Research map / Large language models / Language, safety and interpretability
To Steer or Not to Steer? Mechanistic Error Reduction with Abstention for Language Models
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 MERA, a closed-form, calibrated activation-steering method that uses linear error probes to steer LMs toward correct answers only when a calibration-validated threshold confidently improves accuracy, and abstains otherwise.
Problem
Fixed, manually tuned steering strengths in activation steering cause under- or oversteering, and existing steering methods are not designed to reduce continuous prediction errors on supervised LM tasks.
Approach
Train linear error estimators on residual-stream activations, steer along the probe weights with a closed-form strength that moves activations just to a predicted-error threshold alpha, and calibrate alpha on a held-out set with a Hoeffding/Bonferroni bound so steering is applied only when it provably improves accuracy, with abstention otherwise.
Key results
- MERA turns contrastive steering SPI from -0.05 to +0.52 on Yes/No and from -0.09 to +0.21 on MMLU-hs.
- MERA-steered probe reaches +0.64 SPI on Sentiment and +0.53 on Yes/No, making it the best-performing method in those settings.
- Probe-weight steering without calibration is too weak (+0.00 SPI on MMLU-hs and SMS Spam).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MERA's calibrated, adaptive error-based steering rather than no steering | language models · evaluated tasks and models | SPI | Helped |
| Replace uncalibrated contrastive steering with MERA calibration | contrastive steering · Yes/No | SPI: -0.05 -> +0.52 | Helped |
| Replace uncalibrated contrastive steering with MERA calibration | contrastive steering · MMLU-hs | SPI: -0.09 -> +0.21 | Helped |
| Use MERA calibration with an error probe rather than direct probe-based steering | error-probe steering · Sentiment | SPI: +0.64 | Helped |
| Use MERA calibration with an error probe rather than direct probe-based steering | error-probe steering · Yes/No | SPI: +0.53 | Helped |
| Use logistic probe steering without MERA calibration | logistic probe steering · Sentiment | SPI: -0.85; +0.83 | Mixed |
| Extract activations at the exact answer token position instead of the last input position (ablation) | linear error probes · most model families, including LLaMA and Gemma | RMSE | Helped |
| Use sparse SAE activations instead of original activations (ablation) | linear error probes · evaluated models and tasks | RMSE | No clear effect |
| Replace the error-probe direction with contrastive means (k=100) (ablation) | MERA · Yes/No | SPI: +0.52 | Helped |
| Replace the error-probe direction with contrastive means (k=100) (ablation) | MERA · SMS Spam | SPI: +0.77 | Helped |
| Replace the error estimator with a logistic regression probe (ablation) | MERA · Yes/No | SPI: +0.00 | No clear effect |
| Replace the error estimator with a logistic regression probe (ablation) | MERA · Sentiment | SPI: +0.00 | No clear effect |
Surprising findings
- Unsteered logistic-probe and contrastive steering often degrade accuracy substantially (e.g., -0.85 SPI on Sentiment), showing that uncalibrated steering is unreliable.
- MMLU subsets are overall difficult to steer, with gains much smaller than on binary tasks.
Limitations
- The framework is linear; non-linear relationships between activations and error are not captured.
- Evaluated only on MCQA/classification tasks; effects on generation or general capabilities are unknown.
Experimental setup
Datasets: SMS Spam Yes/No Sentiment MMLU-hs MMLU-prof
Benchmarks: SMS Spam Yes/No Sentiment MMLU-hs MMLU-prof
Metrics: steering performance impact (SPI) accuracy RMSE
Baselines: prompt-based steering (Think before you answer) contrastive steering logistic probe steering probe-weight steering
Builds on: contrastive activation addition probe-based steering sparse autoencoders
Closest papers by meaning
- Steering When Necessary: Flexible Steering Large Language Models with Backtracking (NeurIPS 2025 · similarity 0.90)
- Analysing the Generalisation and Reliability of Steering Vectors (NeurIPS 2024 · similarity 0.89)
- Correctness-Optimized Residual Activation Lens (CORAL): Transferrable and Calibration-Aware Inference-Time Steering (ICML 2026 · similarity 0.88)
- Improved Representation Steering for Language Models (NeurIPS 2025 · similarity 0.88)
- Improving Instruction-Following in Language Models through Activation Steering (ICLR 2025 · similarity 0.88)
- Focus On This, Not That! Steering LLMs with Adaptive Feature Specification (ICML 2025 · 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.