- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper tackles the performance–efficiency tradeoff of LLM agents for open-world tool use. It presents PARAACT, a structured parallel action loop alternating between parallel exploration for capability gaps and DAG-based parallel execution of dependent tools. PARAAGENT learns this loop via two-stage training: multi-agent demonstration SFT for cold start, plus reinforcement learning with decoupled hierarchical rewards at step, phase and trajectory levels. A simulated environment TOOLENV built from real API data supports low-cost training. Tested on ToolBench and API-Bank, the 4B PARAAGENT surpasses many larger baselines including GPT-4.1, achieving strong multi-tool performance within ~3.74 turns per task. Improvements stem from structured action planning instead of more interactions or bigger models. Ablations validate the proposed training and reward design. Limitations are the 4B policy and reliance on simulation; future work scales models and extends to skill ecosystems.
Critical strengths
-
The paper proposes an effective parallel planning method that improves model efficiency and tool-calling performance. The framework design and task decomposition demonstrate a degree of novelty.
-
The detailed experimental analysis supports the design choices and demonstrates the contributions of individual modules.
Critical weaknesses
- The experiments use only one relatively small model, leaving the method’s generalizability across model sizes and capabilities unclear.
Questions for authors
-
How does the proposed method perform on widely used tool-use benchmarks such as BFCL and τ-bench? Evaluation on these benchmarks would help assess its generalizability to more challenging tasks.
-
The proposed method involves additional training, whereas the part of baselines do not. Can the authors provide comparisons under matched training conditions to clarify whether the improvements come from the proposed framework or the additional training?
4 comments
Response to Reviewer azwX (part 2)
Q3: The proposed method involves additional training, whereas the part of baselines do not. Can the authors provide comparisons under matched training conditions to clarify whether the improvements come from the proposed framework or the additional training?
Thanks for the suggestion. To better understand where the gains come from, we conduct three controlled analyses:
- (1) Training-controlled comparison (Table 4). We train ReAct-4B with RL using the same RL data, settings, and outcome reward as ParaAgent. (i) ParaAgent achieves higher performance at lower cost under matched training conditions. (ii) It also outperforms ToolOmni-4B and DeepAgent-32B (SOTA models trained with RL ). Thus, with comparable training conditions, the framework still wins.
- (2) Training-free comparison (Table 5). Under the same prompting, ParaAct on Qwen3-4B/32B outperforms ReAct with fewer turns, whereas naive parallel underperforms ReAct. On Qwen3-4B, ParaAct approaches DFS with fewer turns. These results show that the framework can improve the performance–interaction trade-off without training.
- (3) Joint view of framework × training (Table 6). Training also improves ReAct and narrows the overall ToolBench gap, but benefits ParaAct more: ParaAgent gains +17.5 Avg with fewer turns, while RL increases ReAct's turns. ParaAct's advantage grows on the hardest subset, and its API-Bank deficit turns into a gain, showing that our training method amplifies the framework's structural advantages.
Overall, our framework and training are complementary: ParaAct provides basic gains, and ParaAgent training realizes them fully, especially on complex tasks. We will add these in the revision.
| Method | ToolBench I1 ↑ | I2 ↑ | I3 ↑ | Avg. ↑ | Turns ↓ | Path ↑ | API-Bank Suc. ↑ | Acc. ↑ | Turns ↓ |
|---|---|---|---|---|---|---|---|---|---|
| ToolOmni-4B | 29.1 | 28.3 | 21.3 | 26.2 | 11.01 | 39.1 | 20.0 | 46.6 ± 0.0 | 10.22 |
| DeepAgent-32B | 39.7 | 37.8 | 27.9 | 35.1 | 13.34 | 31.9 | 24.0 | 59.9 ± 2.7 | 7.76 |
| ReAct-4B + direct RL | 57.17 | 62.61 | 44.26 | 54.68 | 6.67 | 32.1 | 22.00 | 56.49 | 5.00 |
| ParaAgent-4B | 58.4 | 65.7 | 55.7 | 59.9 | 3.30 | 42.9 | 24.0 | 58.0 ± 0.0 | 4.19 |
Table 4. RL-trained baselines (Success: Pass@2).
