Research map / Vision-language and multimodal / Multimodal LLMs

Linking Perception, Confidence and Accuracy in MLLMs

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual 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 CDRL, an RL method using original-noised image pairs and a confidence-based reward to calibrate MLLM confidence, plus CA-TTS, a confidence-aware test-time scaling framework coordinating self-consistency, self-reflection, and visual self-check via a Gemini-2.5-Pro expert.

Problem

MLLMs show severe confidence miscalibration: their confidence stays high even when visual evidence is degraded, so they do not know when they do not know.

Approach

Use CLIP-attention-based noised image pairs in GRPO with a reward combining a confidence-change perception term and a calibration term tied to answer correctness, then use the calibrated confidence to drive adaptive test-time scaling with an expert model acting as planner, voter, and critic.

Key results

• CDRL+CA-TTS reaches 79.5 on MathVista, 42.4 on MathVision, 71.3 on MMStar, and 66.3 on MMMU with Qwen2.5-VL-7B, versus 64.7, 23.0, 60.2, 48.8 for Pass@1.
• On MathVision, CA-TTS scaling slope is 3.65 versus 1.64 for majority voting and 1.19 for DeepConf.
• CDRL raises confidence drop under noised images from -0.32 to -1.39 on Math-Vision for Qwen2.5-VL-7B and lowers ECE on the origin condition from 64.57 to 62.24.

Surprising findings

• The base Qwen2.5-VL-7B shows near-zero or positive confidence drop under occlusion, viewpoint, and mosaic perturbations, indicating confidence is insensitive to visual degradation.
• Even Qwen2.5-VL-7B itself as expert gives gains over majority voting, suggesting the framework is not heavily dependent on expert strength.

Limitations

• Main results rely on Gemini-2.5-Pro as an external expert, making gains partly dependent on a proprietary model.
• Training set is small (1936 samples) and the main-table comparisons may mix settings with different evaluation splits (testmini vs test).

Experimental setup

Datasets: MathVerse DynaMath LogicVista RealWorldQA MMBench MMMU-Pro

Benchmarks: MathVista MathVision MMStar MMMU

Metrics: accuracy ECE AUC confidence drop scaling slope

Baselines: Majority Voting DeepConf DreamPRM R1-Onevision VL-Rethinker We-Think ToT

Builds on: GRPO DeepConf Visual Contrastive Decoding Self-Consistency Tree of Thoughts

Closest papers by meaning

• Be Confident: Uncovering Overfitting in MLLM Multi-Task Tuning (ICML 2025 · similarity 0.90)
• Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.89)
• Unleashing Perception-Time Scaling to Multimodal Reasoning Models (ICLR 2026 · similarity 0.88)
• PDCR: Perception-Decomposed Confidence Reward for Vision-Language Reasoning (CVPR 2026 · similarity 0.88)
• On Robustness and Chain-of-Thought Consistency of RL-Finetuned VLMs (ICML 2026 · similarity 0.87)
• Bad Seeing or Bad Thinking? Rewarding Perception for Multimodal Reasoning (ICML 2026 · similarity 0.87)