Research map / Vision-language and multimodal / Multimodal LLMs

Thinking in Uncertainty: Mitigating Hallucinations in MLRMs with Latent Entropy-Aware Decoding

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 LEAD, a training-free plug-and-play decoding method that switches between discrete token embeddings and probability-weighted latent embeddings based on token entropy, plus a visual anchor injection, to reduce hallucinations in multimodal large reasoning models.

Problem

Multimodal large reasoning models (MLRMs) hallucinate, and hallucinations concentrate after high-entropy transition words (e.g., because, however, wait) during explicit reasoning.

Approach

Under high token entropy, feed the model a probability-weighted mixture of token embeddings instead of a sampled one-hot token, switch back to discrete embeddings as entropy drops (with a persistence window and switch budget), and inject a visual special-token embedding at the onset of each high-entropy phase.

Key results

• LEAD on R1-Onevision-7B raises MMHalu from 3.52 to 3.80 and Bingo from 3.65 to 3.84 and VStar from 66.5 to 71.2.
• On Vision-R1-7B, LEAD improves VStar from 78.5 to 81.7 and MMHalu from 3.64 to 3.89.
• LEAD improves average accuracy by +2.0% on math benchmarks and +3.2% on scientific benchmarks for R1-Onevision-7B, and shortens reasoning length on MathVision.

Surprising findings

• Masking high-entropy tokens hurts reasoning more than masking low-entropy tokens, and early high-entropy tokens have the strongest influence on the trajectory.
• Hallucinated high-entropy tokens receive lower visual attention than non-hallucinated ones.

Limitations

• Gains are modest on some benchmarks (e.g., +0.7 to +2.0 on several math and POPE scores).
• Relies on hyperparameters (entropy threshold dynamics, persistence window, switch budget C_max=5, injection strength lambda=0.4) tuned on the evaluated models.

Experimental setup

Datasets: MMEval-Pro MMVP RealWorldQA VMCBench VStar Bingo MMHalu POPE MathVision MathVista MathVerse VisuLogic

Benchmarks: VStar RealWorldQA MMVP MMEval-Pro VMCBench MMHalu Bingo POPE MathVision MathVista MathVerse VisuLogic

Metrics: accuracy MMHalu score (0-6) Bingo score (1-5) perplexity pass@k

Baselines: VCD MemVR SID

Builds on: contrastive decoding MemVR superposed reasoning

Closest papers by meaning

• Seeing Far and Clearly: Mitigating Hallucinations in MLLMs with Attention Causal Decoding (CVPR 2025 · similarity 0.91)
• Joint Evaluation of Answer and Reasoning Consistency for Hallucination Detection in Large Reasoning Models (AAAI 2026 · similarity 0.89)
• More Thinking, Less Seeing? Assessing Amplified Hallucination in Multimodal Reasoning Models (NeurIPS 2025 · similarity 0.89)
• In-Context Sharpness as Alerts: An Inner Representation Perspective for Hallucination Mitigation (ICML 2024 · similarity 0.89)
• Mitigating Hallucinations in Large Language Models via Causal Reasoning (AAAI 2026 · similarity 0.89)
• SLVR: Structured Latent Visual Reasoning via Human-like Reasoning Flows (NeurIPS 2026 · similarity 0.89)