A library for making RepE control vectors

[ICLR 2025] General-purpose activation steering library

Steering vectors for transformer language models in Pytorch / Huggingface

A resource repository for representation engineering in large language models

[🏆 CHI26 Best Paper] CoBRA: Reproducible control of LLM agent behavior via classic social science experiments

KV Cache Steering for Controlling Frozen LLMs

Lightweight representation engineering dataflow operations for agent developers.

Agents that talk through model internals — activations & J-space — instead of text. No words pass between them. The lab on top of brainscope + hidden-directions.

[🔥 ICLR 2026] - Misaligned Roles, Misplaced Images: Structural Input Perturbations Expose Multimodal Alignment Blind Spots

[Under Review] Not All Tokens Are Equally Useful for Steering: Robust Directions and Prefix Steering

Activation steering and trait monitoring for HuggingFace transformers

Official code for "Activation Steering for Accent Adaptation in Speech Foundation Models" (Interspeech 2026). Parameter-free accent adaptation via mean-shift steering vectors — no weight updates, consistent WER reductions across 8 accents.

CRSM (Continuous Reasoning State Model): An asynchronous "System 2" architecture that implements Hierarchical State Sovereignty within a Mamba backbone. Unlike traditional search wrappers, CRSM uses Forward-Projected Planning and Sparse-Gated Injection to steer latent manifolds in real-time, decoupling strategic reasoning from token generation.

Mechanistic interpretability experiments on political control circuits, refusal behavior, concept steering, and late-decoder interactions in open LLMs.

Phase-aware LLM activation steering and linear probing. A memory-efficient, practical implementation of Representation Engineering (RepE) for safety research.

Steer2Adapt: data-efficient inference-time LLM adaptation by composing steering vectors via Bayesian optimization over a semantic prior subspace.

Pre-generation tool-call gating via linear probes on LLM hidden states. F1 ≈ 0.91–0.94 on BFCL v4, 14–22× faster than full generation. Cross-architecture transfer across Llama / Qwen / Phi / Mistral (3B–7B) with ≥96% retention.

Steering vectors with receipts: make one, catch one, deploy a calibrated one. pip install hidden-directions

Turn a knob inside a small open model instead of writing a prompt. A reproducible RepE and CAA steering harness with an honest benchmark, including the failures.