steering-vectors

Runtime control of LLM agent behaviors through activation steering vectors. More calibrated than prompting.

Iterative Sparse Matrix Steering: Closed-Form Subspace Alignment for Multi-Layer LLM Control (No SGD required).

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

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

Functional emotional architecture for LLMs — 42 systems, 1994 tests, 27 psychological theories. Emergent emotions via 7 ANIMA pillars: predictive processing, global workspace, autobiographical memory, ontogenic development, motivational drives, emotional discovery, computational phenomenology.

Mechanistic interpretability on Apple Silicon: steering vectors, residual capture, and SAE analysis for MLX models

CUDA-graph-safe per-request activation steering plugin for vLLM. pip install hotwire-vllm

Official implementation of "Beyond Multiple Choice: Evaluating Steering Vectors for Summarization" (Findings of EACL 2026).

SAE-steered browser agents with a live interpretability HUD. Two SAE feature edits + a 1-line prompt lift Llama-3.1-8B from 10% to 75% on a 60-trial benchmark, closing 72% of the cross-scale gap to Llama-3.3-70B baseline at ~1/8 the inference cost.

一种通过替换模型自答为教师回答来提取和施加行为向量的工具包。 GHOST: Ghost-Host Output Steering Toolkit — Extract and apply behavior vectors from language models by swapping their own answers with teacher answers.

Investigating honesty, deception and steering in large language models. Replicating and extending the MASK honesty benchmark on frontier models, working toward internal representation analysis and activation steering for honesty.

Extract and inject activation steering vectors in HuggingFace LLMs — mean-diff & PCA, MPS-friendly

Repository for paper "Understanding (Un)Reliability of Steering Vectors in Language Models" by Joschka Braun, Carsten Eickhoff, David Krueger, Seyed Ali Bahrainian, Dmitrii Krasheninnikov.

[Mechanistic Interpretability] global Jane Street challenge to steer model's behaviour in "model.pt" with controlled behavioural probing

Representation Engineering Lab — observation, instrumentation, findings, intervention, application on frozen LLM weights. Home of SFP, RepSNI, rl-drift, SDE, Joi.

Qwen3-8B に対して、日本語・ルー語・英語の同一意味の短文 triplet から hidden state の差分方向を作り、ルー語方向が英語方向とどの程度重なるかを観察する小規模実験です。

Qwen3-0.6B activation steering: style vectors, lens contamination eval, CPRR methodology

Mechanistic interpretability experiments: raw GPT-2 inference, activation steering, and manual backprop MLP.