Editing explanation texts steers reasoning models efficiency and accuracy
VERS: Steering Reasoning Models by Editing Their Activation Explanations
Computation and LanguageMachine Learning
Summary
Reasoning language models write explanations called chains of thought before answering. Many parts of these explanations don't affect the final answer. The authors studied if editing these explanation texts can guide how the model reasons. They created VERS, a technique that edits explanation texts at uncertain points to steer the model and reduce the amount of generated text while keeping accuracy mostly stable. VERS works well on math problem benchmarks but performs worse on competition problems and struggles to identify which explanation parts are necessary.
What this means in practice
- •For applied ai engineers: Cut down reasoning-related text generation in language models by nearly half while keeping accuracy almost unchanged on standard math benchmarks.
- •For machine learning infrastructure teams: Reduce compute and token usage for large reasoning models by applying learned directions to steer reasoning efficiency across model sizes.
Authors
Ihor Vitenko
Abstract
Reasoning language models write a chain of thought before they answer, and many checks in it do not change the final answer. Natural language autoencoders (NLAs) explain a model's activations in text and rebuild activations from text, and editing an explanation can steer a model at a single token. We study whether such edits can also steer how a model reasons. We build VERS (Verbalize, Edit, Reconstruct, Steer): it turns explanation edits at the doubt points of one reasoning trace into one fixed direction, and adds this direction at paragraph breaks. On the full GSM8K test, VERS cuts generated tokens by 45.8% at an accuracy change of -0.82 percentage points (pp); on MATH-500 it saves 31.3% at -3.05 pp. At half strength, its trade-off between tokens and accuracy lies between those of the published methods SEAL and DEER. With fitted affine maps and a strength tuned for each model, the direction transfers to larger Qwen3 models, up to 235B-A22B, and saves 30-35% of tokens. On competition problems, however, VERS loses 3-7 pp at every model size, and a fixed direction did not separate necessary checks from unneeded ones.