL https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html

    2023

  5. [5]

    and Toubia, O

    Brucks, M. and Toubia, O. Prompt architecture can induce methodological artifacts in large language models. Available at SSRN 4484416, 2023

    2023

  6. [6]

    D., Webb, T

    Budny, N., Ghods, K., Campbell, D., Marjieh, R., Joshi, A., Kumar, S., Cohen, J. D., Webb, T. W., and Griffiths, T. L. Visual serial processing deficits explain divergences in human and vlm reasoning. arXiv preprint arXiv:2509.25142, 2025

    arXiv 2025

  7. [7]

    N., Ghods, K., Joshi, A., Ku, A., Frankland, S., Griffiths, T., Cohen, J

    Campbell, D., Rane, S., Giallanza, T., De Sabbata, C. N., Ghods, K., Joshi, A., Ku, A., Frankland, S., Griffiths, T., Cohen, J. D., et al. Understanding the limits of vision language models through the lens of the binding problem. Advances in Neural Information Processing Systems, 37: 0 113436--113460, 2024

    2024

  8. [8]

    Superficial alignment, subtle divergence, and nudge sensitivity in llm decision-making

    Cherep, M., Singh, N., and Maes, P. Superficial alignment, subtle divergence, and nudge sensitivity in llm decision-making. In NeurIPS 2024 Workshop on Behavioral Machine Learning, 2024

    2024

  9. [9]

    A framework for studying ai agent behavior: Evidence from consumer choice experiments

    Cherep, M., Ma, C., Xu, A., Shaked, M., Maes, P., and Singh, N. A framework for studying ai agent behavior: Evidence from consumer choice experiments. arXiv preprint arXiv:2509.25609, 2025 a

    arXiv 2025

  10. [10]

    Llm agents are hypersensitive to nudges

    Cherep, M., Maes, P., and Singh, N. Llm agents are hypersensitive to nudges. arXiv preprint arXiv:2505.11584, 2025 b

    Pith/arXiv arXiv 2025

  11. [11]

    Behavioral systems require behavioral tests

    Cherep, M., Singh, N., and Maes, P. Behavioral systems require behavioral tests. In NeurIPS 2025 Workshop on Bridging Language, Agent, and World Models for Reasoning and Planning, 2025 c

    2025

  12. [12]

    Choi, Y., Kim, D., Baek, J., and Hwang, S. J. Multimodal prompt optimization: Why not leverage multiple modalities for mllms. arXiv preprint arXiv:2510.09201, 2025

    arXiv 2025

  13. [13]

    F., Dideriksen, T., Arora, H., Guillaumin, M., and Malik, J

    Collins, J., Goel, S., Deng, K., Luthra, A., Xu, L., Gundogdu, E., Zhang, X., Yago Vicente, T. F., Dideriksen, T., Arora, H., Guillaumin, M., and Malik, J. Abo: Dataset and benchmarks for real-world 3d object understanding. CVPR, 2022

    2022

  14. [14]

    Cuesta-Vali \ n o, P., Kazakov, S., Guti \'e rrez-Rodr \' guez, P., and Rua, O. L. The effects of the aesthetics and composition of hotels’ digital photo images on online booking decisions. Humanities and Social Sciences Communications, 10 0 (1): 0 1--11, 2023

    2023

  15. [15]

    C., Wu, W., and Liu, Z

    Fu, J., Li, S., Jiang, Y., Lin, K.-Y., Qian, C., Loy, C. C., Wu, W., and Liu, Z. Stylegan-human: A data-centric odyssey of human generation. In European Conference on Computer Vision, pp.\ 1--19. Springer, 2022

    2022

  16. [16]

    J., Keuper, M., and Keuper, J

    Gavrikov, P., Lukasik, J., Jung, S., Geirhos, R., Mirza, M. J., Keuper, M., and Keuper, J. Can we talk models into seeing the world differently? arXiv preprint arXiv:2403.09193, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    J., Shlens, J., and Szegedy, C

    Goodfellow, I. J., Shlens, J., and Szegedy, C. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014

