Illustration, AI-generated from the story text. Tap a panel to read the story through it.

1 / 3Training a draft model to be rejected lifts agent scores by up to seven points, a new paper reports×

The method

Adversarial Speculative Reasoning, from Subconscious Systems with MIT, Princeton and Northeastern authors, trains the small draft model used in speculative decoding to fail at key decision points, so the main model reasons there itself.

The gain

GLM-5.3 rose from 60.6% to 67.6% on DeepSWE, for 1.8% less decoding speed.

The caveat

DeepSeek V4.1 Flash gained four points, with a confidence interval that includes zero.

Sources and the full story

‹›

Training a draft model to be rejected lifts agent scores by up to seven points, a new paper reports

Adversarial Speculative Reasoning, from Subconscious Systems with MIT, Princeton and Northeastern authors, trains the small draft model used in speculative decoding to fail at key decision points, so the main model reasons there itself. GLM-5.3 rose from 60.6% to 67.6% on DeepSWE, for 1.8% less decoding speed. DeepSeek V4.1 Flash gained four points, with a confidence interval that includes zero. SourcesA

Sources

• AAdversarial Speculative Reasoning (Hugging Face Papers)huggingface.co

This story ran in the Afternoon Brief of November 4th 2026.