Agents pick each action before acting: Turn-BoN lifts terminal-task pass rates by 5–10% and halves the turns
Synopsis
Long-horizon agents execute one action per turn, sampled from a stochastic policy, so the same task sometimes succeeds and sometimes fails. Turn-BoN has a generator model propose several candidate actions at each turn and an off-the-shelf open-weight LLM selector choose one before the environment executes it. On Terminal-Bench v2.1 terminal tasks with Qwen3.6-35B and K2-Horizon-375B as generators, mean pass@1 rises by 5–10%, consistency across repeated attempts improves, and solving a task takes about 40–50% fewer turns than the baseline, at roughly 3–6 times the baseline's total tokens per trajectory.
Interpretation
Before each action is executed, a second LLM picks one of several candidate actions and only that action reaches the environment, so the change acts on the distribution of intermediate actions rather than only on the final answer. Trajectory-level best-of-N scores complete trajectories with an outcome reward model or LLM judge and selects afterward, so it cannot change the intermediate actions within an attempt; depth-first search and Monte Carlo tree search explore intermediate states but require checkpointing and restoring environment states. On Terminal-Bench v2.1 with the mini-SWE-agent harness, using Qwen3.6-35B and K2-Horizon-375B as generators and Kimi-K3 and Deepseek-v4.1-Flash as selectors, mean pass@1 averaged over four trials improves by 5–10% over the no-selector baseline, and the pass@k curve covers more tasks at small k.
The selector sees only a list of candidate actions and chooses the one most likely to make genuine progress, without training a separate value or critic model. Earlier guided search relied on an action-value critic trained from policy rollouts to score candidates, whereas this uses an off-the-shelf open-weight model as a generative critic that compares all candidates listwise in one call. The selector favors actions that do real work such as editing a file, testing the last edit, and running a test or check, reduces read-only inspection, and rejects malformed or empty tool calls and doom loops; at sampling widths N=4 and N=8, solving a task takes about 40% and 50% fewer turns than the baseline.
Trajectory-level behavior scoring shows the selector changes more than pass rates: failed trajectories also carry fewer bad behaviors. Baseline trajectories often rerun failed commands, retry without diagnosis, or rely on tests that cannot fail, and these behaviors are hidden by pass-rate metrics alone. An LLM judge annotated 360 trajectories against 28 behavior rubrics; looping drops by 14–18 percentage points, and failed trajectories average 2.7 flagged bad behaviors for K2-Horizon versus 3.9 at baseline, and 6.8 for Qwen3.6 versus 8.9 at baseline.
Perspective
The demonstrated scope is Terminal-Bench v2.1 terminal tasks with the mini-SWE-agent harness, generators Qwen3.6-35B and K2-Horizon-375B, and selectors Kimi-K3 and Deepseek-v4.1-Flash, where the selector is more capable than the generator; the authors propose applying Turn-BoN at only a few turns as a possible next step to reduce cost.
The authors frame the results as preliminary experiments and a thought experiment, with the main concern that generating several candidates and calling a selector every turn costs 3–6 times the baseline's total tokens; gains are capped by the generator's capability, how a weaker selector would cap them was not tested, and whether a trained small selector would generalize is unclear.
