The Safety Decision Before the Answer
ShallowDefend shows how explicit safety commitments improve multi-turn jailbreak resistance, while exposing the cost of rejecting legitimate requests.
ShallowDefend shows how explicit safety commitments improve multi-turn jailbreak resistance, while exposing the cost of rejecting legitimate requests.
Why more LLM assessments and formal-verifier feedback can fail to improve software assurance, and what deterministic evidence checks actually contribute.
TagInsert improves low-resource sequence tagging over left-to-right decoding, but BERT's accuracy, rare-label failures, and training costs limit its practical advantage.
Smaller recurrent-depth models compete with larger clinical language models, but task-specific results, schema transfer, and unmeasured serving costs complicate deployment decisions.
Value Probes improve multi-turn AI safety auditing, but much of the reported gain comes from structured conversation search before learned ranking is introduced.
Why successful simulations cannot establish almost-sure reachability, and what mathematical assumptions make certification of an unknown decision system possible.
A scalable method for reading and steering model concepts shows why classification accuracy alone cannot establish a stable or reliable internal feature.
A shared-autonomy framework separates proposals, evaluation, and execution, making reusable approval logic explicit without requiring a new planning algorithm.
A study of parameter-based AI memory finds that selecting the right information to store is a larger obstacle than retrieving it, with important limits for enterprise adoption.
BULLPEN shows how a model's actual responses can improve behavioral profiling at lower evaluation cost, while revealing why no single profiling channel is sufficient.