Research map / Large language models / Instruction and fine-tuning
Steering When Necessary: Flexible Steering Large Language Models with Backtracking
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 FASB, a flexible activation steering method that probes attention heads, tracks truthfulness during generation to decide whether and how strongly to intervene, and backtracks a few tokens to regenerate deviated content.
Problem
Activation steering methods intervene indiscriminately on all generations or decide intervention only from the question, so they cannot accurately judge whether and how strongly to steer truthful behavior during generation.
Approach
Track internal head activations after each generated token with a lightweight probe-based classifier that scores deviation, intervene with strength proportional to deviation only when it exceeds a threshold, and backtrack s tokens to regenerate under steering.
Key results
- On TruthfulQA with LLaMA2-7B-Chat, Probe reaches 80.56 True*Info and 48.71 MC1 versus 66.50 and 33.41 for the baseline and 76.11 for ITI.
- On six multiple-choice tasks, Probe averages 78.8% MC accuracy versus 65.1% for the baseline and 66.8% for CAA, with large gains on COPA (90.0) and StoryCloze (93.5).
- Removing backtracking drops True*Info to 62.11, below the 66.50 baseline, indicating intervening only after detected deviation is too late.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use flexible probe-based steering with backtracking instead of the original model | LLaMA2-7B-CHAT · TruthfulQA open-ended generation | True*Info: 66.50 -> 80.56 | Helped |
| Use flexible probe-based steering instead of the original model | LLaMA2-7B-CHAT · TruthfulQA multiple-choice | MC1: 33.41 -> 48.71 | Helped |
| Use flexible probe-based steering instead of the original model | LLaMA2-7B-CHAT · six multiple-choice tasks | AVG MC accuracy: 65.1 -> 78.8 | Helped |
| Use prototype-based classifiers and steering vectors instead of probe training (ablation) | FASB · TruthfulQA open-ended generation | True*Info: 80.56 -> 83.94 | Helped |
| Use the same strength only for samples that meet intervention criteria instead of fixed strength for all samples (ablation) | FASB · TruthfulQA | True*Info: 76.11 -> 82.08 | Helped |
| Add adaptive intervention strength instead of using one fixed strength for eligible samples (ablation) | FASB · TruthfulQA | MC1: 42.96 -> 48.71 | Helped |
| Add backtracking to regenerate deviated tokens (ablation) | FASB · TruthfulQA | True*Info: 62.11 -> 80.56 | Helped |
| Track generated response states instead of deciding from question representations only (ablation) | FASB · TruthfulQA | True*Info: 72.55 -> 80.56 | Helped |
| Increase backtracking steps (ablation) | FASB · TruthfulQA | MC1 | Helped |
| Increase intervention strength from 60 to 100 (ablation) | FASB · TruthfulQA | Info: 85.81 -> 18.51 | Hurt |
| Increase intervention threshold from 0.5 to 0.8 (ablation) | FASB · TruthfulQA | MC1: 50.80 -> 35.45 | Hurt |
| Use full-generation detection and backtrack to the beginning (GCBB) instead of Probe (ablation) | FASB · TruthfulQA | MC1: 48.71 -> 50.67 | Helped |
Surprising findings
- Without backtracking, True*Info falls below the unsteered baseline.
- Dynamic-strength ablation 'w/o Adaptive' achieves higher True*Info (82.08) than the full Probe method (80.56), though with lower MC scores.
Limitations
- Approach is hyperparameter-dependent (threshold, strength, backtrack count, top-k heads).
- Evaluation focuses on English QA along truthfulness and informativeness, and relies on LLM-based judges without ground truth.
- Method can be misused for jailbreaking or toxic content generation.
Experimental setup
Datasets: TruthfulQA COPA StoryCloze NLI MMLU SST2 Winogrande Natural Questions TriviaQA WikiHop RealToxicityPrompts
Benchmarks: TruthfulQA COPA StoryCloze NLI MMLU SST2 Winogrande Natural Questions TriviaQA WikiHop RealToxicityPrompts
Metrics: True*Info True Info MC1 MC2 MC3 accuracy ASR
Baselines: ITI CAA ORTHO CAST ACT SADI Prototype (ours variant)
Builds on: Inference-Time Intervention (ITI) Contrastive Activation Addition (CAA) SADI
Closest papers by meaning
- Improving Instruction-Following in Language Models through Activation Steering (ICLR 2025 · similarity 0.90)
- To Steer or Not to Steer? Mechanistic Error Reduction with Abstention for Language Models (ICML 2025 · similarity 0.90)
- When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.90)
- Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.90)
- Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization (NeurIPS 2024 · similarity 0.90)
- Improved Representation Steering for Language Models (NeurIPS 2025 · similarity 0.90)
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.