pith. sign in

arxiv: 2606.21409 · v1 · pith:6C3IVU7Inew · submitted 2026-06-19 · 💻 cs.AI

Don't Blindly Trust It: How Unreliable Feedback Breaks Tool-Using LLM Agents

Chubin Zhang , Zhenglin Wan , Xingrui Yu , Pengfei Zhou , Wangbo Zhao , Jingxuan Wu , Yaxin Zhou , Ivor Tsang This is my paper

Pith reviewed 2026-06-26 14:20 UTC · model grok-4.3

classification 💻 cs.AI
keywords tool-augmented agentsunreliable feedbackvalue inversionLLM agentsquestion answeringfact verificationno-feedback baseline
0
0 comments X

The pith

Persistent misleading feedback makes tool-using LLM agents perform worse than receiving no feedback at all.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Tool-augmented LLM agents are normally evaluated only under reliable external feedback, leaving open whether they would do better with no task evidence when feedback is unreliable. The paper runs a matched-loop experiment that holds the agent loop, prompt, action space, and decoding fixed while varying only the returned observation as faithful, misleading, or absent. Across question answering and fact verification tasks, persistently misleading feedback creates a value inversion in which agents that improve with clean tools fall below the no-feedback baseline. On HotpotQA, Qwen2.5-7B scores 44.8 F1 with clean retrieval and 22.3 F1 with no feedback, but only 4.7 F1 under shuffled retrieval. The inversion survives stronger clean retrieval and locally plausible distractors, yet later clean evidence can sometimes repair trajectories.

Core claim

When feedback is persistently misleading, agents that benefit from clean tools can perform worse than the matched no-feedback fallback. This inversion appears across question answering and fact verification, persists under stronger clean retrieval and plausible distractors, and weakens only when later clean evidence can repair the trajectory. Early trajectory signals predict many failures, but simple repairs such as rejecting bad evidence help only when the exposed no-feedback fallback is itself reliable.

What carries the argument

The matched-loop comparison, which fixes the agent loop, prompt, action space, and decoding while varying only the returned observation between faithful, misleading, or absent.

If this is right

  • Clean-tool gains can overstate tool value when unreliable feedback is possible.
  • Matched no-feedback fallback controls are required to evaluate tool-augmented agents.
  • Early trajectory signals predict many failures caused by bad feedback.
  • Rejecting bad evidence helps only when the no-feedback fallback is reliable.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Agent designs could add early detection of unreliable observations to avoid locking into bad trajectories.
  • Existing tool-use benchmarks should add unreliable-feedback conditions to avoid overstating agent capabilities.
  • The inversion finding may apply to other agent architectures or domains beyond the tested QA and verification tasks.
  • Re-evaluating prior tool-augmented agent results that lacked no-feedback controls could change assessments of their practical value.

Load-bearing premise

The matched-loop comparison fully isolates the causal effect of feedback reliability without confounds from prompt sensitivity or decoding variance.

What would settle it

A replication of the matched-loop experiment on the same tasks and models in which misleading feedback no longer produces performance below the no-feedback baseline.

read the original abstract

Tool-augmented agents are typically evaluated by their gains under reliable external feedback. Yet these gains leave open a key counterfactual: when feedback is unreliable, would the agent be better off receiving no task evidence? We study this question with a controlled matched-loop comparison that fixes the agent loop, prompt, action space, and decoding, while varying only the returned observation: faithful, misleading, or absent. Across question answering and fact verification, persistent misleading feedback produces a value inversion: agents that benefit from clean tools can perform worse than the matched no-feedback fallback. On HotpotQA, Qwen2.5-7B reaches 44.8 F1 with clean retrieval and 22.3 F1 with no feedback, but drops to 4.7 F1 under shuffled retrieval. The inversion persists under stronger clean retrieval and locally plausible distractors, but weakens when later clean evidence can repair the trajectory. Early trajectory signals predict many failures, yet simple repairs remain fallback-limited: rejecting bad evidence helps only when the exposed fallback is reliable. These results show that clean-tool gains can overstate tool value, and that matched no-feedback fallback controls are necessary for evaluating tool-augmented agents.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript claims that tool-augmented LLM agents exhibit a value inversion under persistent unreliable feedback: agents that improve with clean tool observations can perform worse than a matched no-feedback baseline. This is demonstrated via controlled experiments that fix the agent loop, prompt, action space, and decoding while varying only the returned observation (faithful, misleading via shuffled retrieval, or absent) on HotpotQA and fact-verification tasks. Key result: Qwen2.5-7B reaches 44.8 F1 with clean retrieval, 22.3 F1 with no feedback, and 4.7 F1 with shuffled retrieval. The inversion holds under stronger retrieval and plausible distractors but weakens with repairable trajectories; early signals predict failures, yet simple repairs are fallback-limited.

