Research map / Large language models / Reasoning and chain of thought
Semantics-Adaptive Activation Intervention for LLMs via Dynamic Steering Vectors
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 SADI, a training-free activation intervention that masks top-K contrastive-difference elements (heads, neurons, or hidden states) and scales them by the input's own activations at inference, so the steering direction adapts to each input.
Problem
Activation intervention methods use a fixed steering vector for all inputs, which can misalign with the semantics of individual inputs and hurt task performance, especially when the mismatch is large.
Approach
Build a binary mask from top-K mean activation differences between positive and negative contrastive pairs, then at inference scale the masked elements of each input's own activations (A' = A + δ(A ⊙ M)) instead of adding a fixed mean-difference vector, making the intervention input-dependent.
Key results
- On seven multiple-choice tasks with LLaMA2-7b-chat, SADI-Head reaches 70.08 average accuracy versus 64.85 for the baseline and 68.42 for CAA.
- On StoryCloze, SADI-Head improves over the baseline by +14.69 points (79.75 vs 65.06).
- On TruthfulQA generation, SADI-Head raises True×Info from 66.50 to 76.58; SADI-Neuron lowers ToxiGen toxicity from 49.71 to 17.14.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SADI-Head instead of CAA | LLaMA2-7b-chat · StoryCloze | accuracy: 79.75 vs 74.65 | Helped |
| use SADI-Head instead of the baseline | LLaMA2-7b-chat · TruthfulQA generation | True×Info: 76.58 vs 66.50 | Helped |
| use SADI-Neuron instead of the baseline | LLaMA2-7b-chat · ToxiGen | toxicity: 17.14 vs 49.71 | Helped |
| use SADI-Head instead of the baseline | LLaMA2-7b-chat · COPA | accuracy: 78.80 vs 70.80 | Helped |
| use SADI-Hidden instead of the baseline | LLaMA2-7b-chat · TruthfulQA generation | True×Info: 25.38 vs 66.50 | Hurt |
| select top elements by activation difference rather than random elements (ablation) | SADI-Head · COPA | accuracy: 72.40 -> 78.80 | Helped |
| select top elements by activation difference rather than random elements (ablation) | SADI-Head · StoryCloze | accuracy: 71.87 -> 79.75 | Helped |
| select top elements by activation difference rather than random elements (ablation) | SADI-Head · Winogrande | accuracy: 52.88 -> 53.04 | Helped |
| select top elements by activation difference rather than random elements (ablation) | SADI-Head · SST2 | accuracy: 91.48 -> 92.20 | Helped |
| select top elements by activation difference rather than random elements (ablation) | SADI-Head · TriviaQA | EM: 42.90 -> 44.00 | Helped |
| use input-adaptive steering rather than a fixed steering vector (ablation) | SADI-Head · COPA | accuracy: 71.20 -> 78.80 | Helped |
| use input-adaptive steering rather than a fixed steering vector (ablation) | SADI-Head · StoryCloze | accuracy: 69.42 -> 79.75 | Helped |
Surprising findings
- SADI-Hidden performs worse than the baseline on StoryCloze generation-style evaluation and TruthfulQA generation, suggesting hidden-state intervention is sensitive to task format.
- Replacing the input-dependent scaling with a fixed mean-difference steering vector, or randomly choosing K elements, lowers accuracy, but the gap to fixed steering is modest on some tasks (e.g., Winogrande 53.04 vs 52.56).
Limitations
- Hyperparameters K and δ must be tuned per task on validation data, with optimal settings varying widely across tasks.
- Gains from head intervention are smaller in few-shot settings.
- Benefits vary across models and configurations; SADI-Hidden sometimes degrades performance.
Experimental setup
Datasets: COPA StoryCloze NLI MMLU SST2 SST5 BoolQ Winogrande TriviaQA TruthfulQA ToxiGen RealToxicityPrompts
Benchmarks: COPA StoryCloze NLI MMLU SST2 SST5 BoolQ Winogrande TriviaQA TruthfulQA ToxiGen XCOPA
Metrics: accuracy exact match multiple-choice accuracy (MC1/MC2/MC3) truthful and informative judge scores toxicity (HateBERT classifier)
Baselines: ITI CAA SFT baseline (no intervention)
Builds on: Inference-Time Intervention Contrastive Activation Addition linear representation hypothesis
Closest papers by meaning
- Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.90)
- Focus On This, Not That! Steering LLMs with Adaptive Feature Specification (ICML 2025 · similarity 0.89)
- Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · similarity 0.88)
- Improving Instruction-Following in Language Models through Activation Steering (ICLR 2025 · similarity 0.88)
- Activation Control for Efficiently Eliciting Long Chain-of-thought Ability of Language Models (NeurIPS 2025 · similarity 0.87)
- LoFiT: Localized Fine-tuning on LLM Representations (NeurIPS 2024 · similarity 0.87)
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.