Language models can resist distractors better without additional training

Across five reasoning models and five benchmarks, targeted weight edits improved accuracy on perturbed problems by an average of 11.8 percentage points.

Chinese Tech

Zhizhang Fu · Lucas Zhao · Michalis Korakakis · Mengru Ding · Lukasz Szpruch · Yue Zhang

Westlake University · Yale University · University of Cambridge · The Alan Turing Institute · Xianyu of Alibaba

Research Digest·10 November 2026·2 min read

Thread:Activation Steering Vectors

Illustration: Language models can resist distractors better without additional training

In one line

Weight editing suppresses spurious reasoning in LLMs, improving accuracy on perturbed inputs by 11.8 percentage points.

Before you read this

Three ideas this paper builds on

mechanistic interpretabilityThe study of how neural networks compute through internal circuits, features, and causal interventions, rather than just input-output behavior.129 papers →counterfactual reasoningComparing actual outcomes with imagined alternatives to estimate the effects of different actions or conditions.104 papers →activation steeringActivation steering modifies a model's internal vector representations during inference to shift behavior, such as editing outputs, removing capabilities, or controlling trust.72 papers →

Read it as

Large language models often rely on misleading shortcuts in the input, such as irrelevant details, instead of true reasoning. This makes them brittle when small changes are made to the question. Fixing this is hard because errors in long reasoning chains can hide the source of the problem.

The researchers developed a method called SWEET that works without extra training. It compares the model's internal activity on the original problem to its activity on a version with a distracting change. This identifies which neurons and attention heads are processing the distraction, then edits their weights to reduce their influence.

SWEET improved accuracy by an average of 11.8 percentage points on tricky inputs and 2 points on normal inputs across five models. It also shrank the performance gap between clean and perturbed problems from 17.6% to 2.5%, making reasoning far more robust to distractions.

How it works

stage by stage

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

mechanistic interpretability counterfactual reasoning weight editing chain of thought activation steering