    Pith/arXiv arXiv 2014

  18. [18]

    The agentic investor: Ai for real estate investment management

    Graham, L. The agentic investor: Ai for real estate investment management. Available at SSRN 5720342, 2025

    2025

  19. [19]

    Grigsby, J., Zhu, Y., Ryoo, M., and Niebles, J. C. Vlm q-learning: Aligning vision-language models for interactive decision-making. arXiv preprint arXiv:2505.03181, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    R., and Price, S

    Gupta, I., Fronsdal, K., Sheshadri, A., Michala, J., Tay, J., Wang, R., Bowman, S. R., and Price, S. Bloom: an open source tool for automated behavioral evaluations, 2025. URL https://github.com/safety-research/bloom

    2025

  21. [21]

    Handa, K., Gal, Y., Pavlick, E., Goodman, N., Andreas, J., Tamkin, A., and Li, B. Z. Bayesian preference elicitation with language models. arXiv preprint arXiv:2403.05534, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Agentic interactions

    Imas, A., Lee, K., and Misra, S. Agentic interactions. Available at SSRN 5875162, 2025

    2025

  23. [23]

    How to distinguish ai-generated images from authentic photographs

    Kamali, N., Nakamura, K., Chatzimparmpas, A., Hullman, J., and Groh, M. How to distinguish ai-generated images from authentic photographs. arXiv preprint arXiv:2406.08651, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    H., Zheng, W., Zhou, Y., Mai, Y., Roberts, J

    Lee, T., Tu, H., Wong, C. H., Zheng, W., Zhou, Y., Mai, Y., Roberts, J. S., Yasunaga, M., Yao, H., Xie, C., et al. Vhelm: A holistic evaluation of vision language models. Advances in Neural Information Processing Systems, 37: 0 140632--140666, 2024

    2024

  25. [25]

    Feedback descent: Open-ended text optimization via pairwise comparison

    Lee, Y., Boen, J., and Finn, C. Feedback descent: Open-ended text optimization via pairwise comparison. arXiv preprint arXiv:2511.07919, 2025

    arXiv 2025

  26. [26]

    L., Touchet, J

    Li, C., Camu \ n as, J. L., Touchet, J. T., Andreas, J., Lapedriza, A., Torralba, A., and Shaham, T. R. Automated detection of visual attribute reliance with a self-reflective agent. arXiv preprint arXiv:2510.21704, 2025

    arXiv 2025

  27. [27]

    P.-W., Qiu, J., Wang, Z., Yu, H., Chen, Y., Zhang, G., and Lo, B

    Lo, F. P.-W., Qiu, J., Wang, Z., Yu, H., Chen, Y., Zhang, G., and Lo, B. Ai hiring with llms: A context-aware and explainable multi-agent framework for resume screening. In Proceedings of the Computer Vision and Pattern Recognition Conference, pp.\ 4184--4193, 2025

    2025

  28. [28]

    Improving text-to-image consistency via automatic prompt optimization

    Ma \ n as, O., Astolfi, P., Hall, M., Ross, C., Urbanek, J., Williams, A., Agrawal, A., Romero-Soriano, A., and Drozdzal, M. Improving text-to-image consistency via automatic prompt optimization. arXiv preprint arXiv:2403.17804, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Automatically interpreting millions of features in large language models

    Paulo, G., Mallen, A., Juang, C., and Belrose, N. Automatically interpreting millions of features in large language models. arXiv preprint arXiv:2410.13928, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Discovering language model behaviors with model-written evaluations

    Perez, E., Ringer, S., Lukosiute, K., Nguyen, K., Chen, E., Heiner, S., Pettit, C., Olsson, C., Kundu, S., Kadavath, S., et al. Discovering language model behaviors with model-written evaluations. In Findings of the Association for Computational Linguistics: ACL 2023, pp.\ 13387--13434, 2023

    2023

  31. [31]

    gradient descent

    Pryzant, R., Iter, D., Li, J., Lee, Y. T., Zhu, C., and Zeng, M. Automatic prompt optimization with" gradient descent" and beam search. arXiv preprint arXiv:2305.03495, 2023

