Research map / Vision-language and multimodal / Multimodal LLMs

VISCO: Benchmarking Fine-Grained Critique and Correction Towards Self-Improvement in Visual Reasoning

CVPR 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces VISCO, a benchmark for fine-grained step-wise critique and correction by LVLMs, and proposes VISCO-Critic (revisit-the-image question verification) as a baseline that improves critique by up to 13.5%.

Problem

LVLMs' ability to critique and correct their own visual reasoning, needed for self-improvement, has not been systematically analyzed; existing critique benchmarks use scalar, text-only, or non-correction-oriented evaluation.

Approach

Dense, step-wise binary critique with natural-language explanations for LVLM chain-of-thought responses, evaluated with a hierarchical F1-based metric (VISCO score) and correction gain, plus a training-free critique method that decomposes CoT steps into atomic visual questions and re-verifies them against the image.

Key results

• GPT-4o achieves the best VISCO critique score of 52.4 among evaluated models, versus 86.5 for human experts.
• Human-annotated critiques enable correction of up to 76% of errors; self-generated critiques often yield negative correction gains for open models.
• VISCO-Critic improves InternVL2-26B critique score from 25.2 to 38.7 (+13.5), with perception gains of +22.2.

Measured effects reported in the paper

Surprising findings

• Self-generated critiques are less helpful and sometimes detrimental to correction, with many open models showing negative correction gains.
• LVLMs predict 'Incorrect' less often than ground truth, showing a reluctance to say no; LLaVA-1.6-34B flags only 5.7% of inputs as incorrect.

Limitations

• CoT constrained to at most five reasoning steps, limiting generalization to complex or multi-turn reasoning.
• Dataset is relatively small (1645 questions) and critique annotation is costly; human baseline measured on only 265 points.
• LLM-based explanation evaluation with GPT-4o has moderate agreement (0.61 kappa with manual judgments).

Experimental setup

Datasets: VQA datasets across 18 sources (reasoning and perception)

Benchmarks: VISCO

Metrics: VISCO score (geometric mean of answer, step, explanation F1) correction gain (PCR - NCR) F1 at answer, step, explanation levels

Baselines: Prometheus-Vision LLaVA-Critic Gemini-1.5-Pro Claude-3.5-Sonnet GPT-4o

Builds on: chain-of-thought prompting LLaVA-Critic Prometheus-Vision

Closest papers by meaning

• Through the Lens of Contrast: Self-Improving Visual Reasoning in VLMs (ICLR 2026 · similarity 0.91)
• What, Whether and How? Unveiling Process Reward Models for Thinking with Images Reasoning (AAAI 2026 · similarity 0.90)
• Sherlock: Self-Correcting Reasoning in Vision-Language Models (NeurIPS 2025 · similarity 0.90)
• VisuLogic: A Benchmark for Evaluating Visual Reasoning in Multi-modal Large Language Models (ICLR 2026 · similarity 0.90)
• VisRes Bench: On Evaluating the Visual Reasoning Capabilities of VLMs (CVPR 2026 · similarity 0.90)
• From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.90)