Significance. If the results hold, the work is significant for agent evaluation practices because it shows that clean-tool gains can overstate tool utility and that no-feedback controls are required to avoid misleading assessments. The matched-loop design isolates the causal role of feedback reliability without confounding prompt or decoding changes, providing a reproducible template for future studies. The empirical demonstration of inversion on concrete tasks adds falsifiable evidence to discussions of tool reliability in LLM agents.

major comments (2)
  1. [Experiments] Experiments section (HotpotQA results): the central value-inversion claim rests on the F1 scores (44.8 clean vs. 22.3 no-feedback vs. 4.7 shuffled), yet the manuscript does not report the number of runs, variance, or statistical significance tests; without these the magnitude and reliability of the inversion cannot be assessed.
  2. [Experimental setup] Experimental setup: the construction of shuffled retrieval (how distractors are sampled and whether they remain locally plausible) is load-bearing for the claim that misleading feedback is realistic; the current description leaves open whether the 4.7 F1 drop is an artifact of implausible noise rather than representative unreliability.
minor comments (2)
  1. [Introduction] The introduction should explicitly define 'value inversion' with a short formal statement before the empirical results.
  2. [Figures] Figure captions for trajectory plots should include the exact agent model and task split used.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment, recognition of the matched-loop design's value, and recommendation for minor revision. The comments identify areas where additional rigor and detail will strengthen the manuscript. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Experiments] Experiments section (HotpotQA results): the central value-inversion claim rests on the F1 scores (44.8 clean vs. 22.3 no-feedback vs. 4.7 shuffled), yet the manuscript does not report the number of runs, variance, or statistical significance tests; without these the magnitude and reliability of the inversion cannot be assessed.

    Authors: We agree that reporting the number of runs, variance, and statistical significance is necessary to substantiate the reliability of the reported F1 scores. The current manuscript presents point estimates from single executions. In the revised version we will add results aggregated over multiple independent runs with different random seeds, include standard deviations, and report statistical significance tests (e.g., paired t-tests or Wilcoxon tests) comparing the clean, no-feedback, and shuffled conditions. These additions will allow readers to evaluate the stability of the value inversion. revision: yes

  2. Referee: [Experimental setup] Experimental setup: the construction of shuffled retrieval (how distractors are sampled and whether they remain locally plausible) is load-bearing for the claim that misleading feedback is realistic; the current description leaves open whether the 4.7 F1 drop is an artifact of implausible noise rather than representative unreliability.

    Authors: We acknowledge that a more explicit description of distractor sampling is required. The manuscript already states that the inversion persists under "locally plausible distractors," but the precise sampling procedure is not fully detailed. We will expand the experimental setup section to specify how distractors are drawn (random selection from the same retrieval corpus, conditioned on topical overlap with the query to preserve local plausibility) and will include an example of a shuffled observation. This will clarify that the misleading feedback reflects realistic retrieval errors rather than arbitrary or implausible noise. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper is an empirical study that reports direct experimental measurements (F1 scores on HotpotQA and other tasks) obtained via a matched-loop comparison that holds agent loop, prompt, action space, and decoding fixed while varying only the observation. No equations, fitted parameters, or derivations appear in the manuscript. The central claim of value inversion follows immediately from the controlled measurements without any reduction to self-defined quantities, self-citation chains, or ansatzes. The design is self-contained and externally falsifiable.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The study is purely empirical and introduces no new free parameters, axioms beyond standard ML evaluation practices, or invented entities.

