See2Think: Do Multimodal Models Really Use Intermediate Visual States?

• URL: http://arxiv.org/abs/2607.26769v1
• Date: Wed, 29 Jul 2026 11:10:34 GMT
• Title: See2Think: Do Multimodal Models Really Use Intermediate Visual States?
• Abstract summary: We introduce See2Think, a unified evaluation framework comprising See2ThinkBench and Visual Action-of-Thought (VAoT)<n>VAoT records textual thoughts, visual actions, rendered states, and subsequent reasoning under four controlled inference settings.<n>We find that visual reasoning is strongly model- and environment-dependent, with no single setting consistently dominating across tasks.
• Score: 12.082470543980422
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Multimodal large language models increasingly use sketches, annotations, tools, and intermediate images during reasoning, but it remains unclear whether they truly rely on these visual states. Existing benchmarks are limited both by task collections with narrow coverage or partially text-solvable samples and by evaluations that emphasize final answers without diagnosing how intermediate visual states are generated, rendered, and used. We introduce See2Think, a unified evaluation framework comprising See2ThinkBench and Visual Action-of-Thought (VAoT). See2ThinkBench contains 1,200 open-ended, visually dependent problems across 12 task categories spanning 2D structured, 3D scene, and real-world reasoning. VAoT records textual thoughts, visual actions, rendered states, and subsequent reasoning under four controlled inference settings. Evaluating representative proprietary and open-source multimodal models, we find that visual reasoning is strongly model- and environment-dependent, with no single setting consistently dominating across tasks. Process analysis further shows that models usually select relevant visual operations, while faithful rendering remains the clearest bottleneck and high feedback uptake does not necessarily translate into accuracy gains. Under task-relevant corrupted feedback, models exhibit behavioral dependence on visual states, with accuracy dropping by over 10 percentage points in controlled interventions.

