Challenge: Existing methods require explicit safety labels or contrastive data, yet visual inputs enable harmful outputs.
Approach: They propose a visual self-fulfilling alignment mechanism that fine-tunes vision-language models on neutral VQA tasks without any safety labels.
Outcome: The proposed approach reduces attack success rate, improves response quality, and mitigates over-refusal while preserving general capabilities.

Similar Papers

Reimagining Safety Alignment with An Image (2025.emnlp-main)

Copied to clipboard

Challenge: Existing approaches to large language models face inefficiency, fragility, or architectural constraints, resulting in inefficient performance and heightened over-refusal in cross-modal tasks.
Approach: They propose an optimization-driven visual prompt framework that enhances security and reduces over-refusal at the same time.
Outcome: The proposed framework enhances security and reduces over-refusal while maintaining robust safety while reducing unnecessary denials.
MLLM-Protector: Ensuring MLLM’s Safety without Hurting Performance (2024.emnlp-main)

Copied to clipboard

Challenge: MLLMs are deployed on limited image-text pairs, which makes them more vulnerable to catastrophic forgetting of their original abilities during safety fine-tuning.
Approach: They propose a plug-and-play strategy that detects harmful visual inputs and transforms harmful ones into harmless ones.
Outcome: The proposed approach mitigates the risks posed by malicious visual inputs without compromising the original performance of MLLMs.
Re-Align: Aligning Vision Language Models via Retrieval-Augmented Direct Preference Optimization (2025.emnlp-main)

Copied to clipboard

Challenge: emergence of large Vision Language Models (VLMs) has broadened the capabilities of single-modal Large Language Model (LLM) but VLMs are prone to significant hallucinations, especially in the form of cross-modal inconsistencies.
Approach: They propose a new alignment framework that leverages image retrieval to integrate both textual and visual preference signals.
Outcome: The proposed framework mitigates hallucinations more effectively than previous methods . it maintains robustness and scalability across a wide range of VLM sizes and architectures .
Unraveling and Mitigating Safety Alignment Degradation of Vision-Language Models (2025.findings-acl)

Copied to clipboard

Challenge: LLaVA-7B demonstrated a decline in safety alignment ability on multi-modal inputs compared to its LLM backbone.
Approach: They propose a method to recover alignment ability from LLM backbone while preserving functional capabilities of VLMs.
Outcome: The proposed framework recovers alignment ability that is inherent in the LLM backbone with minimal impact on fluency and linguistic capabilities of pre-trained VLMs.
VFA: Empowering Multilingual MLLMs via Vision-Free Adaptation (2026.acl-long)

Copied to clipboard

Challenge: Multimodal large language models have advanced rapidly, yet most remain English-centric . scaling multilingual multimodal instruction tuning is limited by the scarcity and high cost of non-English image–text supervision.
Approach: They propose a framework that decouples multilingual language enhancement from visual alignment by composing complementary task vectors over a shared LLM backbone.
Outcome: The proposed framework achieves competitive performance with a fully multimodally trained model using less than 2% of the text data.
Empowering Reliable Visual-Centric Instruction Following in MLLMs (2026.findings-acl)

Copied to clipboard

Challenge: Existing benchmarks for evaluating instruction-following capabilities focus on verbal instructions in the textual modality.
Approach: They propose to incorporate vision-dependent constraints into instruction design to enable a more rigorous assessment of how well MLLMs align their outputs with both visual input and textual instructions.
Outcome: The proposed benchmark incorporates vision-dependent constraints into instruction design, enabling a more rigorous and fine-grained assessment of how well MLLMs align their outputs with both visual input and textual instructions.
Do VLMs Have a Moral Backbone? A Study on the Fragile Morality of Vision-Language Models (2026.findings-acl)

Copied to clipboard

Challenge: Vision-Language Models (VLMs) have advanced multimodal learning, driving progress in cross-modal reasoning.
Approach: They propose to examine moral robustness of vision-language models by analyzing their moral stances under multimodal perturbations.
Outcome: The proposed model-agnostic multimodal perturbations expose VLMs to a variety of moral vulnerabilities, including a sycophancy trade-off where stronger instruction-following models are more susceptible to persuasion.
VLMGuard-R1: Proactive Safety Alignment for VLMs via Reasoning-Driven Prompt Optimization (2026.findings-acl)

Copied to clipboard

Challenge: integrating vision and language models with safety standards is essential to mitigate multimodal complexity . integrating visual inputs with vision and text unveils subtle threats beyond the reach of conventional safeguards .
Approach: They propose a framework that combines vision and language to provide a multimodal reasoning-driven prompt rewriting.
Outcome: The proposed framework outperforms baseline models on five benchmarks with six VLMs.
Improving Alignment in LVLMs with Debiased Self-Judgment (2025.findings-emnlp)

Copied to clipboard

Challenge: Existing methods for aligning LVLMs rely on external datasets, human annotations or complex post-processing.
Approach: They propose a method that generates a debiased self-judgment score for LVLMs . this self-evaluation metric is created internally by the model without external resources .
Outcome: The proposed approach outperforms existing methods in reducing hallucinations and safety concerns.
When Helpers Become Hazards: A Benchmark for Analyzing Multimodal LLM-Powered Safety in Daily Life (2026.findings-acl)

Copied to clipboard

Challenge: Safety impact of Multimodal Large Language Models (MLLMs) on human behavior is evaluated in this study.
Approach: They propose a safety-warning-based evaluation framework that encourages models to provide clear and informative safety warnings, rather than generic refusals.
Outcome: The proposed safety-warning-based evaluation framework encourages models to provide clear and informative safety warnings, rather than generic refusals.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations