AGENTIC-SYSTEMSApril 3, 20261 min read14 views

Robust Multimodal Safety via Conditional Decoding

Researchers have introduced CASA, a novel conditional decoding strategy that reduces attack success rates on multimodal AI models by over 97% without requiring modality-specific fine-tuning.

Computer Science > Machine Learning

Title:Robust Multimodal Safety via Conditional Decoding

View PDF HTML (experimental)Abstract:Multimodal large-language models (MLLMs) often experience degraded safety alignment when harmful queries exploit cross-modal interactions. Models aligned on text alone show a higher rate of successful attacks when extended to two or more modalities. In this work, we propose a simple conditional decoding strategy, CASA (Classification Augmented with Safety Attention) that utilizes internal representations of MLLMs to predict a binary safety token before response generation. We introduce a novel safety attention module designed to enhance the model's ability to detect malicious queries. Our design ensures robust safety alignment without relying on any external classifier or auxiliary head, and without the need for modality-specific safety fine-tuning. On diverse benchmarks such as MM-SafetyBench, JailbreakV-28k, and adversarial audio tests, CASA lowers the average attack success rate by more than 97% across modalities and across attack types. Our empirical evaluations also show that CASA maintains strong utility in benign inputs, a result validated through both automated and human evaluations (via 13 trained annotators). Together, these results highlight CASA as a simple and generalizable framework to improve multimodal LLM safety.

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

Source: arXiv cs.AI Recent

More in this category

agentic-systems

Probabilistic Concept-Aware Steering for Trustworthy LLM Inference

Researchers have introduced Probabilistic Concept-Aware Steering (PCS), an inference-time intervention framework for LLMs. PCS provides fine-grained, safety-oriented semantic steering while preserving original task competence.

agentic-systems

MUX: Continuous Reasoning via Multiplexed Tokens

Researchers introduced MUX, a novel latent reasoning method that compresses discrete text-based reasoning steps into continuous multiplexed tokens. By enabling lossless superposition, MUX significantly boosts LLM reasoning efficiency and speed across complex problem-solving tasks.

agentic-systems

Semantic Cooperative Games for Contribution Attribution in LLM-Based Multi-Agent Systems

Researchers introduce Semantic Cooperative Games (SCG) and the SLIC algorithm to evaluate agent contributions in LLM-based multi-agent systems without re-running models, reducing computational costs by 93.3%.

agentic-systems

ProbSPARQL: Querying Knowledge Graphs with Multi-dimensional, Uncertain Numeric Data

Researchers have introduced ProbSPARQL, an upward-compatible SPARQL extension designed to query multi-dimensional and uncertain numeric measurement data within Knowledge Graphs, providing significant query performance gains for complex industrial applications.

agentic-systems

Position: AI/ML Deepfake Research is Misaligned with AI-Generated Non-Consensual Intimate Imagery (AIG-NCII)

Current AI/ML research on deepfakes focuses primarily on epistemic harms like fake news and scams, leaving a dangerous gap in addressing AI-generated non-consensual intimate imagery (AIG-NCII) and its dignity harms to victims.

NOW LET US Related – Cross-Dialect Generalization Without Retraining: Benchmarks and Evaluation of Schema-Derived Constrained Decoding for MLIR

agentic-systems

Cross-Dialect Generalization Without Retraining: Benchmarks and Evaluation of Schema-Derived Constrained Decoding for MLIR

Researchers introduced schema-derived constrained decoding for MLIR, enabling a 1.7B small language model to match or beat 15B-34B LLMs without retraining, while generating code 8x-25x faster.