Reasoning is not needed at every turn
Most LLM agents write a reasoning step before every action across a long interaction. That is expensive, and the paper argues it is often unnecessary: reasoning produced at one turn can keep supporting actions several turns later. The authors call this cross-turn reasoning persistence. Some reasoning steps help only the next turn, while others remain useful across multiple turns, because plans, constraints, or intermediate judgments stay relevant after the turn that produced them.
The practical question is therefore not whether an agent should reason, but when its existing reasoning has become insufficient. A new reasoning step is warranted when changes in the interaction context make earlier reasoning unable to support the next action. The paper's contribution is a way to detect that condition without repeatedly generating candidate actions to test each reasoning turn.
A likelihood drop stands in for action recovery
The paper's key observation is empirical. If later reasoning is removed and a subsequent reference action can still be recovered from earlier reasoning, the likelihood of that reference action stays relatively stable. If the action is no longer recoverable, its likelihood drops. The authors measure this as the mean per-token log-likelihood of the reference action under the full context minus its log-likelihood under a truncated context that keeps only earlier reasoning.
This matters because it replaces generation with scoring. Testing whether an action survives removal of a reasoning step by regenerating actions is costly; scoring a known reference action under two contexts only requires batched teacher-forced passes. The paper reports that small likelihood decreases generally correspond to reliable recovery and larger decreases to lower recovery rates, which is why the decrease can serve as a lightweight proxy for cross-turn action support.
LoGiC finds reasoning turns that can be dropped
RACE builds on a procedure called Likelihood-Guided Progressive Reasoning Cover Detection, or LoGiC. It progressively identifies reasoning turns whose removal has limited impact on the current and subsequent reference actions, using the likelihood-decrease signal rather than repeated generation. The output is a set of removal signals marking where reasoning was not carrying its weight.
Those signals are then used in training in two stages. In supervised fine-tuning, the authors apply LoGiC to successful trajectories annotated by DeepSeek-V4-Pro-0813, using Qwen3.5-9B and a likelihood-decrease threshold of 0.001, and fine-tune on the resulting adaptive reasoning trajectories. In agentic reinforcement learning, training starts from the corresponding RACE-SFT checkpoint and LoGiC is applied only to successful on-policy trajectories, again with a 0.001 threshold. The policy thus learns from examples of when reasoning could be removed without breaking the actions that followed.
Four benchmarks measure success and reasoning tokens
Evaluation covers ScienceWorld, WebShop, AppWorld, and DeepSearch, with episodes capped at 30 interaction turns and up to 65,536 tokens per generation. The two reported quantities are task success rate and the average total number of reasoning tokens per trajectory, so reasoning cost is measured directly rather than inferred from wall-clock time.
The success definitions differ by benchmark and should not be conflated. ScienceWorld uses 1,819 test instances across 30 task types, with success only when the final environment score reaches 100. WebShop uses 500 crowd-sourced instructions from the official test split, with success at environment reward 1. AppWorld reports overall strict task success over 585 tasks across Test-Normal and Test-Challenge. DeepSearch uses 400 questions scored by Exact Match. A gain on one of these is not interchangeable with a gain on another.
Reasoning cost falls while success holds or improves
Across the four benchmarks, the authors report that RACE substantially reduces reasoning cost while maintaining or improving task performance. The claim is stated at the level of the full benchmark suite and in the paper's own two metrics: fewer reasoning tokens per trajectory, with success rate not sacrificed.
The qualitative behavior is visible in the paper's worked trajectories. In a WebShop example, the RACE 9B model reasons when constructing the search query and when checking whether a product description mentions individual wrapping, but emits empty reasoning blocks for straightforward clicks such as selecting an item, opening a page, or buying. In an AppWorld example, it reasons to plan credential retrieval and contact lookup, then acts directly once the plan is established. These examples illustrate the intended pattern rather than quantify it; the quantitative claim rests on the benchmark-level token and success measurements.
The signal is validated within the paper's own setup
The likelihood proxy is supported by the paper's own analysis comparing likelihood decrease with generation-based recovery rate, and the training recipe is tied to specific choices: Qwen3.5-9B as the base model, DeepSeek-V4-Pro-0813 annotations for SFT data, a 0.001 likelihood-decrease threshold, and benchmark-specific learning rates and batch sizes. The reported results come from these four agent benchmarks under the stated turn and token caps.
Readers should also keep the two stages distinct. SFT learns from removal signals on annotated successful trajectories, while RL applies LoGiC only to successful on-policy trajectories and adds a closure-loss term weighted at 0.02 for ScienceWorld, WebShop, and AppWorld and 0.01 for DeepSearch. The evidence supports the combined recipe as evaluated; it does not separately establish how much each component contributes, nor how the threshold would transfer to other models or task suites.
Reason when earlier reasoning stops covering the action
The paper's distinctive priority is to treat reasoning as something with a persistence span rather than a per-turn requirement. The decision rule it supports is conditional: keep acting directly while the likelihood of the next reference action remains stable without fresh reasoning, and insert a reasoning step when that likelihood would drop sharply. Because the check is a scoring pass rather than a generation, it can be applied during training at a fraction of the cost of verifying each turn by regenerating actions.
The most consequential comparison is between that likelihood-guided signal and turn-by-turn necessity checks that ignore cross-turn persistence. The paper's argument is that the former captures whether earlier reasoning still covers later actions, while the latter assesses each turn in isolation and requires repeated generations to obtain removal signals. For readers deciding how to train a cheaper agent, the relevant threshold is the likelihood decrease at which reference actions stop being recoverable, and the relevant evidence is the paired reporting of reasoning tokens and task success across the four benchmarks.
Sources used in this answer
When Should Agents Think? Adaptive Reasoning via Cross-Turn Estimation
RACE uses decreases in reference-action likelihood after removing later reasoning as a lightweight proxy for whether earlier reasoning still supports subsequent actions, then trains agents via SFT and RL to skip unnecessary reasoning turns, substantially reducing reasoning cost while maintaining or improving task performance on four agent benchmarks.