concept
active
concept:inference-time-steeringInference-Time Steering
Subtracting a scaled persona vector from hidden states at each decoding step to reduce trait expression after finetuning
Neighborhood — ranked by edge-count
Papers (1)
paper
Concepts (1)
concept
- Inference-Time Controlrelated_toApproach to steering LLM behavior without modifying weights, offering flexibility and composability
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Paradigm of improving model outputs through more computation at inference time; VS is presented as a diversity-oriented alternative
- Method by Li et al. 2023a that adds static vectors to model activations at inference time to steer behavior
- Inference-Time Intervention: Eliciting Truthful Answers from a Language Model (Li et al., 2023)concept0.777Safety intervention that relies on activation modification, which ESR might undermine
- The process of inferring causes of sensory inputs, a key aspect of the free-energy minimization scheme.
- The perspective that LLM inference decomposes into distinct computational stages, which the paper extends to looped models
- Current steering applies fixed strength; dynamic uncertainty-aware steering during inference is an open gapquestion0.759Research gap identified in limitations/future work section connecting uncertainty findings to practical improvement
- Algorithmic framework for probabilistic inference in graphical models.