Skip to content

When Thinking Hurts: Epistemic Signals in the Reasoning Chains of Visual Language Models

🕒 Published (v1): 2026-07-09 02:19 UTC · Source: Arxiv · Venue: ICML 2026 · link

Why this paper was selected

ICML 2026; first empirical characterization of reasoning-chain uncertainty in thinking-mode VLMs — calibration insight

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Thinking-mode VLMs that generate explicit <think> chains before answering collapse their answer-token entropy, rendering conventional uncertainty quantification (UQ) methods useless under greedy decoding. This paper provides the first controlled, cross-family empirical characterization of this collapse, and shows that entropy and length of the thinking chain itself restore discriminative power at zero additional inference cost. A practical abstention gate built on chain signals raises accuracy from 71.0% to 93.8% at 62.7% coverage on POPE adversarial.

Problem

Standard UQ for VLMs estimates uncertainty from the answer token distribution \(H(a \mid x)\). Thinking-mode VLMs condition the answer on an explicit reasoning chain \(z\), so the relevant quantity becomes \(H(a \mid x, z)\). When the chain fully pre-commits to an answer, \(H(a \mid x, z) \approx 0\), destroying variance in the answer distribution. On Qwen3-VL-8B-Thinking, answer entropy AUROC for hallucination detection drops from 0.899 (same family, non-thinking) to 0.492 (chance), despite identical questions, images, and evaluation protocol. Every existing output-level UQ method (semantic entropy, UMPIRE, HALP) is therefore inapplicable to this growing model class under greedy decoding.

Method

The study runs four models — Qwen3-VL-8B-Thinking, Qwen3-VL-8B-Instruct (non-thinking control, same vision encoder), GLM-4.1V-9B-Thinking, and InternVL3-8B — on 1,000 POPE adversarial samples and 951 HallusionBench questions under greedy decoding (do_sample=False).

Three signals are extracted from a single forward pass at no extra cost: - \(H_{\text{think}} = \frac{1}{|z|}\sum_t H_t\), mean per-token Shannon entropy \(H_t = -\sum_v p_t(v)\log p_t(v)\) over thinking tokens - \(H_{\text{ans}}\): mean entropy over answer tokens - \(L\): number of thinking tokens (chain length)

Hallucination detection is framed as binary classification (false-positive vs. true-positive within yes-predictions, i.e., conditioned on answer direction to remove the answer-direction confound) and evaluated with AUROC. Unconditional AUROCs are also reported for deployment realism. A VQAv2 300-sample pilot tests generalization to open-ended answers. Bootstrap 95% CIs (10,000 resamples) are provided for pooled estimates.

Key Contributions

  • First three-family empirical characterization of answer entropy collapse in thinking-mode VLMs, identifying three qualitatively distinct patterns: complete collapse (Qwen, AUROC 0.492), no collapse (GLM, 0.716), and selective thinking (InternVL3, chains on only 50% of samples).
  • Demonstration that thinking chain entropy \(H_{\text{think}}\) and chain length \(L\) are consistently superior or competitive uncertainty signals compared to \(H_{\text{ans}}\) across all thinking-mode models (pooled AUROC 0.786 and 0.776 respectively vs. pooled \(H_{\text{ans}}\) lying entirely below chance at CI = [0.386, 0.478]).
  • Documentation of structured abstention (12–22% of queries) with asymmetry toward absent-object queries and mechanistically distinct failure modes across benchmarks.
  • A zero-cost abstention gate: refusing unclear samples and the top 20% of committed predictions by chain length raises POPE accuracy from 71.0% to 93.8% at 62.7% coverage.

Results

  • Qwen3-VL-8B-Thinking vs Instruct (POPE adversarial, n=1,000): \(H_{\text{ans}}\) AUROC drops from 0.899 → 0.492; \(H_{\text{think}}\) = 0.647.
  • GLM-4.1V-9B-Thinking (POPE): \(H_{\text{ans}}\) = 0.716, \(H_{\text{think}}\) = 0.759 (no collapse, but chain still improves).
  • InternVL3-8B (POPE): \(H_{\text{ans}}\) = 0.675 (full), \(H_{\text{think}}\) = 0.608 (thinking-only subset, \(n_{\text{FP}}=17\), margin not statistically reliable); FP rate twice as high when chain is absent (13.1%) vs. present (6.0%).
  • Pooled conditioned AUROC (163 FPs, 1,951 samples): \(H_{\text{think}}\) = 0.786 [0.745, 0.825], chain length \(L\) = 0.776 [0.735, 0.815]; \(H_{\text{ans}}\) = 0.432 [0.386, 0.478] — entirely below chance.
  • Unconditional (POPE): \(H_{\text{think}}\) error AUROC = 0.636, chain length = 0.686.
  • HallusionBench (Qwen): \(H_{\text{think}}\) = 0.768 (\(|r|=0.53\)), \(H_{\text{ans}}\) = 0.602; FP chains are 83% longer (322.6 vs. 176.2 tokens) and 47% more uncertain than TP chains.
  • VQAv2 pilot (n=300): \(H_{\text{think}}\) = 0.680 vs. \(H_{\text{ans}}\) = 0.595; free-form subset: 0.733 vs. 0.467.
  • Abstention gate (POPE): 71.0% (baseline) → 91.1% (committed only) → 93.8% at 62.7% coverage; instruct model with answer entropy achieves 96.7% at matched coverage (gap narrows from 16 pp → ~3 pp at 40% coverage).
  • Epistemic hierarchy confirmed monotonic: committed-correct (L=59.9, \(H_{\text{think}}\)=0.218) < committed-wrong (74.4, 0.262) < abstaining-with-chain (124.3, 0.338) on POPE.

Limitations

  • Only 8–9B parameter models evaluated; 32B and 72B thinking-mode VLMs remain untested.
  • All experiments use greedy decoding; stochastic sampling (temperature > 0) may partially rehabilitate \(H_{\text{ans}}\) by preventing full pre-commitment and is identified as the single most actionable remaining experiment.
  • InternVL3-8B is only characterized at the headline POPE level; conditioned AUROCs, epistemic hierarchy, and abstention gate analyses are Qwen/GLM only.
  • "Confident hallucinations" — fast, low-entropy chains committing to wrong answers — represent 11/34 POPE FPs and 14/129 HallusionBench FPs that chain signals cannot detect.
  • No empirical comparison against sampling-based baselines (e.g., self-consistency); the contribution is zero-cost single-pass signals only.
  • Chain entropy may partly reflect inherent image ambiguity (aleatoric uncertainty) rather than model ignorance; epistemic/aleatoric decomposition not addressed.
  • Chain length thresholds do not transfer across benchmarks; thinking entropy is more transferable but still requires held-out calibration data per task.
  • Open-ended VQA evaluation is limited to a 300-sample VQAv2 pilot.

Relevance to Vision-Language Models

This paper directly challenges the assumption that standard entropy-based UQ applies uniformly across VLM architectures, identifying a concrete failure mode specific to the increasingly common thinking-mode design. For practitioners deploying thinking-mode VLMs (Qwen3-VL, GLM-4.1V, InternVL3), it provides immediately actionable guidance: discard answer entropy, use \(H_{\text{think}}\) or \(L\) from the same greedy pass already being run. The finding that collapse is training-recipe-dependent (Qwen collapses; GLM does not despite both generating explicit chains) raises open questions for model training that are directly relevant to anyone designing or fine-tuning VLMs with reasoning capabilities. The work also connects to broader concerns about CoT inducing overconfidence in VLMs, providing a principled uncertainty-theoretic framing for why thinking chains that improve accuracy can simultaneously degrade calibration signals.