aiqbench analysis
TRACE introduces a novel safety framework that monitors and intervenes in real-time during LLM generation by editing token activations based on learned latent concepts. This approach provides an auditable record of risk detection while maintaining high performance on standard harmful content benchmarks like WildGuard and S-Eval. The method effectively reduces attack success rates without significantly degrading the model's ability to generate safe responses.
TRACE: Temporal Risk Assessment and Concept Editing for Safer Language Generation
arXiv:2610.15365v1 Announce Type: new Abstract: Large language models (LLMs) can produce harmful content while generating a response. A useful safeguard should not only detect this risk during generation but also intervene while leaving safe responses largely unchanged and provide an auditable record of the internal patterns supporting its decisions. We propose TRACE (Temporal Risk Assessment and Concept Editing), a framework that tracks harmful-response risk during generation, selectively edits token activations, and leaves an auditable trace of its detection and steering decisions. TRACE learns a codebook of recurring activation patterns, with each codebook vector representing a latent concept. At inference time, it assigns each token activation to the nearest codebook vector to obtain the corresponding concept. A gated recurrent unit (GRU) processes the activations and their assigned concept features to predict completed-response harmfulness and a monotone cumulative streaming risk. During steering, activations assigned to harmful concepts are steered along directions defined by nearby benign targets. An external LLM labels the learned concepts post hoc, making detection sequences and harmful-to-benign edit records interpretable. Experiments across three generators and the WildGuard and S-Eval benchmarks show that TRACE improves streaming F1 over Kelp in five of six comparisons while remaining competitive on completed-response detection. TRACE's selective steering reduces Harmbench attack success with little change in refusal on originally safe responses. The post-hoc audit reveals recurring semantic structure in the learned concepts and distinct traces associated with harmful detection, safe generation, and activation edits.
Abstract:Large language models (LLMs) can produce harmful content while generating a response. A useful safeguard should not only detect this risk during generation but also intervene while leaving safe responses largely unchanged and provide an auditable record of the internal patterns supporting its decisions. We propose TRACE (Temporal Risk Assessment and Concept Editing), a framework that tracks harmful-response risk during generation, selectively edits token activations, and leaves an auditable trace of its detection and steering decisions. TRACE learns a codebook of recurring activation patterns, with each codebook vector representing a latent concept. At inference time, it assigns each token activation to the nearest codebook vector to obtain the corresponding concept. A gated recurrent unit (GRU) processes the activations and their assigned concept features to predict completed-response harmfulness and a monotone cumulative streaming risk. During steering, activations assigned to harmful concepts are steered along directions defined by nearby benign targets. An external LLM labels the learned concepts post hoc, making detection sequences and harmful-to-benign edit records interpretable. Experiments across three generators and the WildGuard and S-Eval benchmarks show that TRACE improves streaming F1 over Kelp in five of six comparisons while remaining competitive on completed-response detection. TRACE's selective steering reduces Harmbench attack success with little change in refusal on originally safe responses. The post-hoc audit reveals recurring semantic structure in the learned concepts and distinct traces associated with harmful detection, safe generation, and activation edits.
Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2610.15365 [cs.AI]
(or arXiv:2610.15365v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2610.15365 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
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.
Full text under the publisher's open licence · 331 words
Related coverage
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · Sep 18, 2026
[AINews] The Future of Latent Space
Latent Space · Sep 25, 2026
[AINews] Opus 5.5 is good at explainer videos
Latent Space · Sep 29, 2026
Assessing Information Ecosystems : How Governments Can Get Ahead of Hybrid Threats
carnegieendowment.org · Aug 5, 2026