Skip to content

Understanding and Mitigating Hallucinations in Multimodal Chain-of-Thought Models

🕒 Published (v1): 2026-03-28 08:56 UTC · Source: Arxiv · Venue: CVPR 2026 · link

Why this paper was selected

CVPR 2026; diagnoses root cause and mitigates hallucination in multimodal chain-of-thought

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Multimodal Chain-of-Thought (MCoT) models hallucinate primarily during divergent thinking steps—associative reasoning segments that rely on internal priors rather than visual evidence—at ~5× the rate of normal reasoning steps. The authors introduce visual entropy, a training-free, entropy-based score that detects these steps at inference time and penalizes high-entropy token choices via a modified decoding objective. The method outperforms all baselines on hallucination benchmarks while simultaneously improving perception and reasoning accuracy.

Problem

Prior work attributes MCoT hallucinations to progressive visual attention decay as reasoning chains lengthen, treating MCoT models as functionally equivalent to traditional LVLMs with implicit reasoning. This paper argues MCoT models have a unique hallucination mechanism: fabricated content is not uniformly distributed across reasoning steps but concentrated in specific associative reasoning steps, a phenomenon unexplained and unaddressed by attention-decay-based methods.

Method

Hallucination analysis. The authors separate model outputs into thinking (<think>) and answer (<answer>) segments. They show via co-occurrence analysis (\(\rho > 0.96\), \(R^2 > 0.92\)) that answer hallucinations are strongly predicted by thinking hallucinations, and via attention visualization that answer tokens allocate <0.04 attention ratio to image tokens, directing attention almost entirely to the thinking context.

Divergent thinking identification. Reasoning chains are segmented and labeled as normal vs. divergent (associative reasoning that generates content beyond visual evidence) using GPT-5-assisted annotation verified by human annotators. Divergent steps show ~5Ă— higher hallucination rate.

Visual entropy. For token \(y_t\), the visual activation probability \(p_v(y_t) \in \mathbb{R}^m\) is extracted from the language head applied to visual hidden states. Visual entropy is:

\[E(y_t, v) = -\frac{\sum_{i=1}^{m} p_{v,i}(y_t)\log p_{v,i}(y_t)}{\log m}\]

normalized to \([0,1]\). High visual entropy indicates the model is uncertain about which visual token drives the prediction—a proxy for divergent thinking.

Entropy-constrained decoding. When \(E(y_t, v) > \gamma\) (detected divergent step), the decoding distribution is modified:

\[\hat{p}_t(\cdot|v,q,y_{<t}) = p_t(\cdot|v,q,y_{<t}) \cdot e^{-\alpha \cdot E(\cdot,v)}\]

penalizing tokens with high visual entropy. At normal thinking steps the distribution is unchanged. Visual token probabilities \(p_v\) are pre-computed during prefilling, so decoding overhead is negligible. No training or model weight modification is required.

Key Contributions

  • Empirical finding that hallucinations propagate from thinking to answers via self-attention, with divergent thinking steps being ~5Ă— more hallucination-prone than normal steps.
  • Visual entropy metric that localizes divergent thinking steps with pseudo-\(R^2 > 0.9\) (McFadden) in a logistic regression classifier.
  • Training-free, entropy-constrained decoding intervention compatible with greedy, nucleus sampling, and beam search.
  • Demonstrated compatibility and complementarity with existing LVLM hallucination mitigators (DoLa, VCD), boosting their performance on MCoT models without degrading reasoning.

Results

  • GRIT-3B: CHAIRS \(23.8 \to 16.0\) (best prior: FlexAC 19.2); CHAIRI \(10.5 \to 5.5\) (47.6% relative reduction over backbone).
  • PixelReasoner-7B: CHAIRS \(22.0 \to 15.4\) (FlexAC: 18.8); MathVista \(62.6 \to 64.0\); VSR \(82.4 \to 84.3\).
  • R1-Onevision-7B: CHAIRS \(23.2 \to 15.8\) (FlexAC: 18.4); MMStar \(52.8 \to 54.2\).
  • Visual entropy intervention alone reduces baseline CHAIRS by ~30%.
  • Divergent-only intervention outperforms both all-thinking and normal-only intervention variants (Table 3), confirming targeting specificity matters.
  • DoLa+Ours and VCD+Ours recover reasoning degradation caused by those methods while further reducing hallucinations.

Limitations

  • Divergent thinking labels were produced semi-automatically (GPT-5 + human annotators); generalizing the concept to new model families may require re-annotation.
  • Two hyperparameters (\(\gamma\), \(\alpha\)) must be set; while claimed to be model-agnostic, the sensitivity analysis is limited to the paper's three backbone models.
  • Evaluation is restricted to 7B-scale models; behavior at larger scales (e.g., 70B+ MCoT models) is untested.
  • The method requires access to the language head's visual token probability \(p_v\), which may not be exposed in all inference frameworks or closed-source deployments.

Relevance to Vision-Language Models

This work directly addresses a failure mode of the emerging MCoT paradigm (R1-style reasoning in VLMs), which is increasingly the dominant architecture for complex visual reasoning. The finding that explicit reasoning chains can amplify hallucinations—rather than suppress them—has important implications for VLM safety and reliability. The visual entropy mechanism provides a mechanistic, interpretable bridge between internal model states and observable hallucination behavior, a gap that most post-hoc or contrastive decoding approaches in the VLM literature have not addressed. For VLM researchers, the compatibility result with DoLa and VCD suggests visual entropy can serve as a lightweight plug-in for any MCoT deployment.