Adversarial speculative decoding improves long-horizon agent reasoning
Original title: Adversarial Speculative Reasoning Improves Long Horizon Agents
Training draft models to be rejected at critical points raises end-to-end success on complex agentic coding benchmarks without changing target model output.
Speculative decoding is recast as an active allocator for context precision in long-horizon agents. Numerical noise grows within verification windows and draft acceptance can alter model output distribution after 30 steps. The approach trains draft models to favor adversarial preference at identified decision points so critical tokens are sampled from the target model under least noisy conditions.
- Draft acceptance changes output distribution after 30 steps
- Training 0.06% to 0.28% of parameters improves accuracy
- Critical reasoning tokens sampled from target model
- Verification restarts at decision points
- Raises success rate on complex agentic coding benchmarks
From the original article
1 min read at huggingface.co
Speculative decoding has been treated purely as a systems optimization that speeds up large language model decoding without changing the model behavior. Under this view, the growing effort to train better draft models has been focused solely on acceptance rate. We argue that this view does not hold true for long-horizon agent reasoning.
By Hongyin Luo, Wei Fang, Lijie Yang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.