New Papers Probe LLM Safety Inside the Model, Not Just at the Interface

A cluster of new arXiv preprints examines how to make large language models safer by looking inside the model rather than only wrapping it in external guardrails. One paper argues that current safety infrastructure adds latency and compute overhead, which is a problem for resource-constrained, time-critical systems; it proposes detecting harm via latent states. Another independently makes a similar case in the automotive domain, where a voice assistant must decide whether a natural-language request should be authorized to control vehicle functions.

A third paper takes the internal-state idea further, showing that local sparsity can enable unsupervised safety detection. That matters because supervised methods assume access to unsafe training data, yet new attacks and harm categories keep emerging. The authors of that work claim their approach can catch harms not captured by models trained on known examples.

Two other papers focus on the structure of harm itself. One asks what role fine-grained harm signals play, noting that internal harmfulness representations vary across risk categories while also sharing a common general component. Another proposes a modular framework for targeted harm mitigation rather than broad alignment tuning. A sixth preprint, SAFARI, introduces an industrial benchmark for LLM-assisted hazard analysis in automotive functional-safety workflows, suggesting that LLM reliability in regulated engineering settings remains underexplored.

Where the papers agree is on the value of internal representations and on the inadequacy of purely external or purely supervised safety measures. They differ in emphasis: some target deployment efficiency, others focus on unsupervised detection, and one is explicitly an industrial benchmark. None of the six appears to be a direct continuation of another, but together they point to a research direction that treats safety as a property of the model's internal states, not just its outputs.