    Pith/arXiv arXiv 2023

  32. [32]

    R., and Jagersand, M

    Qin, X., Zhang, Z., Huang, C., Dehghan, M., Zaiane, O. R., and Jagersand, M. U2-net: Going deeper with nested u-structure for salient object detection. Pattern recognition, 106: 0 107404, 2020

    2020

  33. [33]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp.\ 8748--8763. PmLR, 2021

    2021

  34. [34]

    Quantifying language models' sensitivity to spurious features in prompt design or: How i learned to start worrying about prompt formatting

    Sclar, M., Choi, Y., Tsvetkov, Y., and Suhr, A. Quantifying language models' sensitivity to spurious features in prompt design or: How i learned to start worrying about prompt formatting. arXiv preprint arXiv:2310.11324, 2023

    Pith/arXiv arXiv 2023

  35. [35]

    Intriguing properties of neural networks

    Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., and Fergus, R. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013

    Pith/arXiv arXiv 2013

  36. [36]

    Wan, X., Zhou, H., Sun, R., Nakhost, H., Jiang, K., Sinha, R., and Ar k, S. \"O . Maestro: Self-improving text-to-image generation via agent orchestration. arXiv preprint arXiv:2509.10704, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    H., Jiang, Z., Zheng, Z., Wu, Z., Chen, M., and Xiao, C

    Wang, J., Liu, Z., Park, K. H., Jiang, Z., Zheng, Z., Wu, Z., Chen, M., and Xiao, C. Adversarial demonstration attacks on large language models. arXiv preprint arXiv:2305.14950, 2023

    Pith/arXiv arXiv 2023

  38. [38]

    C., Sheikh, H

    Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13 0 (4): 0 600--612, 2004

    2004

  39. [39]

    Qwen-image technical report

    Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., Yin, S.-m., Bai, S., Xu, X., Chen, Y., et al. Qwen-image technical report. arXiv preprint arXiv:2508.02324, 2025

    Pith/arXiv arXiv 2025

  40. [40]

    Webshop: Towards scalable real-world web interaction with grounded language agents

    Yao, S., Chen, H., Yang, J., and Narasimhan, K. Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems, 35: 0 20744--20757, 2022

    2022

  41. [41]

    How do visual attributes influence web agents? a comprehensive evaluation of user interface design factors

    Yu, K., Yu, N., Wang, H., Yang, R., and Zhang, H. How do visual attributes influence web agents? a comprehensive evaluation of user interface design factors. arXiv preprint arXiv:2601.21961, 2026

    arXiv 2026

  42. [42]

    Optimizing generative ai by backpropagating language model feedback

    Yuksekgonul, M., Bianchi, F., Boen, J., Liu, S., Lu, P., Huang, Z., Guestrin, C., and Zou, J. Optimizing generative ai by backpropagating language model feedback. Nature, 639 0 (8055): 0 609--616, 2025

    2025

  43. [43]

    Fine-tuning large vision-language models as decision-making agents via reinforcement learning

    Zhai, S., Bai, H., Lin, Z., Pan, J., Tong, P., Zhou, Y., Suhr, A., Xie, S., LeCun, Y., Ma, Y., et al. Fine-tuning large vision-language models as decision-making agents via reinforcement learning. Advances in neural information processing systems, 37: 0 110935--110971, 2024

    2024

  44. [44]

    A., Shechtman, E., and Wang, O

    Zhang, R., Isola, P., Efros, A. A., Shechtman, E., and Wang, O. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 586--595, 2018

    2018

  45. [45]

    F., Zhu, H., Zhou, X., Lo, R., Sridhar, A., Cheng, X., Ou, T., Bisk, Y., Fried, D., et al

    Zhou, S., Xu, F. F., Zhu, H., Zhou, X., Lo, R., Sridhar, A., Cheng, X., Ou, T., Bisk, Y., Fried, D., et al. Webarena: A realistic web environment for building autonomous agents. arXiv preprint arXiv:2307.13854, 2023

    Pith/arXiv arXiv 2023

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.