cs.CVNov 16, 2026

Words Speak Louder? Exploring the Language Shortcut of LLM-Driven Incomplete Multimodal Learning

Authors: Qiangchang Wang, Chong Xiao, Xiaokun Bai, Rundong He, Zhuang Qi, Xusheng Liang, Zhibin Wu, Wenhao Li

Organizations: Shandong University · The Hong Kong Polytechnic University · National University of Singapore · City University of Hong Kong · 

Abstract

Is a face or a voice still heard once the words are known? Incomplete multimodal learning increasingly uses large language models (LLMs) as fusion engines: modality features enter as token blocks and missing modalities become placeholders. We find that LLM-driven fusion can override non-verbal evidence it can read. Contrasting its full-input prediction with counterfactual views, the same model shown one modality or no content, we identify language shortcuts. We build MR-Bench: MR-Bench-A types every test sample of four sentiment benchmarks; MR-Bench-H contains 900 human-validated clips (620 language-shortcut, LS; 280 non-verbal-shortcut, NS) from seven sources, and every model is scored on all of them. Zero-shot MLLMs solve at most 32.7% of the LS clips; adding the transcript to the muted video lowers vision-language accuracy by 46.1-53.1 points. Yet inverting a text-only model's answers solves 94.4%, so LS accuracy alone cannot establish use of non-verbal evidence. Inside the fusion LLM, fused states align increasingly with the text view late in the decoder. Counterfactual-View Reliability learning (CoVeR) trains views as unimodal judges and withholds fused supervision where they disagree. It achieves state-of-the-art results on five benchmarks without extra inference cost in accuracy mode. On MR-Bench-H, it reaches 61.7% LS and 93.6% NS accuracy; robust mode raises LS accuracy to 63.0% while lowering NS accuracy to 84.8%. Among LS clips solved without text, adding text causes errors on 30.1% with CoVeR versus 32.2% with LLM-Fusion. Project page: https://peacelwh.github.io/CoVeR/.

Figures & tables

Explore similar work

Sep 29, 2026

Seeing What Should Be Heard: Diagnosing and Repairing Cross-Modal Shortcuts in Omni-Modal LLMs

Omni-modal large language models (LLMs) are expected to answer a question using the modality it explicitly refers to. However, existing training paradigms rarely verify whether models actually follow this modality, because multimodal inputs from the same sample often provide redundant evidence for the same answer. In this work, we uncover a pervasive cross-modal shortcut in omni-modal LLMs: when asked an audio-related question, models rely on the image as much as on the audio, and sometimes even more. To systematically diagnose this behavior, we introduce the Factorized Modality Diagnostic, which independently swaps audio and images between samples to isolate each modality's causal contribution. Across two model families in different settings, we find that this shortcut persists throughout supervised fine-tuning and reinforcement learning post-training, while judge-based RL may further amplify such reliance on irrelevant visual information. Based on this finding, we propose DMC-Repair, which trains models on the same kind of cross-modal swapped samples while assigning supervision according to the modality specified by the question. This prevents models from exploiting the spurious correspondence between modalities within the same clip. Experiments demonstrate that DMC-Repair reduces the image-induced share of the answer effect by 59.9%, effectively suppressing the cross-modal shortcut without compromising audio-question answering performance. The reduction in shortcut reliance generalizes across two model families and zero-shot to an unseen dataset and an unseen benchmark, and persists through subsequent post-training. Code is available at https://anonymous.4open.science/r/DMC-Repair.
Jun 16, 2026

MLLMs Get It Right, Then Get It Wrong: Tracing and Correcting Late-Layer Textual Bias

When vision contradicts text, multimodal large language models (MLLMs) consistently favor text, even when images provide clear evidence otherwise. This bias poses risks for applications requiring visual grounding, yet its cause remains unclear. In this paper, we uncover a surprising finding: models often get it right initially, forming correct vision-based predictions in their intermediate layers, before changing their minds and favoring text in the final output. We call this "late-layer textual override". The visual information is encoded, it simply does not survive to the output. More intriguingly, we find that how predictions change reveals whether they're correct: 85% of failures shift toward text, while 89% of successes shift toward vision. This directional signature enables a simple but powerful intervention: when we detect a confident visual prediction being suppressed, we restore it. We propose CALRD (Conflict-Aware Layer Reference Decoding), a training-free method that recovers overridden predictions at inference time. Experiments across five MLLMs of varying architectures demonstrate up to 9.4% absolute improvements on conflict benchmarks while largely preserving standard performance, without training or external knowledge. It recovers what the model already knew but failed to preserve.
Jul 28, 2026

Seeing or Knowing? Visual Context Sensitivity in Multimodal Large Language Models

Multimodal Large Language Models (MLLMs) achieve strong performance by integrating visual inputs with the rich priors of pretrained language models. However, they often fail on vision-centric tasks, especially when visual evidence conflicts with pretrained knowledge. We explore these failures separately using two diagnostic paradigms: (1) probing whether visual information is available, via image reconstruction, and (2) measuring multimodal context sensitivity, the extent to which the model follows visual context versus the language prior. To support the second, we introduce the WhatIfVis, a benchmark spanning five coarse-grained dimensions (spatial-temporal, color, count, size, and weight) whose questions admit answers from either the image or the prior. Our analysis yields three findings: (i) Coarse-grained visual evidence is preserved, as these attributes can be reconstructed from the final-layer image tokens of frozen MLLMs. Failures on questions about these attributes therefore point to post-perceptual utilization, rather than to degraded visual encoding during perception. (ii) Even when explicitly instructed to use or ignore visual evidence, vanilla models (without supervised fine-tuning on the WhatIfVis) show unstable visual context sensitivity. Supervised fine-tuning (SFT) improves this controllability and generalizes across domains, and activation patching further localizes the vision-versus-prior trade-off at architecture-specific depths across all six models. (iii) The vision-versus-prior trade-off is controllable along a learned vector. Applying this steering vector, even without any intent instruction, improves controllability over the vanilla model. Together, these results relocate the bottleneck, indicating that for the coarse attributes we study, MLLMs encode the visual evidence but cannot reliably control their reliance on it.