Paper replications¶
The replications/
directory reproduces published steering papers with EasySteer notebooks. Each folder
contains a README, the notebook(s), and the extracted vectors.
| Folder | One-liner | Category |
|---|---|---|
bipo/ |
Bi-directional preference optimization vectors steering power-seeking behavior | Personalization |
cast/ |
Conditional activation steering to program refusal (CAST) | Safety |
controlingthinkingspeed/ |
Speeding up / slowing down reasoning-model thinking on MATH500 | Reasoning |
creative_writing/ |
Steering LLMs to evaluate and amplify creativity | Style |
fractreason/ |
Fractional reasoning via latent steering vectors for inference-time compute | Reasoning |
improve_reasoning/ |
Representation-engineering vectors that improve reasoning performance | Reasoning |
lm_steer/ |
Word embeddings as steers for language models (LM-Steer, GPT-2) | General |
loreft/ |
ReFT: train and apply LoReFT representation finetuning | General |
refusal_direction/ |
Refusal is mediated by a single direction (DiffMean ablation) | Safety |
sae_entities/ |
SAE entity-knowledge directions and hallucination awareness | Reality |
sake/ |
SAKE: steering activations for knowledge editing | Knowledge |
seal/ |
SEAL: steerable reasoning calibration (execution/reflection/transition vectors) | Reasoning |
sharp/ |
SHARP: steering hallucination in LVLMs via representation engineering (EMNLP 2025) | Reality |
steerable_chatbot/ |
Personalizing LLMs with preference-based activation steering | Style |
steermoe/ |
SteerMoE: expert (de)activation steering of MoE routers on OLMoE-1B-7B (arXiv:2509.09660) | MoE |
Contributions of new replications are welcome — see Contributing.