Tag
Introduces Policy Gradient Steering (PGS), a method that formulates activation steering as a reinforcement learning problem, using policy gradients to construct removable, composable steering vectors from behavioral objectives. Validated in gridworld, chess puzzle, and football environments.
This paper investigates how the choice of source activations influences activation steering in language models, finding that execution-boundary states (where the model is about to produce target behavior) yield stronger signals, and introduces tail subtraction to improve steering stability.
This paper introduces EHR-MPC, a framework that decouples learning patient dynamics from treatment optimization by training a generative digital twin of patients using electronic health records, then applying model predictive control at inference time, achieving comparable or improved performance over RL baselines on a multi-hospital sepsis cohort.
This paper introduces CogniConsole, an architectural abstraction that externalizes inference-time control for LLMs, demonstrating through controlled probes that increasing structural scaffolding reduces output variance and failure rates, challenging the assumption that reliability stems solely from model capability.
Introduces Stochastic Token Steering (STS) and Stochastic Block Steering (SBS) for LLM activation steering, which probabilistically gate steering signals per token or per sequence. Shows that steering only 50% of tokens recovers most of the dense-steering effect while preserving fluency, and that the behavioral outcome is rate-limited by cumulative signal dosage.
This paper introduces a two-stage inference-time budget control method for LLM search agents, using Value-of-Information scores to optimize tool-call and token allocation during multi-hop question answering.