To Steer or Not to Steer? Mechanistic Error Reduction with Abstention for Language Models

TL;DR

MERA addresses the challenge of steering language models to reduce errors by introducing abstention-based calibration and a principled mechanism to select when and how much to intervene. It combines error estimation via linear probes with a closed-form, layer- and token-aware steering policy, enabling selective activation alterations and a safe abstention option when improvements cannot be guaranteed. The framework provides theoretical guarantees on improvement or abstention and can augment existing steering methods, yielding consistent accuracy gains across diverse models and tasks. Practically, MERA offers a lightweight, post-hoc steering tool with broad applicability to error mitigation and potential extensions to other alignment objectives.

Abstract

We introduce Mechanistic Error Reduction with Abstention (MERA), a principled framework for steering language models (LMs) to mitigate errors through selective, adaptive interventions. Unlike existing methods that rely on fixed, manually tuned steering strengths, often resulting in under or oversteering, MERA addresses these limitations by (i) optimising the intervention direction, and (ii) calibrating when, and how much to steer, thereby provably improving performance or abstaining when no confident correction is possible. Experiments across diverse datasets, and LM families demonstrate safe, effective, non-degrading error correction, and that MERA outperforms existing baselines. Moreover, MERA can be applied on top of existing steering techniques to further enhance their performance, establishing it as a general-purpose, and efficient approach to mechanistic activation steering.

Figures (16)

• Figure 1: An illustration of traditional vanilla steering with fixed steering strengths (left), leading to under-, and oversteering, and MERA (right), providing calibrated steering thresholds.
• Figure 2: Layer-wise performance of linear error estimators for different LM families, separated by distinct token position, and input feature strategies. RMSE, and standard error is reported. Extended results are provided in \ref{['sec:sae-extended-results']}.
• Figure 3: Visualisation of MERA methodology to mechanistically steer LMs: 1.cache activations, and errors, 2.train error estimators, 3.calibrate steering thresholds.
• Figure 4: Overview of steering results for various LMs, and datasets. The first, and second panels display model-, and dataset-specific views, respectively, with $\delta=0.05$. The third panel shows error percentiles, aggregated over the "last"n setting.
• Figure 5: Prompt template for MMLUsubsets.