Today we're releasing Adversarial Speculative Reasoning (ASpec), a new technique for making long-horizon agents more accurate.
Before ASpec, speculative decoding draft models have been used purely as a mechanism to improve the throughput of a language model. We demonstrated that by retraining just 0.06% to 0.28%of the draft model's parameters, we can significantly improve the system's accuracy on coding tasks.
We make the model adversarial by training it to guess wrong at the points where an agent makes a key decision like to call a tool or to run a test. When the draft model suggests a rejected token, the main model steps in to write that token itself. Then, the following key decision tokens gets generated with the least noise.
Wrong decisions are critical on long-horizon agentic work because a small mistake early on may compound into much larger problems. With ASpec, we're able to more accurately make correct decisions while also significantly increasing the throughput of the system.
ASpec is live in the Subconscious API today with custom draft models supporting inference for GLM 5.3 and DeepSeek V4.1 Flash 1017.
Read the full ASpec paper on arXiv: Adversarial Speculative Reasoning Improves Long Horizon Agents.
How speculative decoding works
If you're not familiar with speculative decoding, this video walks through the basics and how our approach with adversarial speculative decoding improves both speed and accuracy.
Results
We ran both GLM 5.3 and DeepSeek V4.1 Flash on DeepSWE, a benchmark of complex software engineering tasks. With ASpec, the models successfully solved more tasks, and did so more consistently across a number of sample runs.
| Model | Success | Success with ASpec |
|---|---|---|
| GLM 5.3 (NVFP4) | 60.6% | 67.6% |
| DeepSeek V4.1 Flash | 67.8% | 71.8% |
Live on Subconscious today
ASpec is live on on the Subconscious API and powering the GLM 5.3 and DeepSeek V4.1 Flash 1017 models. Combined with our enhancements in runtime context compression and suffix caching, ASpec provides another gain in accuracy and performance for long-running agents.
Immediately, our users will feel a boost in ability with no change to the throughput of the system. Subconscious is an excellent choice for coding agents, cybersecurity agents, and any task consuming 200k+ tokens.
