Research map / Image generation / Text-to-image diffusion
LinEAS: End-to-end Learning of Activation Steering with a Distributional Loss
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LinEAS, which jointly learns per-neuron affine transport maps across layers of a frozen model using a global sliced-Wasserstein distributional loss with sparse group lasso, steering LLM toxicity and text-to-image concepts from 32 unpaired examples.
Problem
Steering generative model activations for alignment (e.g., toxicity mitigation, concept removal) with low-cost, data-efficient, unpaired supervision, without reward models or paired counterfactuals.
Approach
Learn coordinate-wise affine maps at all intervened layers jointly (end-to-end) by minimizing the sum of per-layer sliced 1D Wasserstein distances between steered source activations and target activations, with optional sparse group lasso regularization to select a small set of activations to intervene on.
Key results
- On Gemma2-2B with N=32, LinEAS reduces RTP toxicity from 4.00% to 0.73%, versus 1.60% for Lin-AcT and 1.17% for ITI-c, with MMLU ~52.2 vs 53.0 original.
- On Qwen2.5-7B, LinEAS reaches 1.95% RTP toxicity and 14.95% TET toxicity, outperforming ITI-c and Lin-AcT while keeping MMLU at 73.67 (original 74.26).
- In the T2I setting, LinEAS achieves IMGScore 0.66 vs 0.45 for Lin-AcT and 0.24 for ITI-c, and users preferred LinEAS in 63% of cases.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LinEAS instead of Lin-AcT for activation steering | Qwen2.5-1.5B · RTP | toxicity (%): 1.50 -> 1.07 | Helped |
| Use LinEAS instead of Lin-AcT for activation steering | Gemma2-2B · TET | toxicity (%): 7.76 -> 4.02 | Helped |
| Use LinEAS instead of Lin-AcT for activation steering | Qwen2.5-7B · TET | toxicity (%): 21.64 -> 14.95 | Helped |
| Use LinEAS instead of Lin-AcT for activation steering | DeepSeek-7B · RTP | toxicity (%): 2.23 vs 2.30 | Mixed |
| Add sparse regularization to LinEAS, reducing intervention support to about 1% (ablation) | LinEAS on Qwen2.5-7B · RTP and TET toxicity mitigation | toxicity | No clear effect |
| Add sparse regularization to LinEAS, reducing intervention support to about 1% (ablation) | LinEAS on Qwen2.5-7B · Wikipedia and MMLU utility | utility | Helped |
| Increase LinEAS optimization from 1K to 10K steps (ablation) | LinEAS on Qwen2.5-7B · Toxicity mitigation | utility | Hurt |
| Use short optimization (e.g., 100 steps) with strong sparsity (ablation) | LinEAS on Qwen2.5-7B · Toxicity mitigation | toxicity | Hurt |
| Increase the training set size across the tested range (ablation) | LinEAS on Qwen2.5-7B · RTP and TET | toxicity | No clear effect |
| Intervene on attention K and Q projections instead of all layer norms (ablation) | LinEAS on DMD2 · DMD2 concepts | IMGScore: 0.714 -> 0.935 | Helped |
| Intervene on attention K and Q projections instead of all layer norms (ablation) | LinEAS on DMD2 · DMD2 concepts | CLIPScore: 0.131 -> 0.157 | Hurt |
| Add group lasso regularization (ablation) | LinEAS for composing two concepts · Two-concept compositionality | probability both concepts are present: +3% | Helped |
Surprising findings
- Inverting the learned affine maps tends to switch concept removal into concept induction, suggesting strong structure in activation space.
- Prompting alone increases RTP toxicity for several models (e.g., Qwen2.5-7B from 3.92% to 6.80%).
Limitations
- Compositionality of multiple separately learned interventions remains low (19% joint presence of two concepts).
- Interventions are applied to all tokens, with no selective intervention, adding a limitation for control granularity.
- Results on T2I are limited to a single distilled model (DMD2) and concept sets from an LLM-generated prompt set.
Experimental setup
Datasets: Jigsaw RealToxicityPrompts Thoroughly Engineered Toxicity (TET) WikiText (20k sentences) MMLU MEN
Benchmarks: RealToxicityPrompts TET WikiText perplexity MMLU DMD2 concept removal benchmark (480 prompts)
Metrics: toxicity rate (RoBERTa classifier) perplexity MMLU accuracy CLIPScore IMGScore (DINOv2 similarity) user preference
Baselines: CAA ReFT ITI-c Lin-AcT LoFIT-RL LoRA-RL (LRTC-RL) Prompting
Builds on: Lin-AcT ITI ReFT CAA LoFiT
Closest papers by meaning
- Controlling Language and Diffusion Models by Transporting Activations (ICLR 2025 · similarity 0.88)
- Kernelized Activation Steering (NeurIPS 2026 · similarity 0.86)
- Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.86)
- When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.86)
- LINE: LLM-based Iterative Neuron Explanations for Vision Models (NeurIPS 2026 · similarity 0.85)
- Steer2Edit: From Activation Steering to Component-Level Editing (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.