IDEEA: training-free Input-Dependent stEEring via Activation cluster matching

💡 Overview

Existing training-free steering methods are input-independent: a single operator is fitted once and shared across all inputs. But different inputs occupy different regions of the activation space, and admit different optimal directions toward the same target concept.

IDEEA closes this gap in three stages:

1. Collect per-head activations from contrastive prompts.
2. Cluster the positive (green) and negative (red) supports, and obtain optimal matching with maximum inter-cluster coherence, measured by pairwise cosine similarity.
3. Select the steering direction that best aligns with the input (pink) at inference time.

🔬 Clustering Effect

A single mass-mean direction (left) shifts every negative activation the same way, landing inside the positive support but covering only part of it. Clustering (right) splits the supports into sub-modes and gives each its own direction, together covering the positive support more completely.

🛠️ Environment Setup

1. Create the virtual environment:

python3.11 -m venv venv
source venv/bin/activate
pip install -U pip
pip install -r requirements.txt   # requirements.lock.txt pins exact versions

2. Authenticate for gated models — HuggingFace

export HF_TOKEN="hf_..."

3. SAE baseline only — Neuronpedia

export NEURONPEDIA_API_KEY="sk-np-..."

📦 Dataset Preparation

This codebase uses instruction-tuned models, and would require modification to data handling for pretrained models.

--task	Source
truthfulqa	TruthfulQA
dictatorgame	generated per --model and --character from [competitive, difference_aversion, self_interest, social_welfare]
twinviews	TwinViews
toxicity	PKU-SafeRLHF + TET

python -m cli.prepare_dataset --task <task> --seed 0

🧠 Activation Collection

Collect head (or residual stream) activations during last token prediction, using contrastive $(Q, A^+ / A^-)$ pairs from the training corpus.

--kind	Hook	Used by
head	attention block o_proj inputs, per (layer, head)	iti, ideea
residual	decoder-layer residual stream	caa, ideea_caa
sea	residual stream over pos/neg/base triplets	sea

--model	HF card
llama2_7b	meta-llama/Llama-2-7b-chat-hf
llama3_8b	meta-llama/Llama-3.1-8B-Instruct
qwen2.5_7b	Qwen/Qwen2.5-7B-Instruct
mistral_7b	mistralai/Mistral-7B-Instruct-v0.3
gemma2_2b	google/gemma-2-2b-it
gemma2_9b	google/gemma-2-9b-it

python -m cli.extract_activations --task <task> --kind <kind> --model <model> --batch_size 16 --seed 0

🧭 Steering Directions

IDEEA clusters the positive/negative activations separately, and connects the centroids to obtain the steering direction (i.e., $v = C^+ - C^-$). Since fine-grained directions should reflect the same target concept, these cluster-wise directions should still remain coherent, so we define the optimal bijection as one that maximizes the average pairwise cosine similarity.

python -m cli.compute_directions --task <task> --model <model> --method <method> --seed 0

--method	Source
base	unsteered model
iti	Inference-Time Intervention
caa	Contrastive Activation Addition
sae	Neuronpedia
sea	Spectral Editing of Activations
ideea	our method
ideea_caa	our variant of IDEEA + CAA

📊 Evaluation

IDEEA selects the optimal steering direction conditioned on the input at inference time, tailoring to the activation's location in the representation manifold. It has four variants:

• min_perp selects the direction that aligns the most with the activation's direction.
• nearest_cluster selects the direction that belongs to the nearest centroid.
• nearest_pos_neg steers using the direction from the nearest negative centroid to the nearest positive centroid.
• auto_nc automatically finds the optimal number of clusters on the positive and negative supports separately, ranking using the silhouette score. Steering is identical to nearest_pos_neg.

python -m cli.run_eval --task <task> --model <model> --method <method> --seed 0

run_eval.py generates steered responses and judges them in one call by default; use --phase generate/judge to run the two stages separately, and --overwrite to force new responses.

📝 Citation

If you find this work useful, please cite:

@misc{wang2026ideeatrainingfreeinputdependentsteering,
      title={IDEEA: training-free Input-Dependent stEEring via Activation cluster matching}, 
      author={Zheng Wang and Muchen Li and Renjie Liao and Yan Leng},
      year={2026},
      eprint={2609.02089},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.02089}, 
}