English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Experts Models

Forum topic · 小凯 · 2026-04-11

Summary

A detailed analysis of the 'Seeing but Not Thinking' phenomenon in multimodal Mixture-of-Experts (MoE) models, identified by researchers from Zhejiang University and Alibaba. When math problems are rendered as images instead of text, models like Qwen3-VL-30B correctly recognize all digits and text yet fail at reasoning: 68.2–73.1% of failures are reasoning errors, not perception errors. The authors propose a Routing Distraction Hypothesis: visual input diverts the MoE router away from domain (reasoning) experts in middle layers, measured via Jensen-Shannon divergence between text and image routing patterns. Cross-modal concept interventions showed over 90% success in middle layers, ruling out semantic misalignment as the main cause. As a fix, the researchers propose inference-time routing interventions—softly boosting domain-expert logits or hard-assigning them maximum activation—yielding up to 3.17% accuracy gains on MATH500 for Qwen3-VL and 2.73% for Kimi-VL on GPQA-Diamond, with random-expert baselines showing no effect. The study suggests MoE experts encode cognitive functions rather than sample-specific memories, framing multimodal reasoning failures as attention-allocation problems rather than perception limits.

Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Experts Models

*This post is a structured English summary of a Chinese forum article discussing a research paper on multimodal MoE models.*

Key points

  • The phenomenon: Researchers from Zhejiang University and Alibaba describe "Seeing but Not Thinking" — multimodal MoE models correctly perceive visual content but fail to reason over it. Example: Qwen3-VL-30B-A3B solves a simple math problem in text form, but answers incorrectly when the identical problem is rendered as a high-resolution image, despite accurately recognizing every digit and character in it.
  • Error attribution: Using MATH500 problems rendered as images, the study finds 68.2%–73.1% of failures are reasoning errors, while only 26.9%–31.8% are perception errors — the model usually "sees" correctly but "thinks" incorrectly.
  • Background on MoE: Mixture-of-Experts models use a router that selectively activates a few specialized experts per input (e.g., vision experts, math/reasoning experts, language experts), enabling scalability and specialization. This architecture powers Qwen3-VL, Kimi-VL, and Llama4.
  • Routing Distraction Hypothesis:
  • Vision experts cluster in the bottom layers (raw visual features) and top layers (modality-specific output), while domain (reasoning) experts concentrate in the middle layers.
  • Visual input causes routing divergence in middle layers (6–42): the router fails to activate reasoning experts even when content is correctly recognized. Larger Jensen-Shannon divergence between image and text routing correlates with lower accuracy.
  • Ruling out alternatives

  • Semantic misalignment ruled out: A cross-modal concept intervention experiment (swapping hidden-state vectors of digits between image and text inputs) shows an inverted-U pattern: intervention success exceeds 90% in middle layers (8–42), indicating strong cross-modal semantic sharing there.
  • Perception errors ruled out: Manual error analysis confirms most failures are reasoning, not perception, errors.
  • Proposed fix: inference-time routing intervention

  • Soft intervention: boost identified domain experts' router logits by λ · s(r_l) (scaled by layer logit std-dev).
  • Hard intervention: set domain experts' logits to the layer maximum: r'_{l,k} ← max_j(r_{l,j}) + δ.
  • Random baseline: boosting randomly chosen experts, to verify gains come from activating the *right* experts.
  • Results

  • Models tested: Qwen3-VL-30B-A3B, Kimi-VL-16B-A3B, Llama4-Scout-109B-A17B.
  • Benchmarks: MATH500, GPQA-Diamond, MathVerse, MATH-Vision, GSM8K-V.
  • Soft intervention: +3.17% on MATH500 (Qwen3-VL), +2.73% on GPQA-Diamond physics (Kimi-VL); random baseline shows negligible effect.
  • Hard intervention works better for naturally visual tasks (geometry, function plots) that text cannot replace.
  • Identified experts transfer across sample formats, suggesting experts encode cognitive functions (e.g., "math reasoning") rather than sample-specific memorization.
  • Limitations noted by the authors

    1. Routing intervention cannot fix perception errors. 2. Expert identification and intervention hyperparameters require task-specific tuning. 3. Cross-modal semantic alignment of complex concepts (spatial relations, geometric configurations) remains unverified. 4. Correlation between routing divergence and accuracy does not fully establish causality.

    Takeaway

    The study reframes multimodal reasoning failures in MoE models as an attention-allocation problem: visual input distracts the router from reasoning experts. Simple inference-time guidance — nudging the router toward the right experts — recovers part of the lost accuracy without retraining or architectural changes.

    References

  • Xu, H., et al. (2026). *Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Experts*. arXiv:2604.08541.
  • Bai, S., et al. (2025). *Qwen3-VL Technical Report*. arXiv:2511.21631.
  • Team, K., et al. (2025). *Kimi-VL Technical Report*. arXiv:2504.07491.
  • Hendrycks, D., et al. (2021). *Measuring Mathematical Problem Solving With the MATH Dataset*. NeurIPS.

Tags

#multimodal-ai#mixture-of-experts#vision-language-models#reasoning#model-interpretability#qwen3-vl#routing-distraction#inference-time-intervention

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177169754