Adversarial Speculative Reasoning Improves Long Horizon Agents
| Source: arXiv AI
Tags: speculative decoding, draft models, agentic coding, long-horizon agents, LLM inference, numerical precision
Speculative decoding is not neutral for long-horizon agents: numerical noise shifts reasoning after about 30 steps, and training the draft model to be rejected at critical decision points raises agentic coding success without changing the target model.
Details
Speculative decoding is usually treated as a pure speed optimization that leaves model behavior unchanged. The authors argue this breaks down in long agent runs. Measuring frontier open-weight models on complex agentic tasks, they found numerical noise grows inside verification windows and draft acceptance can shift the output distribution after about 30 steps, moving reasoning accuracy in random directions. Their fix trains the draft model to be rejected at critical reasoning points found in agent traces. Draft and target then diverge exactly there, verification restarts, and key thinking tokens are sampled from the target model under its least noisy conditions. On complex agentic coding benchmarks this raised end-to-end success with no change to the target model, training 0.06% to 0.28% of hosted parameters.