Elamur Research⌕
☰
home / papers / 2026 / Reinforced Agent: Inference-Time Feedback for Tool-Calling Agents

Reinforced Agent: Inference-Time Feedback for Tool-Calling Agents

Paper · 2026 · Anh Ta, Junjie Zhu and Shahin Shayandeh

Summary

LLM agents that call tools often choose the wrong function, fill in arguments incorrectly, or miss that no available tool fits a request. Usual trajectory evaluation catches these mistakes only afterwards, when a harmful action may already have run. This Apple paper moves the check into the execution loop. A separate reviewer agent inspects each provisional tool call before it runs. It then either sends corrective feedback or picks the best of several candidates.

Three strategies are compared:

  • Progressive feedback (rN): the reviewer critiques and the agent revises, for up to N rounds.
  • Best-of-N selection (sN): the reviewer picks one of N candidate calls.
  • Best-of-N grading (gN): the reviewer scores each candidate and the top one is used.

The base agent, GPT-4o, is never retrained. Only the reviewer's model and prompt are tuned, and the prompts are refined automatically with GEPA. The authors also define Helpfulness-Harmfulness metrics, which weigh the base-agent errors a reviewer fixes against the correct answers it breaks.

On BFCL and τ2-Bench the method gives clear gains, and reasoning models make better reviewers. The work offers a modular, training-free path to more reliable agents, at a measurable latency cost.

Key points

  • Accuracy gains: On BFCL, irrelevance detection rises from 84.9% to 90.4% (+5.5) and the relevance suite from 90.9% to 92.5%. On τ2-Bench (airline, retail, telecom), pass rate rises from 48.7% to 55.8% (+7.1).
  • Reviewer choice: o3-mini as reviewer reaches 36.8% helpfulness against 11.7% harmfulness, roughly 3:1. GPT-4o as its own reviewer manages only about 2.1:1.
  • Prompt optimization: GEPA with a GPT-5 mini reviewer adds +1.5% on relevance and +2.8% on irrelevance. Both GEPA and the Helpfulness-Harmfulness analysis were run on BFCL only.
  • Latency cost: On BFCL, average latency grows 6.2× (1.27s → 7.87s). On τ2-Bench, episode duration grows 2.4× (158.7s → 384.3s), because the overhead is spread over about 40 turns per episode.
  • Main failure mode and limits: The reviewer tends toward over-skepticism, for example flagging valid calls as incomplete because execution results are missing. In early runs, 23% of cases had needless review iterations. Only GPT-4o was tested as the base agent.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

Cite

Cite as Anh Ta, Junjie Zhu and Shahin Shayandeh. 2026. Reinforced Agent: Inference-Time Feedback for Tool-Calling Agents. Proceedings of the Fifth Workshop on Generation, Evaluation and Metrics (GEM).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.gem-main.13

@inproceedings{ta2026reinforced,
  title = {{Reinforced Agent: Inference-Time Feedback for Tool-Calling Agents}},
  author = {Anh Ta, Junjie Zhu and Shahin Shayandeh},
  booktitle = {Proceedings of the Fifth Workshop on Generation, Evaluation and Metrics (GEM)},
  year = {2026},
  doi = {10.18653/v1/2026.gem-main.13},
  url = {https://doi.org/10.18653/v1/2026.gem-main.13}
}

References

15 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Lakshya A Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl-Ong, Arnav Singhvi, Herumb Shandilya, Michael J Ryan, Meng Jiang, Christopher Potts, Koushik Sen, Alexandros G. Dimakis, Ion Stoica, Dan Klein, Matei Zaharia, and Omar Khattab. 2025. Gepa: Reflective prompt evoluti… find
  2. Victor Barres, Honghua Dong, Soham Ray, Xujie Si, and Karthik Narasimhan. 2025. τ 2 -bench: Evaluating conversational agents in a dual-control environment. Preprint, arXiv:2506.07982. find
  3. Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. 2015. Distilling the knowledge in a neural network. Preprint, arXiv:1503.02531.
  4. Namgyu Ho, Laura Schmid, and Se-Young Yun. 2023. Large language models are reasoning teachers. Preprint, arXiv:2212.10071.
  5. Shirley Kokane, Ming Zhu, Tulika Awalgaonkar, Jianguo Zhang, Thai Hoang, Akshara Prabhakar, Zuxin Liu, Tian Lan, Liangwei Yang, Juntao Tan, Rithesh Murthy, Weiran Yao, Zhiwei Liu, Juan Carlos Niebles, Huan Wang, Shelby Heinecke, Caiming Xiong, and Silivo Savarese. 2025. Toolscan: A benchmark for cha… — [arXiv:2411.13547](https://arxiv.org/abs/2411.13547) find
  6. Jiarui Lu, Thomas Holleis, Yizhe Zhang, Bernhard Aumayer, Feng Nan, Felix Bai, Shuang Ma, Shen Ma, Mengyu Li, Guoli Yin, Zirui Wang, and Ruoming Pang. 2025. Toolsandbox: A stateful, conversational, interactive evaluation benchmark for llm tool use capabilities. Preprint, arXiv:2408.04682.
  7. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. 2023. Self-refine: Iterative refinement with… find
  8. OpenAI. 2024. Learning to reason with llms. Accessed: 2025-02-01. find
  9. Krista Opsahl-Ong, Michael J Ryan, Josh Purtell, David Broman, Christopher Potts, Matei Zaharia, and Omar Khattab. 2024. Optimizing instructions and demonstrations for multi-stage language model programs. Preprint, arXiv:2406.11695.
  10. Example execution: User Query: “What’s the weather in New York City?” Loop 1: • GPT-4o generates: get_weather(location="NYC", temp_unit="celsius") find
  11. Shishir G. Patil, Tianjun Zhang, Xin Wang, and Joseph E. Gonzalez. 2023. Gorilla: Large language model connected with massive apis. Preprint, arXiv:2305.15334.
  12. o3-mini reviews: “Error: For US cities, temperature should use Fahrenheit by default.” Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. 2024. Berkeley function calling leaderboard. https://gorilla.cs.berkeley.edu/leaderboard.html. ↗
  13. Loop 2: • GPT-4o revises with get_weather(location="New temp_unit="fahrenheit") find
  14. Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. Preprint, arXiv:2303.11366.
  15. Yunhao Tang, Zhaohan Daniel Guo, Zeyu Zheng, Daniele Calandriello, Rémi Munos, Mark Rowland, Pierre Harvey Richemond, Michal Valko, Bernardo Ávila Pires, and Bilal Piot. 2024. Generalized preference optimization: A unified approach to offline alignment. Preprint, arXiv:2402.05749. — [arXiv:2402.05749](https://arxiv.org/abs/2402.05749) find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ta-2026-reinforced-agent-inference-time-feedback/