Get Started
Home
Topics
Search
Library
Research questionHow can inference-time hidden-state steering of LLMs become efficient and extensible for deployment?Inference-time steering changes an LLM’s hidden states to alter its behavior without retraining. Existing frameworks can be computationally inefficient and difficult to extend or control, complicating applications such as mitigating overthinking and hallucinations.
AI
Alignment & Safety
Evaluation & Benchmarks
Inference Optimization
LLM Pretraining & Post-training
Mechanistic Interpretability
Research Paper
Latest papersRecent research connected to this question, newest first.EasySteer: A Unified Framework for High-Performance and Extensible LLM SteeringThe source describes EasySteer, a vLLM-integrated framework with pluggable interfaces for analysis- and learning-based steering, fine-grained parameter control, precomputed steering vectors across eight application domains, and an interactive demonstration system. It reports 10.8–22.3× speedups over existing frameworks and experiments on overthinking mitigation, hallucination reduction, and related applications.research paper · Sep 3, 2026IDEEA: training-free Input-Dependent stEEring via Activation cluster matchingApplies to LLM activation steering using selected attention-head activations. The source describes clustering positive and negative activation supports, matching clusters to construct conditional directions, and selecting a direction at inference time; evidence is reported on TruthfulQA using truth × information rate against an input-independent steering baseline.research paper · Sep 2, 2026
Related questions
How can we tell whether LLM hidden-state geometry reflects reasoning operations rather than lexical or positional cues?How can LLM orchestrators preserve continuous state when collaborating with non-language agents?How can deployable LLM guardrails maintain robust safety moderation without explicit reasoning’s latency and token overhead?How can LLM agents jointly adapt reasoning policies and hierarchical skill libraries during reinforcement learning?