Related papers

• Not Another Text Benchmark: Putting the "Visual" Back in Visual Question Answering for Large Video Models [97.0559653915626]
  We introduce three new vision-centric evaluation benchmarks in temporal frame retrieval, video future prediction, and causal memory distortion.<n>We show that current frontier models exhibit significant weakness when attempting to reason through visual queries, rather than text.
  arXiv  Detail & Related papers  (2026-09-15T12:41:08Z)
• MNIST-PRO: MNIST is Back as a Partially Observable World for AI Agents [74.83535434786145]
  MNIST-PRO is a benchmark that isolates agentic perception by converting MNIST digit recognition into a sequential, glimpse-based search task with lookback constraints.<n>We evaluate ten multimodal models across four memory representations, including raw visual history, textual states, structured metric grid maps, and a consolidated visual canvas.
  arXiv  Detail & Related papers  (2026-08-31T16:05:39Z)
• Seeing or Knowing? Visual Context Sensitivity in Multimodal Large Language Models [14.5102059776721]
  Multimodal Large Language Models (MLLMs) achieve strong performance by integrating visual inputs with the rich priors of pretrained language models.<n>However, they often fail on vision-centric tasks, especially when visual evidence conflicts with pretrained knowledge.<n>We explore these failures using two diagnostic paradigms: (1) probing whether visual information is available, via image reconstruction, and (2) measuring multimodal context sensitivity.
  arXiv  Detail & Related papers  (2026-07-28T22:52:34Z)
• Seeing Isn't Knowing: Do VLMs Know When Not to Answer Spatial Questions (and Why)? [72.2500547961037]
  spatial reasoning is a fundamental capability for vision-language models (VLMs) deployed in real-world environments.<n>We introduce two types of observation challenges: (1) occlusion, which hides target information, and (2) perspective ambiguity, which produces misleading visual cues.<n>For each configuration, we design spatial questions that are answerable under clean observations but require abstention under the introduced challenges.
  arXiv  Detail & Related papers  (2026-05-28T20:44:47Z)
• Beyond Accuracy: Evaluating Grounded Visual Evidence in Thinking with Images [34.324634481264034]
  We propose ViEBench, a process-verifiable benchmark designed to evaluate faithful visual reasoning.<n>Comprising 200 multi-scenario high-resolution images with expert-annotated visual evidence, ViEBench categorizes tasks by difficulty into perception and reasoning dimensions.<n>Our experiments yield several interesting observations: (1) VLMs can sometimes produce correct final answers despite grounding on irrelevant regions, and (2) they may successfully locate the correct evidence but still fail to utilize it to reach accurate conclusions.
  arXiv  Detail & Related papers  (2026-01-14T07:25:15Z)
• When Visualizing is the First Step to Reasoning: MIRA, a Benchmark for Visual Chain-of-Thought [118.71264263478083]
  We propose MIRA, a new benchmark designed to evaluate models in scenarios where generating intermediate visual images is essential for successful reasoning.<n>We include 546 multimodal problems, annotated with intermediate visual images and final answers.
  arXiv  Detail & Related papers  (2025-11-04T18:00:51Z)
• BLINK-Twice: You see, but do you observe? A Reasoning Benchmark on Visual Perception [67.89135437537179]
  We introduce BLINK-Twice, a vision-centric reasoning benchmark grounded in challenging perceptual tasks.<n>Instead of relying on external knowledge, our tasks require models to reason from visual content alone.<n>Compared to prior perception benchmarks, it moves beyond shallow perception and requires fine-grained observation and analytical reasoning.
  arXiv  Detail & Related papers  (2025-10-10T13:14:13Z)
• Probing Vision-Language Understanding through the Visual Entailment Task: promises and pitfalls [0.10923877073891446]
  This study investigates the extent to which the Visual Entailment task serves as a reliable probe of vision-language understanding in multimodal language models.<n>We conduct experiments across zero-shot, few-shot, and fine-tuning settings, exploring how factors such as prompt design might affect VE performance.<n>Fine-tuning yields strong results, achieving an accuracy of 83.3% on the e-SNLI-VE dataset and outperforming the state-of-the-art OFA-X model.
  arXiv  Detail & Related papers  (2025-07-23T12:46:51Z)
• SpatialViz-Bench: An MLLM Benchmark for Spatial Visualization [44.427830927596204]
  SpatialViz-Bench is a comprehensive benchmark for spatial visualization with 12 tasks across 4 sub-abilities, comprising 1,180 automatically generated problems.<n>Our evaluation of 33 state-of-the-art MLLMs reveals wide performance variations and uncovers counter-intuitive findings.
  arXiv  Detail & Related papers  (2025-07-10T10:27:20Z)
• Can MLLMs Guide Me Home? A Benchmark Study on Fine-Grained Visual Reasoning from Transit Maps [56.76175383189738]
  We introduce ReasonMap, a benchmark designed to assess the fine-grained visual understanding and spatial reasoning abilities of MLLMs.<n>ReasonMap encompasses high-resolution transit maps from 30 cities across 13 countries and includes 1,008 question-answer pairs spanning two question types and three templates.<n> Comprehensive evaluations of 15 popular MLLMs, including both base and reasoning variants, reveal a counterintuitive pattern.
  arXiv  Detail & Related papers  (2025-05-24T12:33:52Z)
• VHELM: A Holistic Evaluation of Vision Language Models [75.88987277686914]
  We present the Holistic Evaluation of Vision Language Models (VHELM) VHELM aggregates various datasets to cover one or more of the 9 aspects: visual perception, knowledge, reasoning, bias, fairness, multilinguality, robustness, toxicity, and safety. Our framework is designed to be lightweight and automatic so that evaluation runs are cheap and fast.
  arXiv  Detail & Related papers  (2024-10-09T17:46:34Z)
• LOIS: Looking Out of Instance Semantics for Visual Question Answering [17.076621453814926]
  We propose a model framework without bounding boxes to understand the causal nexus of object semantics in images. We implement a mutual relation attention module to model sophisticated and deeper visual semantic relations between instance objects and background information. Our proposed attention model can further analyze salient image regions by focusing on important word-related questions.
  arXiv  Detail & Related papers  (2023-07-26T12:13:00Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.