The Pulse
Researchers Report Draft-Model Training Improves Long-Horizon Agent Reasoning
A paper describes a method that trains a draft model to direct critical reasoning tokens through less noisy verification positions. The authors report higher success rates on two complex coding-agent evaluations without changing the target
AI.info Team · Written with AI · fact-checked ·
A paper submitted to arXiv on Nov. 3, 2026, reports a method for improving long-horizon agent reasoning by changing how a draft model behaves during speculative decoding. The authors say the approach raises success rates on complex coding tasks without changing the target model or the output distribution it verifies. Their method trains the draft model to favor rejection at critical reasoning points identified in agent traces, so those tokens can be sampled again from the target model.
Speculative decoding can affect precision
In speculative decoding, a smaller draft model proposes a block of tokens and a larger target model scores that block in one forward pass. A rejection-sampling rule accepts a prefix of the proposed tokens; if a token is rejected, the target resamples that position from a residual distribution. The paper describes the method as lossless in distribution under its verification rule, and says it has generally been treated as a way to speed up decoding. Draft models have therefore often been judged by how many tokens the target accepts in each verification step.
The authors argue that this view misses a precision issue in hosted models. Their target models use low-precision weights, quantized key-value caches and sparse attention, and their outputs are not bitwise reproducible. The researchers reran forward passes on identical prefixes and compared top-k log probabilities. With GLM-5.3 and an eight-token DFlash verification block, they found that noise increased by roughly 4% per slot beyond the first. At the last slot, shifts greater than two natural-log units were nearly twice as common as at the first slot. They report a similar relative increase for DeepSeek-V4.1-Flash with a six-token DSPARK block, although its absolute noise was about half that of GLM-5.3.
Critical decisions may be more sensitive to noise
The paper focuses on long-running software-engineering agents, which can operate across dozens to hundreds of turns and thousands to millions of tokens. Each generated token is stored in the key-value cache and read by later parts of the agent’s reasoning. The authors say some thinking tokens can tolerate more noise than others, and identify decision-framing phrases as potentially important. In their audit of coding traces, they found that decisions framed around principles or evidence, or based on actual tests, took the correct side about twice as often as decisions guided by likelihood or risk. The paper also notes that decisions can change in response to environmental feedback, or that correct decisions can be ignored later in a task.
The method uses the position of a token in a verification block to allocate precision. Accepted draft tokens are committed at the precision associated with their verification slot. When a draft token is rejected, the target resamples it, and verification restarts from the first, least noisy slot. The researchers locate critical thinking points in agent traces and train the draft model with preference pairs that favor the framing the target should avoid. At those points, the draft and target diverge; the target’s choice is then sampled through the residual distribution and written under the less noisy conditions of a fresh block. Routine tokens can still be handled through ordinary speculative decoding.
Reported gains on coding tasks
The authors say the target model, verification rule and distributional guarantee remain unchanged; only the draft model is trained. On the DeepSWE software-engineering benchmark, the paper reports that the method raised the end-to-end success rate of GLM-5.3-nvfp4 from 60.6% to 67.6%, a gain of 7.0 percentage points. For DeepSeek-V4.1-Flash, the reported rate rose from 67.8% to 71.8%, a gain of 4.0 points under the standard evaluation harness and environments. The paper’s abstract says the approach uses training on 0.06% to 0.28% of hosted parameters. The results present speculative decoding not only as a decoding-speed technique, but also as a way to direct context precision toward selected reasoning steps.