- The paper introduces KATE, a framework that retrieves concrete tool-use trajectories, activates them through four-way parallel sampling and aggregation, and optionally applies knowledge-augmented reinforcement learning, improving Qwen3-8B by 15.5 points on BFCL-V3 over direct function calling.
- The paper finds that instance-level execution trajectories generally outperform intent-level summaries, while wider parallel sampling activates tool-use knowledge more reliably than deeper prompt-based reasoning and helps reduce planning errors and premature termination.
- The paper shows that KATE can outperform larger baseline models on selected benchmarks with relatively modest overhead, but performance declines on difficult AppWorld tasks when incorrect sampled trajectories overwhelm useful consensus.
Overview
This paper presents a systematic study of how experiential knowledge affects multi-turn tool use by LLMs, organized along three stages: knowledge acquisition, knowledge activation at inference time, and knowledge internalization through post-training. Building on the analysis, the authors propose KATE (Knowledge-Augmented Tool Execution), a framework that combines instance-level trajectory knowledge retrieved from an external knowledge base with width-based parallel sampling and aggregation, optionally followed by knowledge-augmented fine-tuning. Experiments on BFCL-V3 and AppWorld show consistent gains over strong baselines across model scales, including a 15.5-point average improvement on BFCL-V3 for Qwen3-8B over direct function calling (2606.10875).
The paper's framing departs from prior work that treats tool use primarily as prompt design, API documentation engineering, or alignment: it argues that many tool-use failures stem from missing concrete, executable experience—parameter constraints, scenario-specific operation patterns, error recovery strategies—rather than from reasoning incapacity alone.
Knowledge acquisition: instance-level versus intent-level knowledge
The authors construct four forms of experiential knowledge, differing in granularity:
- Scenario Trajectory Knowledge (ST): ground-truth tool execution trajectories used directly as in-context exemplars.
- Experience Summary Knowledge (ES): LLM-generated textual summaries of paired queries and ground-truth trajectories.
- Script-Style Intent Clustering Knowledge (SIC): semi-structured JSON "standard operating procedures" induced by clustering training examples via inferred user intents (K-Means over embeddings, batch-wise hierarchical summarization with GPT-4o).
- Textual-Style Intent Clustering Knowledge (TIC): natural-language descriptions of each intent cluster's operational strategy.
Retrieval uses all-MiniLM-L6-v2 embeddings with a similarity threshold (p=0.5) and top-K selection; for intent-level knowledge, the model first infers the current intent, which is then encoded as the retrieval query (K=1).
Three findings emerge from the augmentation experiments on BFCL-V3 with Qwen3-8B and Qwen3-32B. First, instance-level knowledge consistently outperforms intent-level knowledge, which the authors attribute to the multi-step abstraction and imperfect intent inference required by intent-level retrieval. Second, ST and ES perform comparably, with no single form universally dominating across tasks and backbones. Third—and notably—naively stacking multiple knowledge types does not guarantee further gains; redundancy or interference can occur, motivating selective activation rather than indiscriminate aggregation. The practical implication is that high-quality execution trajectories alone suffice for substantial improvements, simplifying knowledge base construction considerably.
Knowledge activation: width beats depth
At inference time, the paper contrasts two activation strategies given fixed knowledge.
Depth-based prompt hints append structured guidance after each tool execution, prompting the model to reason explicitly over intent, reflection, and state before the next call. Results are mixed: gains appear in some scenarios but prompts often yield limited improvement or even degrade accuracy, plausibly because constraining reasoning to predefined perspectives restricts flexibility in multi-factorial tool-use settings.
Width-based parallel sampling with aggregation samples N=4 candidate actions per interaction step at temperature 1; if all candidates agree the action executes directly, otherwise an aggregation function decides. Two aggregators are compared: self-consistency (majority voting) and LLM-based aggregation over candidates. Parallel sampling substantially improves tool-calling accuracy, supporting the claim that greedy decoding fails to fully activate latent experiential knowledge. Self-consistency is more stable than LLM-based aggregation and largely insensitive to sampling scale, though the authors note it is less suitable for tasks requiring diverse structured outputs such as code-based agents—a caveat relevant to their own AppWorld evaluation. LLM-based aggregation improves performance but not monotonically with sample count, likely due to limited capacity to reason over long candidate lists.
The central conclusion is that expanding reasoning width is more effective than deepening it for eliciting tool-use knowledge, a finding consistent with test-time scaling trends but here instantiated specifically for multi-step tool invocation.
