Tag
This paper presents the first comparative evaluation of training-free methods for personalizing toxicity sensitivity in language models at inference time, showing that all methods reduce alignment error by 28-47% but reveal a trade-off between alignment, personalization, and language quality.
This paper introduces BlendIn, an inference-time alignment framework that uses probabilistic model blending to assess guidance reliability and proportionally weight model contributions, achieving up to 50% performance improvement by avoiding harmful interventions.
The paper identifies off-manifold drift in guided flow models under compositional rewards and proposes Conflict-Aware Additive Guidance (CAR), a lightweight method that dynamically resolves gradient conflicts to improve generation fidelity without retraining.
This paper studies harness design for LLM agents, separating it into task decomposition and guided execution, and shows that more elaborate harnesses are not uniformly better; it reveals failure modes and proposes partial harnesses as effective.