Research map / Large language models / Language, safety and interpretability
REAL: Reading Out Transformer Activations for Precise Localization in Language Model Steering
Li-Ming Zhan, Bo LIU, Yujie Feng, Chengqiang Xie, Jiannong Cao, Xiao-Ming Wu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| select heads with REAL instead of ITI's probe-based selection | ITI · TruthfulQA (Qwen2.5-7B-Instruct) | MC1: 40.51 -> 64.21 MC2; 24.48 -> 44.43 MC1 | Improved |
| select heads with REAL instead of LoFiT's selection | LoFiT · TruthfulQA (Llama2-7B-Chat) | MC1: 59.56 -> 60.90 | Improved |
| add REAL head selection to ITI | ITI · LLM-as-judge TruthfulQA (Llama2-7B-Chat) | Truth x Info: 3.40 -> 21.39 | Improved |
| add REAL head selection to ITI | ITI · LLM-as-judge TruthfulQA (Qwen2.5-7B-Instruct) | Truth x Info: 14.82 -> 32.11 | Improved |
| add REAL layer selection to SPARE | SPARE · NQSWAP (Llama3-8B) | Contextual EM: 77.69 -> 80.17 | Improved |
| add REAL layer selection to SPARE | SPARE · MACNOISE (Llama3-8B) | Parametric EM: 30.72 -> 32.57 | Improved |
| use 50% of training data instead of 100% for REAL (ablation) | REAL · TruthfulQA MC1 (Qwen2.5-7B) | MC1: 45.59 -> 40.27 | Mixed |
| use 50% of training data instead of 100% for ITI (ablation) | ITI · TruthfulQA MC1 (Qwen2.5-7B) | MC1: 30.72 -> 27.13 | Hurt |
| use adaptive importance-weighted strength instead of ITI constant strength at eps=20 (ablation) | REAL · TruthfulQA MC1 (Llama2-7B-Chat) | MC1: 26.09 -> 42.84 | Improved |
| steer all 1024 heads with REAL instead of ITI (ablation) | REAL · TruthfulQA MC1 (Llama2-7B-Chat) | MC1: 17.82 -> 31.38 | Improved |
| use adaptive strength instead of fixed strength with 100 steered heads (ablation) | REAL · TruthfulQA MC1 (Llama2-7B-Chat) | MC1: 29.07 -> 35.62 | Improved |
| set codebook size to 32 instead of 2 (ablation) | REAL · TruthfulQA MC1 (Qwen2.5-7B) | MC1: 39.78 -> 45.59 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- SPARE — SPARE: Single-Pass Annotation with Reference-Guided Evaluation for Automatic Process Supervision and Reward Modelling (AAAI 2026) · better on 4 of 4 shared results
- LoFiT — LoFiT: Localized Fine-tuning on LLM Representations (NeurIPS 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.90)
- Model Whisper: Steering Vectors Unlock Large Language Models’ Potential in Test-Time (AAAI 2026 · similarity 0.88)
- The Sharp Directions Are Against You: Curvature Analysis of Activation Steering (NeurIPS 2026 · similarity 0.87)
- Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · similarity 0.87)
- Focus On This, Not That! Steering LLMs with Adaptive Feature Specification (ICML 2025 · similarity 0.87)
- Correctness-Optimized Residual Activation Lens (CORAL): Transferrable and Calibration-Aware Inference-Time Steering (ICML 2026 · 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.
