[ICLR 2025] General-purpose activation steering library
-
Updated
Nov 4, 2026 - Python
[ICLR 2025] General-purpose activation steering library
We study whether categorical refusal tokens enable controllable and interpretable safety behavior in language models.
Reproducible, evergreen benchmark for LLM refusal on biological research prompts — 19 models, 141 prompts, 13,389 adjudicated trials
Locating and editing refusal in the J-space workspace with the Jacobian lens: refusal is legible ~10 layers before the first token, and only ~1/3 lives in the verbalizable workspace.
Public Driftmap harness: public-safe CSV suites + rubrics + run logs for drift detection, refusal integrity, injection resistance, and uncertainty tracking.
A probe suite that measures which conversation states an LLM cannot leave. Three arms, because two cannot tell obedience from token statistics; a null only counts when the design had the power to see the effect.
中文企业公开报告 Hybrid RAG:Docling 解析 · Qdrant 稠密/稀疏检索 · 查询理解硬过滤 · 带引用生成与拒答 · 文档生命周期与评测看板
RAG with verifiable citations and measured refusal — retrieval scored separately (TF-IDF beats embeddings here), citations validated against chunks actually retrieved.
Does a model's safety depend on the script Urdu is typed in? Dataset, steering vectors and a pre-registered study of Tiny Aya Fire on Urdu script vs Roman Urdu (Lexsi AI Safety & Alignment Hackathon).
Every RAG eval reports faithfulness. None reports over-refusal beside it. An eval harness and policy engine for refusal correctness in regulated RAG: synthetic KYC/AML corpus, deterministic verifier, and a refusal policy you can commit, hash and audit.
RefusalScope
The Identity-Refusal Effect: LLMs Systematically Refuse First-Person Moral Self-Report, Distorting Moral Foundation Measurement
An experiment in mechanistic interpretability exploring how refusal behaviour is represented inside language models. Using Qwen2.5-1.5B-Instruct, I reproduce the single-direction refusal intervention described by Arditi et al. (2024), investigate where the refusal signal emerges across layers, and test what happens when it is removed.
Is hedging in LLMs mediated by a single direction? Code, data and results for our 2AMM20 paper (TU/e, 2026).
An open reproduction of feature-level activation steering with the prompt set released, showing the capability tax that behavioural metrics miss
Training-time defense that redistributes LLM refusal via mean/covariance matching + KD, raising linear-ablation attack rank from K=1 to K≥16 (Llama-3.2-1B-Instruct)
Public reference interfaces for proof-gated AI action, refusal, authority, and evidence boundaries.
A document copilot that cites what it says and refuses when the evidence is not there: 33 questions, baseline 12/33 to harness 32/33, control 8/8 both ways.
How does this function's cost scale? Counted, not timed — and UNDETERMINED when no complexity class settles.
Activation-level interventions that come with their controls — steer, ablate, patch, and SAE feature analysis for open-weight LLMs.
To associate your repository with the refusal topic, visit your repo's landing page and select "manage topics."