Pivotal Token Search

tokens dataset-generation sae sparse-autoencoder dpo llm reasoning-agent llm-inference direct-preference-optimization llm-steering steering-vector phi-4 reasoning-language-models phi4 reasoning-models mech-interp phi4-mini phi-4-mini pivotal-token-search pivotal-tokens

• Updated Nov 7, 2026
• Python

Steering GPT2-EMGSD less biased & Generating stereotyped text with vanilla GPT2 without fine tuning or prompt engineering

sparse-autoencoder bias-correction stereotype gpt2 bias-mitigation emgsd steering-vector

• Updated Mar 17, 2025
• Jupyter Notebook

UCL COMP0087 Statistical Natural Language Processing - NanoGlot Team

nlp multilingual-nlp llm small-llm steering-vector reasoning-language-models

• Updated Apr 17, 2026
• Jupyter Notebook

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

evaluation summarization language-model abstractive-text-summarization abstractive-summarization steering-vector steering-vectors

• Updated Jan 21, 2026
• Python

Exploration of an alternative approach to extracting steering vectors. Instead of using the classical contrastive method we investigate whether comparing activations between a base model and its fine-tuned deceptive version reveals a more meaningful latent direction.

finetuning llms mechanistic-interpretability steering-vector

• Updated Feb 2, 2025
• Jupyter Notebook

Framework for evaluating cross-model memory insertion and negative concept steering in LLMs via anchored latent inversion.

vector steering steering-ai localai steering-vector niodoo ruffian-l

• Updated Sep 7, 2026
• Python

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

machine-learning language-models unreliability steering-vector steering-vectors

• Updated Jun 10, 2025
• Jupyter Notebook

unsupervised search for interpretable behaviours in Qwen 14b

mechanistic-interpretability qwen-14b steering-vector

• Updated Oct 4, 2025
• Python