Hidden-State Steering
Techniques for controlling language model behavior through targeted interventions on hidden states, enabling systematic black-box optimization.
22 papers
Where this stands
The written synthesis of this thread is for subscribers. Subscribe.
How this thread developed
Oct 2026 · NetEase, Amazon
Internal confidence probe eclipses verbalized scores in calibration for reasoning models
shows that hidden-state probes for confidence can correct overconfident verbalized scores in reasoning models.
1 further paper
Oct 2026 · Microsoft Research
LLM judges often encode better preferences than their verdicts reveal
Reveals that LLM internal representations encode better preferences than their output judgments, motivating steering methods that access hidden states.
2 further papers
Oct 2026 · City University of Hong Kong, Stanford University
LLM persona prompts alter non-target attributes, breaking selective control assumptions
Reveals that persona prompts alter non-target attributes, challenging assumptions of selective control in behavior steering.
1 further paper
Oct 2026 · Cornell University
Positional encoding choice shifts retrieval from positional to semantic mechanisms
Shows that positional encoding choice shapes the internal mechanism used for retrieval, linking PE to the functional role of hidden states.
1 further paper
Oct 2026 · NVIDIA Research, University of Wisconsin–Madison
Steering an LLM feature does not prove the model uses it
Demonstrates that strong steering of a feature does not imply it causally participates in the model's normal computation, adding a critical caveat to steering-based interpretability.
4 further papers
Oct 2026 · Kent State University
Input-blind controls mimic oracle gains for layer programs in multiple-choice evaluation
Layer-skipping and repetition programs as input-blind controls match oracle gains in multiple-choice evaluation, revealing steering dynamics.
1 further paper
Oct 2026 · Carnegie Mellon University, Mila - Quebec AI Institute
Activation-based representations predict LLM value generalization across contexts
Uses activation-based representations to predict how training on one value alters behavior on many others (alignment generalization).
3 further papers
Oct 2026 · St. Jude Children's Research Hospital, Columbia University
Observational discovery of sparse autoencoder features does not predict their causal effects
Shows that observational feature discovery scores (e.g., SAE features) do not predict causal effects, informing intervention-based steering methods.
Oct 2026 · Peking University, Samsung Research
Alternating context edits and logit steering boosts frozen LLM accuracy
Introduces EMO, a training-free framework alternating context edits and confidence-constrained logit steering to boost frozen LLM accuracy.
9 of 22 papers shown