Knowledge internalization through post-training
To move beyond context injection, the authors augment training data by concatenating retrieved experiential knowledge with user instructions, then compare SFT and RL (GRPO with LoRA rank 32). On Qwen3-8B/BFCL-V3, RL alone achieves the best post-training result (48.25 average, +15.50 over FC), exceeding SFT (45.75) and the SFT-then-RL sequence (46.25). This is a somewhat counterintuitive result: for sufficiently strong base models, direct RL better explores and reinforces tool-calling capability than SFT within the same data budget, with reward curves showing consistently better convergence quality for RL. Pre-inserting knowledge into the RL context increases the probability of sampling correct trajectories, improving training efficacy—an application of hint-assisted RL to the tool-use setting.
The KATE framework and main results
KATE integrates the strongest components identified above: ST knowledge for acquisition, parallel sampling with aggregation for activation, and optional knowledge-augmented post-training. Key results on BFCL-V3:
| Model |
Method |
Average |
| GPT-5 |
FC |
37.75 |
| GPT-4.1 |
FC |
44.25 |
| Qwen3-8B |
FC |
32.75 |
| Qwen3-8B |
Memp |
36.00 |
| Qwen3-8B |
KATE |
46.00 (+13.25) |
| Qwen3-8B |
KATE + RL |
48.25 (+15.50) |
| Qwen3-32B |
FC |
46.00 |
| Qwen3-32B |
Memp |
49.00 |
| Qwen3-32B |
KATE |
50.50 (+4.50) |
A striking claim is that KATE-equipped Qwen3 models outperform GPT-4.1 and GPT-5 on this benchmark, suggesting that explicit experiential knowledge can compensate for raw model capability gaps in narrow tool-use domains. Ablations confirm both components matter: removing parallel sampling ("w/o PS") drops Qwen3-8B to 38.00, showing knowledge injection alone underutilizes the injected experience; removing knowledge ("w/o Exp") drops it to 42.75, showing sampling alone has a limited ceiling. Replacing LLM-based aggregation with self-consistency slightly reduces Qwen3-8B performance but matches or exceeds full KATE on Qwen3-32B (50.25), indicating aggregator choice interacts with model scale.
On AppWorld, KATE improves over ReAct substantially (e.g., 10.92 vs. 4.1 average for Qwen3-8B; 12.87 vs. 6.52 for Qwen3-32B), but on Test-Challenge it falls below Memp on some metrics. The authors attribute this to a trade-off between reasoning width and task complexity: when tasks exceed the model's inherent capacity, sampled trajectories are mostly incorrect and aggregation amplifies noise rather than consensus. Error analysis using gpt-5-mini shows planning and reasoning errors dominate across methods and are reduced most by trajectory-level supervision plus parallel sampling, which also mitigates premature termination in long-horizon execution.
Efficiency analysis partially addresses the obvious cost concern: despite four-way sampling plus aggregation, KATE increases token consumption only 1.89× (Qwen3-8B) and 1.29× (Qwen3-32B) relative to FC, because higher accuracy shortens unnecessary reasoning; self-consistency variants approach parity in token cost. Multi-seed replications confirm stability of the width-based gains, and results extend to Llama3.2-3B-Instruct, where KATE reaches 11.25 average versus 2.00 for FC, though the gain over the knowledge-only variant is modest there, possibly due to the small model approaching its ceiling.
Limitations and open questions
The paper concedes several constraints. The evaluation uses a relatively small-scale knowledge base (100 training samples for BFCL-V3; 81 distilled examples for AppWorld), so the effect of scaling the knowledge repository is unexplored. All results come from single runs in the main tables (multi-seed statistics appear only in supplementary analysis), and the study is restricted to text-only tool-use scenarios, leaving multimodal extension open. The AppWorld Test-Challenge result leaves unresolved how to balance sampling width against task difficulty—for instance, whether adaptive sampling budgets conditioned on estimated task difficulty would recover the losses observed on hard tasks. Additionally, self-consistency's unsuitability for code-based agent outputs suggests a need for structure-aware aggregation strategies that the paper does not develop.
Conclusion
This work provides a principled decomposition of experiential knowledge's role in LLM tool calling across acquisition, activation, and internalization stages. Its main empirical contributions are the demonstration that simple instance-level trajectory knowledge suffices for large gains, that reasoning-width expansion via parallel sampling activates latent knowledge more effectively than depth-oriented prompting, and that knowledge-augmented RL internalizes this experience better than SFT. KATE operationalizes these findings into a framework achieving state-of-the-art training-free and training-based results on BFCL-V3 and AppWorld, with favorable token-cost characteristics. The remaining questions—knowledge-base scaling, multimodal extension, and width–complexity trade-offs on very hard tasks—define the boundaries of the current evidence.