Adversarial Speculative Reasoning Improves Long Horizon Agents

Adversarial Speculative Reasoning Improves Long Horizon Agents. By Hongyin Luo and seven others, from Subconscious Systems Technologies, the Massachusetts Institute of Technology, and two other institutions. Published on the third of November.

Speculative decoding has been treated purely as a systems optimization. A draft model's tokens are verified by the target model, decoding gets faster, and the model's behavior is supposed to stay exactly as it was. Under that view, the growing effort to train better draft models has focused solely on acceptance rate. The authors argue that this view does not hold for long-horizon agent reasoning.

They measured frontier open-weight models on complex agentic tasks under different speculative decoding approaches, and found that numerical noise grows within verification windows. The draft acceptance results during the thinking process can change the model's output distribution after thirty steps. That noticeably influences the reasoning accuracy of long-horizon agents, and in random directions. A component everyone treated as neutral turns out to be steering the outcome.

The response is to control that noise instead of ignoring it, and to use it to improve the reasoning of the target model. The overall token distribution verified by the target model cannot be changed. What can be changed is where the precision goes. By training the draft model to be rejected, critical reasoning can be encoded with the highest precision available. Usually a rejected draft is a small failure, a few wasted guesses. Here rejection becomes the goal at chosen moments, because a rejection is what hands those moments back to the target model.

First, the critical points are identified in agent traces. The draft model is then trained to favor an adversarial preference at those points. At inference, the trained draft and the target part ways precisely at these decision points, and verification restarts. As a result, the critical thinking tokens are sampled from the target model itself and encoded under its least noisy conditions. The authors describe the effect as making critical reasoning better indexed and more durable over a long horizon.

On complex agentic coding benchmarks, the approach raises end-to-end success rate, with no change to the target model or to its output distribution. The training touches between zero point zero six percent and zero point two eight percent of hosted parameters. The abstract does not give the success rates themselves, only that they rise. Two properties of the result stand out from that description. The target model is untouched, so whatever improves is coming from how its tokens are drafted and verified. And the trained portion is a tiny slice of what is hosted, well under one percent of the parameters in every case reported.

The reframing matters as much as the gain. Speculative decoding stops being only an accelerator and becomes, in the authors' words, an active allocator for context precision. A draft model trained to be rejected sounds like a contradiction. It stops sounding like one once the rejection is what routes the most critical tokens onto the most precise path, over a horizon long enough for small noise to matter.