Choosing each next move can improve AI agents—but uses far more tokens

A second opinion before every action

An AI agent working in a terminal must repeatedly decide what to do, execute a command and interpret the result. A poor choice can waste time or send an otherwise promising attempt off course. In the preprint On Turn-Level Best-of-N Sampling for Scaling LLM Agents, Jalaj Bhandari and colleagues investigate a straightforward intervention: generate several possible next actions, then ask another language model to choose one.

The paper, submitted as arXiv version 2611.06868v1 on November 4, 2026, calls this method Turn-BoN, short for turn-level best-of-N sampling. Its findings are preliminary benchmark results, not evidence of universal reliability.

How the selection works

At each turn, a generator model proposes N candidate actions. A selector receives the task, the interaction history and all candidates together, without the generator’s private reasoning. It chooses the action it considers most likely to make genuine progress. Only that action is executed; its result becomes part of the next turn’s context.

Neither model needs additional training for this setup. The selector uses an instruction prompt rather than a specially trained numerical scoring function.

This differs from generating several complete attempts and picking the best afterward. Turn-BoN intervenes during an attempt, potentially avoiding mistakes before they happen. It also avoids executing every candidate in a separate environment or restoring earlier environment states. However, it does not test the consequences of unselected actions before choosing.

The authors situate their work alongside earlier guided-search research and concurrent work on per-turn selection, rather than claiming to have invented action selection itself.

Higher success, with limits

The experiments cover 89 Terminal-Bench v2.1 tasks using the mini-SWE-agent harness, the software that manages interactions between models and the terminal. The generators are Qwen3.6-35B and K2-Horizon-375B; the selectors are Kimi-K3 and DeepSeek-v4.1-Flash. Each configuration is evaluated over four trials, primarily with four or eight candidates per turn.

Pooling results across the two selectors, Qwen’s success rate rises from 47.5% without selection to 52.8% with four candidates and 55.2% with eight. K2-Horizon rises from 73.3% to 77.9% with four candidates, but reaches 76.7% with eight. These are percentage-point changes, and the K2 results show that increasing the candidate count does not automatically improve success.

The authors also report better coverage across repeated attempts and more tasks solved consistently. These measure different things: solving a task at least once is not the same as solving it every time.

Fewer turns does not mean cheaper execution

Selected actions more often edit files or run checks, and less often only inspect information. They also tend to combine more shell commands into one response. Together, these changes help explain shorter trajectories.

For example, K2-Horizon averages 69.5 executed turns without selection, versus 33.9 with eight candidates and Kimi-K3. But producing discarded candidates and running the selector substantially increases inference work. Across the reported configurations, combined generator and selector output totals range from 3.6 to 8.7 times the corresponding baseline.

Those figures count output tokens, not a complete financial or computational cost. The study does not establish faster wall-clock execution. Under the same eight-hour limit, timeouts actually become more frequent in several selection configurations.

Better habits, imperfect judgment

A separate analysis examines 360 trajectories from 36 tasks where K2-Horizon’s baseline failed at least once. An LLM judge evaluates 28 undesirable behaviors, with each trajectory judged twice. The authors report less looping and fewer retries without diagnosis, alongside improvements in testing and instruction following that vary by generator. Failed attempts also receive fewer behavioral flags on average.

These annotations are model judgments on a selected task subset, not an independent human assessment. Nor does every behavior improve. One case study shows a selector favoring a script that executes immediately but uses incorrect output keys, then accepting checks aligned with the agent’s own mistake.

A separate model-based audit reports rare reward hacking and no statistically significant increase under Turn-BoN. That does not prove shortcuts are absent: the paper documents a selector choosing actions that retrieve a benchmark reference solution.

What the study establishes

The theoretical analysis gives a conditional explanation: selection improves expected performance when its average mistakes cost less than the benefit available from sampling alternatives. It does not guarantee that an off-the-shelf selector satisfies that condition.

The experiments use selectors chosen to be more capable than the generators; weaker selectors are not tested. Selector-only performance is also not measured in the same harness. The study therefore leaves open whether this arrangement is preferable to directly using a stronger model at comparable cost.

The central result is narrower but useful: choosing among proposed actions can improve tested agents’ success and behavior without retraining, while spending substantially more tokens. Finding cheaper ways to retain those gains remains future work.