CIPO trains LLM agents with counterfactual branches to adaptively pick tool granularity, raising success and decision compression across three benchmarks and two backbones

It formulates when to use an atomic tool versus a composite skill as adaptive tool granularity selection, supported by pilot analyses: skills are better in 37% of examined states, atomic tools in 21%, and the rest are comparable, while SFT and GRPO + Skills keep skill invocation rates far from that oracle trend. Prior memory, workflow-reuse, and tool-merging approaches mostly treat reuse as context or fixed guidance, do not change the action space, and rarely train state-dependent granularity choice. Evidence comes from pilot statistics on successful TOOLATHLON trajectories (several subsequences appear over 100 times; relaxing mining constraints inflates both library size and candidate overlap), which is motivational analysis rather than a causal experiment.

It introduces budget-constrained BPE-style skill mining plus executable skill instantiation: tool invocations are abstracted by tool name and parameter-key signature and merged under maximum merge rounds, minimum pair frequency, and maximum pattern length, then registered as callable skills with exposed parameters, internal piping, selection control, execution traces, and soft interruption. Unlike direct tool merging or treating subsequences as fixed pipelines, this design bounds library size and candidate length (propositions give at most one candidate per merge round and length within the cap) and keeps skills inspectable and recoverable at runtime. Method-level construction plus two size and length bound propositions; ablations show removing budgeted mining or executable instantiation lowers average F1 and TSR (42.82/31.75 and 42.10/31.05 versus 48.23/36.48 for full CIPO).

It proposes CIPO: on top of GRPO, each base rollout is branched at the first eligible granularity decision, the chosen action is replaced by a feasible action at the other granularity, the same policy completes the branch, and the paired outcome difference is scaled and clipped into a bounded supplementary reward while the task reward remains the main signal. Compared with standard GRPO and GRPO + Skills, CIPO directly compares the consequences of two granularities from the same decision context instead of relying only on rollout-level outcome rewards. Ablations show removing the counterfactual reward while keeping branches yields average F1/TSR of 43.25/31.65, below CIPO-first at 48.23/36.48; on branch-point choice CIPO-first beats CIPO-random (46.72/34.90) and CIPO-late (45.76/34.10).

Across six benchmark-backbone settings CIPO achieves the highest Tool F1 and TSR, with an average Comp./Raw of 0.845, and skill usage rises toward a coverage-derived reference, indicating the gain comes from better granularity selection rather than more skill calls. Relative to tool-composition and skill-construction baselines (e.g., AWO 0.86, LLMCompiler 0.93, SKILL0 1.00), CIPO compresses more while keeping the highest TSR. The main table covers six benchmark-backbone settings; TSR is the mean over five independent LLM-judge evaluations of 100 episodes per setting; under yes-only strict judging CIPO exceeds GRPO + Skills by 6.28 percentage points, and by 7.50 percentage points in the TOOLATHLON execution-based comparison.

The results target tool-use environments with enumerable atomic tools, minable successful trajectories, and registrable skill interfaces, covering long-horizon tasks such as retrieval, file access, structured processing, and API-style operations; at inference the policy selects actions directly from the augmented space without counterfactual branching. For engineering teams aiming to reduce policy-level decisions while preserving task completion, CIPO offers a reusable pipeline: budget-constrained mining controls library size, executable instantiation keeps skills runnable in new states, and counterfactual branches supply state-dependent granularity feedback. The authors also note that in medical, legal, financial, or safety-critical settings where erroneous tool execution could cause material harm, such agents should be paired with human approval, sandboxed execution, tool permission control, logging, and domain-specific safety checks rather than deployed directly.

The skill-usage reference is explicitly framed by the authors as a library-level reference rather than a state-level optimal policy, and training-time online rollouts and final evaluation trajectories come from different distributions, so their skill-usage ratios need not match; the counterfactual reward is positioned as a bounded supplementary signal, not a standalone causal estimate. The confidence-interval table in Appendix J and the strict/execution-based evaluation table in Appendix L.1 are not given with concrete numbers in the main text, so the magnitude of that supplementary evidence is known only from the prose. In addition, skill-mining budget, reward scale, and clipping bound show sensitivity, so suitable values still need recalibration when transferring to a new tool environment.