The paper investigates how multimodal large language models (MLLMs) handle conflicting evidence presented in text, image, or both forms. Across 13 MLLMs and two datasets, the authors find that models are not robust to knowledge conflict: they tend to accept contradictory image evidence more readily than contradictory text, and when both modalities conflict the preference is arbitrary, depending on input order, model, and dataset. The instability degrades multimodal retrieval-augmented generation and can be exploited by adversarial attacks, while simple mitigation techniques such as prompting, steering, and direct preference optimization largely fail, with supervised fine‑tuning offering only moderate improvement.
arXiv:2606. 31876v1 Announce Type: new Abstract: To improve safety in Large Language Models (LLMs) we can either perform post-training alignment or exploit refusal directions in the activation space.
TokenSwap introduces a method to benchmark and reduce the modality gap in multimodal large language models (MLLMs) by interleaving semantically aligned images with text tokens. By transforming text-only benchmarks like MMLU into image‑interleaved versions (TokenSwap‑Bench), the authors reveal a pervasive performance drop—averaging 19.6%—across 42 MLLMs, with reasoning models showing smaller gaps. Training with TokenSwap mitigates this gap while maintaining strong text‑only and vision‑language performance.
The paper investigates why multimodal large language models (MLLMs) struggle with vision‑centric tasks when visual evidence conflicts with pretrained language knowledge. Using image reconstruction and a new WhatIfVis benchmark, the authors show that MLLMs preserve coarse‑grained visual attributes but fail to consistently use them, and that supervised fine‑tuning and activation patching can improve controllability of visual context sensitivity. The study demonstrates that the main bottleneck lies in the models’ inability to reliably regulate their reliance on visual evidence rather than in visual perception itself.
arXiv:2609.00293v1 Announce Type: new Abstract: We investigate how vision-language models (VLMs) handle context-memory conflicts; that is, situations in which the model is given information in contex...
arXiv:2608. 05381v1 Announce Type: new Abstract: Current Multimodal Large Language Models (MLLMs) can process diverse sensory inputs, yet their reasoning remains heavily biased toward a dominant modality, resulting in brittle cross-modal reasoning.
arXiv:2608. 05131v1 Announce Type: cross Abstract: On-Policy Self-Distillation (OPSD) has become a standard post-training approach for improving visual reasoning in multimodal large language models (MLLMs).
arXiv:2609.22234v1 Announce Type: cross Abstract: Instruction hierarchy (IH) alignment teaches language models to prioritize higher-level instructions when inputs conflict. While studied primarily in...
arXiv:2609.36798v1 Announce Type: cross Abstract: Omni-modal large language models (LLMs) are expected to answer a question using the modality it explicitly refers to. However, existing training para...
arXiv:2608. 00076v2 Announce Type: replace-cross Abstract: Multimodal large language models (MLLMs) increasingly support high-stakes decision making by combining complementary information from images and text.
The paper investigates cross‑modal safety drift in multimodal large language models, where a harmless text query paired with a visual image can trigger harmful responses. Empirical analysis identifies unsafe response patterns and shows that visual cues receive limited attention, weakening refusal mechanisms. The authors introduce Safety‑Awareness Representation Transfer (SRT), a lightweight method that transfers safety signals from text processing to mitigate cross‑modal drift while maintaining model utility.
The paper identifies that in multimodal learning, optimization often produces asymmetric certainty gains, with the stronger modality becoming more confident than the weaker one, which leads to imbalanced contributions and suboptimal performance. The authors attribute this issue to unimodal characteristics and propose a Max Confidence Regularization (MaxCR) method that tracks each modality’s semantic confidence via a nonlinear sparsity measure and applies max suppression and excitation to balance confidence levels. Experiments on standard datasets demonstrate that MaxCR improves overall performance compared to state‑of‑the‑art multimodal baselines.
The paper introduces Inverted Asymmetric Fusion (IAF) to address strong-modality collapse in multimodal learning, where dominant modalities are degraded during fusion. IAF preserves the dominant modality by passing it unchanged and letting weaker modalities attend to it, while also strengthening weaker modalities via Modality-Aware Knowledge Distillation. Experiments on MultiHuSE, UR-FUNNY, and MUStARD show that IAF maintains unimodal performance and improves over the best unimodal baseline by up to 8.25%.