TraceRouter++ framework realigns LLM safety pathways using optimal transport

Researchers have introduced TraceRouter++, a novel framework designed to enhance the safety of large foundation models (LFMs) by realigning causal semantic routing pathways. Unlike existing methods that focus on localized neuron suppression, TraceRouter++ identifies and modifies distributed, cross-layer circuits responsible for harmful outputs. The framework employs a Discover-Trace-Realign paradigm, using sparse autoencoders and feature influence scores to pinpoint sensitive pathways and then applying optimal transport-based soft routing realignment to redirect unsafe activations while preserving general utility. Experiments across various LLMs, MLLMs, and diffusion models demonstrate TraceRouter++'s superior adversarial robustness compared to baseline methods, achieving a better safety-utility trade-off. AI

IMPACT This research offers a novel approach to improving AI safety by realigning internal model pathways, potentially leading to more robust and secure AI systems.

RANK_REASON The cluster contains a research paper detailing a new method for AI safety. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• Diffusion Models
• Hugging Face
• Large Foundation Models
• large language models
• Multimodal Large Language Models
• optimal transport
• Shangze Li
• Sparse Autoencoders
• TraceRouter++

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →