Skip to content
Neural Daily

ResearchNews

ASpec: adversarial speculation for long agents

🇭🇺 Olvasd el magyarul →

Subconscious released ASpec, which trains a draft model to guess wrong at key agent decisions so the main model corrects them more accurately. It is live for GLM 5.3 and DeepSeek V4.1 Flash.

· 1 min read

Subconscious research graphic for Adversarial Speculative Reasoning
Image: Subconscious · Press kit · Source

The idea

Speculative decoding usually speeds generation. Subconscious ASpec retrains 0.06–0.28% of a draft model’s parameters so it deliberately guesses wrong at key decisions such as tool calls or running tests. When that draft token is rejected, the main model writes the decision with less noise. The goal is fewer early mistakes in long multi-step agent runs where errors compound.

Results and availability

  • DeepSWE (company numbers): GLM 5.3 NVFP4 60.6% → 67.6%; DeepSeek V4.1 Flash 67.8% → 71.8% with ASpec.
  • Live on the Subconscious API with custom draft models for GLM 5.3 and DeepSeek V4.1 Flash 1017.

Caveat

Scores come from Subconscious’s own harness. The gains matter most on long jobs that burn 200k+ tokens.

    By Mikola Máté Patrik

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles