Research map / Large language models / Language, safety and interpretability
AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders
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.
Introduces AxBench, a synthetic-data benchmark comparing representation-based steering and concept detection methods (SAEs, DiffMean, probes, LAT, PCA, ReFT-r1) against prompting and finetuning on Gemma-2 models, and proposes ReFT-r1.
Problem
There is no common benchmark to directly compare representation-based steering and concept detection methods (such as SAEs) with prompting and finetuning at realistic, open-vocabulary scale.
Approach
Generate synthetic labelled training and evaluation data per concept with an LLM (gpt-4o-mini), then evaluate steering with an LLM judge on long-form Alpaca-Eval generations and concept detection with AUROC/F1; additionally propose ReFT-r1, a rank-1 supervised dictionary that jointly learns detection and steering.
Key results
- For steering, prompting achieves the highest mean overall score (0.894 avg), above all representation methods; ReFT-r1 reaches 0.543 vs SAE 0.165 on average.
- For concept detection, DiffMean (0.942), Probe (0.940) and ReFT-r1 (0.938) have the highest average AUROC, while SAE averages 0.695.
- ReFT-r1 achieves 88.0% winrate against SAEs on steering, with DiffMean at 61.6%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ReFT-r1 instead of SAE for concept detection | Gemma-2 models · Concept500 concept detection | AUROC: 0.695 -> 0.938 | Helped |
| Use DiffMean instead of SAE for concept detection | Gemma-2 models · Concept500 concept detection | AUROC: 0.695 -> 0.942 | Helped |
| Use Prompt instead of representation steering with SAE | Gemma-2 models · Concept500 model steering | overall score: 0.165 -> 0.894 | Helped |
| Use ReFT-r1 instead of SAE for model steering | Gemma-2 models · Concept500 model steering | overall score: 0.165 -> 0.543 | Helped |
| Use DiffMean instead of SAE for model steering | Gemma-2 models · Concept500 model steering | overall score: 0.165 -> 0.239 | Helped |
| Replace SAE decoder-feature addition with clamping (ablation) | SAE steering · Concept500 model steering | overall score: 0.165 -> 0.088 | Hurt |
| Replace SAE decoder-feature addition with minimum clamping (ablation) | SAE steering · Concept500 model steering | overall score: 0.165 -> 0.090 | Hurt |
| Calculate SAE maximum activation from the evaluation dataset instead of Neuronpedia (ablation) | SAE steering · Concept500 model steering | overall score: 0.165 -> 0.152 | No clear effect |
| Increase the steering factor | All steering methods · Concept500 model steering | instruction score | Hurt |
| Increase the number of ReFT-r1 training examples (ablation) | ReFT-r1 · Concept10 concept detection and model steering | performance | Helped |
Surprising findings
- Simple prompting outperforms all representation-based steering methods, including SAEs and supervised steering vectors, on steering.
- SAE-A (AUROC-selected features) slightly underperforms the unsupervised SAE on steering, suggesting better classification does not yield better steering.
Limitations
- SFT results only available for the first 20 concepts due to limited resources.
- Concept lists derived from Neuronpedia auto-interpretability labels, which are skewed toward token-level concepts.
- Relies on a single synthetic data generator (gpt-4o-mini) and LLM judge.
Experimental setup
Datasets: Concept500 Concept16K Alpaca-Eval GemmaScope SAE concept lists (Neuronpedia)
Benchmarks: AxBench Alpaca-Eval
Metrics: AUROC F1 harmonic mean of concept, instruct, fluency scores (0-2) winrate
Baselines: Prompting SFT LoRA LoReFT SAE DiffMean Probe LAT PCA SSV
Builds on: ReFT difference-in-means steering GemmaScope pyvene linear artificial tomography
Closest papers by meaning
- Improved Representation Steering for Language Models (NeurIPS 2025 · similarity 0.86)
- Beyond Interpretability: When, Why, and How Sparse Autoencoders Enable Label-Free Visual Steering (NeurIPS 2026 · similarity 0.86)
- SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretability (ICML 2025 · similarity 0.85)
- MIB: A Mechanistic Interpretability Benchmark (ICML 2025 · similarity 0.85)
- Are Sparse Autoencoders Useful? A Case Study in Sparse Probing (ICML 2025 · similarity 0.85)
- Analysing the Generalisation and Reliability of Steering Vectors (NeurIPS 2024 · 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.