Research map / Vision-language and multimodal / Multimodal LLMs

See It, Say It, Sorted: An Iterative Training-Free Framework for Visually-Grounded Multimodal Reasoning in LVLMs

CVPR 2026 · 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.

Proposes ECRD, a training-free decoding framework that reweights LVLM next-token probabilities using a growing textual visual-evidence pool and invokes a GRIT-3B visual decider only when uncertainty is high.

Problem

Chain-of-thought reasoning in LVLMs suffers from visual hallucination propagation, where an early step inconsistent with the image leads to wrong final answers; existing fixes rely on costly RL training of think-with-images policies.

Approach

At test time, mix the base next-token distribution with an evidence-induced distribution computed from a textual evidence pool (mean-over-prefix likelihoods), with mixing weight equal to the base top probability; when the mixed top-2 margin is small and the knee set has more than one token, a visual decider (GRIT-3B) produces a new textual micro-observation that is added to the pool.

Key results

• On TreeBench, ECRD raises Qwen2.5-VL-7B overall accuracy from 37.0 to 47.9 (+10.9).
• On RH-Bench, RH-AUC rises from 0.51 to 0.58 for Qwen2.5-VL-7B, with Reasoning 39.6 to 46.4 and Perception 50.2 to 57.1.
• ECRD outperforms training-free baselines (Woodpecker 36.3, ViperGPT 38.3, beam search 38.8) and RL-based DeepEyes-7B (37.5) and Pixel-Reasoner-7B (39.0) on TreeBench with Qwen2.5-VL-7B.

Surprising findings

• The GRIT-3B visual decider alone scores lowest on TreeBench (30.1), yet ECRD using it at uncertain steps reaches 47.9, which the authors attribute to the decoding design rather than decider strength.

Limitations

• Adds inference latency from decider calls and evidence scoring, growing with the uncertainty threshold.
• Relies on the quality of the visual decider's textual micro-observations; decider errors can propagate.
• Hyperparameter delta and knee-based trigger rely on empirical choice (delta=0.08).

Experimental setup

Datasets: TreeBench RH-Bench V*Bench MathVista ChartQA OCRBench HallusionBench

Benchmarks: TreeBench RH-Bench V*Bench MathVista ChartQA OCRBench HallusionBench

Metrics: accuracy RH-AUC reasoning score perception score

Baselines: DeepEyes Pixel-Reasoner TreeVGR-7B Woodpecker ViperGPT ControlMLLM GRIT-3B

Builds on: VDGD GRIT TreeVGR

Closest papers by meaning

• Grounded Chain-of-Thought for Multimodal Large Language Models (CVPR 2026 · similarity 0.92)
• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.92)
• Machine Mental Imagery: Empower Multimodal Reasoning with Latent Visual Tokens (CVPR 2026 · similarity 0.92)
• Understanding the Role of Hallucination in Reinforcement Post-Training of Multimodal Reasoning Models (CVPR 2026 · similarity 0.92)
• Perceptual-Evidence Anchored Reinforced Learning for Multimodal Reasoning (CVPR 2026 · similarity 0.92)
• VGR: Visual Grounded Reasoning (ICLR 2026 · similarity 0.92)