| Model / Method | ToolBench I1 ↑ | I2 ↑ | I3 ↑ | Avg. ↑ | Turns ↓ | API-Bank Suc. ↑ | Acc. ↑ | Turns ↓ |
|---|---|---|---|---|---|---|---|---|
| Qwen3-32B / ReAct | 46.8 | 48.3 | 27.9 | 41.0 | 5.16 | 22.0 | 53.4 | 4.34 |
| Qwen3-32B / DFS | 50.6 | 47.4 | 39.3 | 45.8 | 7.07 | 22.0 | 56.5 | 4.57 |
| Qwen3-32B / Parallel | 35.7 | 33.0 | 26.2 | 31.6 | 4.75 | 10.0 | 48.1 | 4.33 |
| Qwen3-32B / ParaAct | 52.53 | 59.13 | 49.18 | 53.61 | 4.47 | 20.0 | 57.25 | 4.22 |
| Qwen3-4B / ReAct | 33.8 | 36.1 | 23.0 | 30.9 | 5.76 | 14.0 | 46.2 | 4.44 |
| Qwen3-4B / DFS | 47.5 | 53.0 | 42.6 | 47.7 | 30.81 | 18.0 | 49.6 | 8.11 |
| Qwen3-4B / Parallel | 31.0 | 32.2 | 24.6 | 29.3 | 4.78 | 18.0 | 28.2 | 4.18 |
| Qwen3-4B / ParaAct | 48.10 | 51.30 | 27.87 | 42.42 | 4.00 | 18.0 | 42.37 | 4.37 |
Table 5. Training-free comparison (Success: Pass@2).
| Setting | Method | ToolBench Avg ↑ | I3 ↑ | Turns ↓ | API-Bank Suc ↑ | Acc ↑ | Turns ↓ |
|---|---|---|---|---|---|---|---|
| Training-free | ReAct | 30.9 | 23.0 | 5.76 | 14.0 | 46.2 | 4.44 |
| Training-free | ParaAct | 42.4 | 27.9 | 4.00 | 18.0 | 42.4 | 4.37 |
| Δ | +11.5 | +4.9 | −1.76 | +4.0 | −3.8 | −0.07 | |
| Matched RL | ReAct | 54.7 | 44.3 | 6.67 | 22.0 | 56.5 | 5.00 |
| Matched RL | ParaAct | 59.9 | 55.7 | 3.30 | 24.0 | 58.0 | 4.19 |
| Δ | +5.2 | +11.4 | −3.37 | +2.0 | +1.5 | −0.81 |
Table 6. Framework × training (Success: Pass@2).
References
[1] Wang, Zhenting, et al. MCP-Bench: Benchmarking Tool-Using LLM Agents with Complex Real-World Tasks via MCP Servers. arXiv:2508.20453, 2025.
[2] Mo, Guozhao, et al. LiveMCPBench: Can Agents Navigate an Ocean of MCP Tools? arXiv:2508.01780, 2025.
Follow-up on Author Rebuttal
Dear Reviewer azwX,
We would appreciate it if you could take another look at our responses and let us know whether your concerns have been addressed.
Thank you for your time and consideration.
Second Follow-up on Rebuttal
Dear Reviewer azwX,
We would like to kindly follow up once again on our rebuttal. We have added experiments on a smaller 1.7B model, two real-world MCP benchmarks, and controlled comparisons under matched training conditions to address your questions.
With the discussion period nearing its end, we would greatly appreciate your feedback on whether these results resolve your concerns.
Thank you for your time and consideration.
Response to Reviewer azwX (part 1)
Q1: The experiments use only one relatively small model, leaving the method’s generalizability across model sizes and capabilities unclear.
We thank the reviewer for the suggestion. Given the time constraints, we add experiments on Qwen3-1.7B, under the same settings as our 4B model. Our method remains effective, improving average success from 4.97 to 29.70 after SFT and further to 42.33 after RL,
demonstrating its generalizability across model sizes. We will add these results to the revised paper and include larger models.Table 1. Model-size comparison (Success: Pass@1; Δ vs. backbone).
Q2: How does the proposed method perform on widely used tool-use benchmarks such as BFCL and τ-bench? Evaluation on these benchmarks would help assess its generalizability to more challenging tasks.
Thank you for the suggestion. Since BFCL and τ-bench (tool states are bound to specific tasks) are difficult to adapt to dynamic tool discovery and evolving execution feedback, we add MCPBench [1] and LiveMCPBench [2], two large-scale benchmarks built on real, running MCP servers. For both, we verified and configured all accessible tool interfaces, adapted them to the open-world setting, and used their official evaluation protocols.
Overall, these results demonstrate
ParaAgent's effectiveness and generalizability to unseen real-world MCP servers, achieving a better performance–efficiency trade-off.Table 2. MCPBench performance.
Table 3. LiveMCPBench performance (Pass@1).