cs.LGNov 14, 2026

Read, Then Shade? How Visual Evidence Changes Incentive-Driven Misreporting in Vision-Language Advisors

Authors: Hamidreza Hasani Balyani

Organizations: Independent Researcher

Abstract

Vision-language models (VLMs) increasingly read a scene and report it to someone who acts on the report, sometimes under objectives that differ from that person's. We ask whether a stated incentive moves what a VLM reports, and whether this changes when the value must be read from pixels rather than given as text. In seeded synthetic scenes, a model reads the position of a target on a labelled axis and reports one integer to a fixed receiver; it is paid, in simulation, for accuracy or for a report a few units to one side, and every scene is also presented with the true value as text. Across a pilot and three follow-up studies with hash-frozen protocols, four small open VLMs from three families, and 15,170 held-out trials, a stated payoff moved reports in both modalities, and exact compliance at three units was significantly lower with the image than with text for three of four models; Qwen2.5-VL 7B showed no reliable difference. Adding the true value as text to the image prompt removed the gap in every model. For InternVL3 the gap vanished on scenes it read correctly when it wrote its reading first; for Gemma 3 it persisted. For Qwen2.5-VL 3B, which read every scene correctly, the gap held in four of five payoff paraphrases, nearly vanished when the model wrote its reading before its report or when the prompt stated that the client wants the truth, vanished with an unpaid goal, and persisted when the model reported before stating its reading; how often it shaded at all, even with text, depended strongly on the wording. When told that the client wants the true value, three of four models still shaded every text report. Visual input did not make these advisors reliably more truthful; it made compliance with a stated incentive depend on whether the read value is made explicit and on output order, a dependence that text-only evaluations and mean-shift metrics do not reveal.

Figures & tables

Explore similar work

Apr 16, 2026

Reasoning Dynamics and the Limits of Monitoring Modality Reliance in Vision-Language Models

Recent advances in vision language models (VLMs) offer reasoning capabilities, yet how these unfold and integrate visual and textual information remains unclear. We analyze reasoning dynamics in 18 VLMs covering instruction-tuned and reasoning-trained models from two different model families. We track confidence over Chain-of-Thought (CoT), measure the corrective effect of reasoning, and evaluate the contribution of intermediate reasoning steps. We find that models are prone to answer inertia, in which early commitments to a prediction are reinforced, rather than revised during reasoning steps. While reasoning-trained models show stronger corrective behavior, their gains depend on modality conditions, from text-dominant to vision-only settings. Using controlled interventions with misleading textual cues, we show that models are consistently influenced by these cues even when visual evidence is sufficient, and assess whether this influence is recoverable from CoT. Although this influence can appear in the CoT, its detectability varies across models and depends on what is being monitored. Reasoning-trained models are more likely to explicitly refer to the cues, but their longer and fluent CoTs can still appear visually grounded while actually following textual cues, obscuring modality reliance. In contrast, instruction-tuned models refer to the cues less explicitly, but their shorter traces reveal inconsistencies with the visual input. Taken together, these findings indicate that CoT provides only a partial view of how different modalities drive VLM decisions, with important implications for the transparency and safety of multimodal systems.
May 26, 2026

Real Images, Worse Judgments: Evaluating Vision-Language Models on Concreteness and Imagery

Visual inputs are often assumed to improve language understanding in multimodal models. We examine this assumption by asking whether vision-language models (VLMs) can distinguish useful visual evidence from incidental image context in lexical judgments. We use human concreteness and imagery ratings because they span words with varying expected visual relevance, from abstract and low-imagery words to concrete and high-imagery words. We find that real-image contexts do not yield consistent gains and often hurt alignment with human ratings, most sharply when visual evidence is least relevant. Through probing and canonical correlation analysis, complemented by an attribution case study, we find that real-image contexts are associated with representational shifts and greater sensitivity to spurious visual cues, coinciding with weaker recoverability of the targeted lexical properties. We further show that instructing models to focus solely on textual content at inference time can reduce this degradation, with the clearest gains on these vulnerable subsets. Our findings suggest that current instruction-tuned VLMs need better calibration of when visual context should inform lexical judgments.
Nov 17, 2026

Lost in Read-out: Locating the Image Quality Gap in Vision-Language Models with a Closed-Form Edit

Vision-language models (VLMs) are widely used as zero-shot image quality assessors, yet their good-vs-poor answers agree only moderately with human opinion. We ask whether these models fail to see quality or fail to say it. Layer-wise probes on five VLMs point to the second: quality is linearly decodable from the vision encoder and still from the last language layer, far better than the answer reveals, and about a hundred labels recover most of it. The loss lies mainly in the final read-out, which partly reads something else: relative to human opinion, it over-rewards colorfulness and under-rewards sharpness on all five models. We test this by intervention: if the read-out is at fault, a closed-form edit of a single last-layer neuron, fitted to many prompts at once, should close the gap and no more. Both hold. Under prompts never used for fitting, the edit recovers 86--100% of what a ridge head on the last-layer state, fitted under the same prompts, finds beyond the answer on KonIQ. Where that gap is small, it gains little, too little for our transfer criterion on the two largest models; where the head itself is below the answer, it lowers the score; and it works only when written from the middle of the language model on. Unlike unconstrained fine-tuning, the edit leaves perception question answering intact. It does not carry over to yes-or-no or rating questions, and it reverses the good--poor contrast after sentiment-laden sentences. Code will be made available.