CITA trains a comparative inference model with a Bayesian tool-graph simulator and LLM comparison judgments, consistently improving Tool F1 and task success across three tool-use benchmarks and multiple backbone LLMs

It proposes CITA, framing the tool-use agent's decision objective as estimating the long-horizon value of a possible next tool invocation before executing it, and trains a Comparative Inference Model (CIM) to output how likely that candidate invocation is to support final task success under the current context. Relative to approaches that rely on final-outcome rewards or step-level rewards, the work places supervision explicitly on comparisons among alternative invocations under the same context rather than scoring only the action that was taken. The abstract states that CITA consistently improves Tool F1 and task success across three tool-use benchmarks and multiple backbone LLMs, and reports analysis showing CIM learns accurate step-level value estimates; specific numbers, benchmark names, and model lists are not given in the abstract.

It designs a combination of paired supervision signals: observed tool behavior, scalable supervision from a Bayesian tool-graph simulator, and semantic judgments from LLM-based comparison. Addressing the fact that logged trajectories contain only the invocation actually taken and thus lack counterfactual candidate supervision, the work uses a simulator for scalable supervision and LLM comparison for semantic judgment to construct paired signals under the same context. The abstract explicitly lists these three signal sources and states they are used to train CIM; it does not give simulator scale, the specific configuration of LLM comparison, or ablation results.

It validates CITA across multiple backbone LLMs and reports consistent improvements in Tool F1 and task success. Relative to reporting results on a single model or benchmark, the work provides cross-model and cross-benchmark consistency evidence across three tool-use benchmarks and multiple backbone LLMs. The abstract reports consistent improvements across three benchmarks and multiple backbone LLMs; it does not provide effect sizes, confidence intervals, or statistical tests.

The work targets long-horizon tool-use agents and applies to settings that require evaluating the value of a possible next tool invocation before execution, such as action selection and training-signal construction in multi-step tool-calling tasks. Its method depends on paired supervision signals, including observed tool behavior, scalable supervision from a Bayesian tool-graph simulator, and semantic judgments from LLM-based comparison, so it fits settings where comparisons among candidate invocations can be constructed or simulated. The abstract's results cover three tool-use benchmarks and multiple backbone LLMs, indicating a reproducible direction for improvement within those evaluation scopes; for agent researchers and practitioners seeking to reduce human or extra-rollout supervision cost, CITA offers a path that replaces direct step labeling with a comparative inference model.

The abstract does not give effect sizes, benchmark names, the list of backbone LLMs, or how the Bayesian tool-graph simulator is built and scaled, nor does it describe the prompt design and consistency controls for LLM comparison judgments. How the accuracy of CIM's step-level value estimates is measured, and under what conditions it degrades, is mentioned only as an analysis conclusion. Because the reading scope here is the abstract only, without the body figures and experimental details, these points remain open questions that require reading the original text to confirm.