[ICLR 2025] General-purpose activation steering library

KV Cache Steering for Inducing Reasoning in Small Language Models

[ACL 2026] - Official repo for the paper: "Selective Steering: Norm-Preserving Control Through Discriminative Layer Selection"

Activation steering and trait monitoring for HuggingFace transformers

[ICLR 2026] ASGuard: Activation-Scaling Guard to Mitigate Targeted Jailbreaking Attack

A bilingual awesome list for refusal suppression research: benchmarks, papers, tools, models, and ecosystem updates.

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.

The paper list related to activation steering

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

Accepted at 19th Conference of the European Chapter of the Association for Computational Linguistics, 2026

Token-time interpretability instrument for language models — measures how generation trajectories move, branch, and respond to perturbation. Deterministic intervention comparisons via SeedCache branchpoints, hysteresis protocol, and SAE-feature steering.

Contrastive Activation Steering in GPT-OSS-20B to explore the study of gender, race, religion, and refusal bias in Bilingual data, extending earlier Llama 2 experiments.

Activation steering toolkit for Llama 3.2 3B — inject sensory-constructed vectors into model activations to alter processing dispositions. Web UI + API. Runs locally on consumer hardware.

🏆[ICML 2026 Spotlight] Official implementation of "DecodeShare: Tracing the Shared Subspace of LLM Decode-Time Decisions"

Contrastive Activation Steering for Bias detection in Llama 2 7B — English & Italian.

🌐 Model adversarial economics and AI alignment using the Panopticon Lattice, a multi-agent simulation exploring hidden collusion and system dynamics.

Pharmakon-Eval: dual-use LLM evaluation via Pharmakon Separability Index (PSI) and Dose-Response Curve (DRC)

Code, vectors, and figures for the paper 'Emotion and authorization steering both move cheat; trained-probe suppression doesn't undo it: a mechanistic study in Gemma-2-2B'

Three matched-control gates localize long-context retrieval failure in sparse attention to a single locus- delivery. Attention redirection acts as transport, while stored-state correction is not reusable. Delivery works reliably, is content-addressable (no positional decay), and achieves a 27-byte attribute→code binding. Mechanism: RAG micro-hints.

Anthropic-style emotion-vector geometry, on any open-weight LLM, in one command. Frozen corpus + unified pipeline + statistical rigor + 5-model reference results.