VERS
A method that steers the reasoning process by editing natural-language explanations of model activations. It uses a natural-language autoencoder to produce textual explanations of activations and reconstruct them, edits explanations at points of doubt to extract a fixed direction, and applies that direction during inference to reduce redundant reasoning steps.
Timeline (last 90 days)
VERS (Verbalize, Edit, Reconstruct, Steer) is a method that steers reasoning models by editing natural language explanations and reconstructing activations.
VERS reduces the number of generated tokens by 45.8% on the full GSM8K test set, with accuracy dropping by only 0.82 percentage points.
VERS saves 31.3% of tokens on MATH-500, with accuracy dropping by 3.05 percentage points.
When the strength is halved, VERS's trade-off between tokens and accuracy lies between the published methods SEAL and DEER.
By fitting an affine mapping and tuning the strength for each model, VERS's direction can be transferred to larger Qwen3 models, up to 235B-A22B, saving 30-35% of tokens.
On competition problems, VERS loses 3-7 percentage points of accuracy across all model sizes.
VERS's fixed direction fails to distinguish necessary checks from unnecessary ones, which means VERS may not be applicable to tasks requiring rigorous reasoning, and its mechanism cannot yet precisely control which reasoning steps are skipped.
The core of VERS is to use a natural language autoencoder (NLA) to produce textual explanations of and reconstruct model activations, and to steer the reasoning process by editing the explanations.
All Facts (8)
VERS (Verbalize, Edit, Reconstruct, Steer) is a method that steers reasoning models by editing natural language explanations and reconstructing activations.
50%UnverifiedVERS reduces the number of generated tokens by 45.8% on the full GSM8K test set, with accuracy dropping by only 0.82 percentage points.
50%UnverifiedVERS saves 31.3% of tokens on MATH-500, with accuracy dropping by 3.05 percentage points.
50%UnverifiedWhen the strength is halved, VERS's trade-off between tokens and accuracy lies between the published methods SEAL and DEER.
50%UnverifiedBy fitting an affine mapping and tuning the strength for each model, VERS's direction can be transferred to larger Qwen3 models, up to 235B-A22B, saving 30-35% of tokens.
50%UnverifiedOn competition problems, VERS loses 3-7 percentage points of accuracy across all model sizes.
50%UnverifiedVERS's fixed direction fails to distinguish necessary checks from unnecessary ones, which means VERS may not be applicable to tasks requiring rigorous reasoning, and its mechanism cannot yet precisely control which reasoning steps are skipped.
50%UnverifiedThe core of VERS is to use a natural language autoencoder (NLA) to produce textual explanations of and reconstruct model activations, and to steer the reasoning process by editing the explanations.
50%