axioms (1)
  • domain assumption F1 score is a valid metric for measuring performance on HotpotQA and fact verification
    Standard practice in the QA literature; invoked implicitly when reporting F1 numbers.

pith-pipeline@v0.9.1-grok · 5767 in / 1186 out tokens · 24700 ms · 2026-06-26T14:20:07.961156+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 12 canonical work pages · 1 internal anchor

  1. [1]

    2022 , url =

    Karpas, Ehud and Abend, Omri and Belinkov, Yonatan and Lenz, Barak and Lieber, Opher and Ratner, Nir and Shoham, Yoav and Bata, Hofit and Levine, Yoav and Leyton-Brown, Kevin and Muhlgay, Dor and Rozen, Noam and Schwartz, Erez and Shachaf, Gal and Shalev-Shwartz, Shai and Shashua, Amnon and Tenenholtz, Moshe , journal =. 2022 , url =

    2022

  2. [2]

    2021 , url =

    Nakano, Reiichiro and Hilton, Jacob and Balaji, Suchir and Wu, Jeff and Ouyang, Long and Kim, Christina and Hesse, Christopher and Jain, Shantanu and Kosaraju, Vineet and Saunders, William and Jiang, Xu and Cobbe, Karl and Eloundou, Tyna and Krueger, Gretchen and Button, Kevin and Knight, Matthew and Chess, Benjamin and Schulman, John , journal =. 2021 , url =

    2021

  3. [3]

    and Cao, Yuan , booktitle =

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik R. and Cao, Yuan , booktitle =. 2023 , url =

    2023

  4. [4]

    Advances in Neural Information Processing Systems , volume =

    Schick, Timo and Dwivedi-Yu, Jane and Dess. Advances in Neural Information Processing Systems , volume =. 2023 , url =

    2023

  5. [6]

    First Conference on Language Modeling , year =

    How Easily do Irrelevant Inputs Skew the Responses of Large Language Models? , author =. First Conference on Language Modeling , year =

  6. [8]

    2026 , url =

    Wang, Ruipeng and Chen, Yuxin and Wang, Yukai and Wu, Chang and Fang, Junfeng and Cai, Xiaodong and Gu, Qi and Su, Hui and Zhang, An and Wang, Xiang and Cai, Xunliang and Chua, Tat-Seng , journal =. 2026 , url =

    2026

  7. [12]

    The Twelfth International Conference on Learning Representations , year =

    Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts , author =. The Twelfth International Conference on Learning Representations , year =

  8. [13]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Benchmarking Large Language Models in Retrieval-Augmented Generation , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2024 , doi =

    2024

  9. [18]

    The Twelfth International Conference on Learning Representations , year =

    Large Language Models Cannot Self-Correct Reasoning Yet , author =. The Twelfth International Conference on Learning Representations , year =

  10. [22]

    The Instruction Hierarchy: Training

    Wallace, Eric and Xiao, Kai and Leike, Reimar and Weng, Lilian and Heidecke, Johannes and Beutel, Alex , journal =. The Instruction Hierarchy: Training. 2024 , url =

    2024

  11. [23]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems , volume =. 2020 , url =

    2020

  12. [24]

    2020 , series =

    Guu, Kelvin and Lee, Kenton and Tung, Zora and Pasupat, Panupong and Chang, Ming-Wei , booktitle =. 2020 , series =

    2020

  13. [25]

    Proceedings of the 40th International Conference on Machine Learning , pages =

    Large Language Models Can Be Easily Distracted by Irrelevant Context , author =. Proceedings of the 40th International Conference on Machine Learning , pages =. 2023 , series =

    2023

  14. [27]

    The Twelfth International Conference on Learning Representations , year =

    Making Retrieval-Augmented Language Models Robust to Irrelevant Context , author =. The Twelfth International Conference on Learning Representations , year =

  15. [28]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  16. [29]

    Publications Manual , year = "1983", publisher =

    1983

  17. [30]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    work page doi:10.1145/322234.322243 1981

  18. [31]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  19. [32]

    Dan Gusfield , title =. 1997

    1997

  20. [33]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  21. [34]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  22. [35]

    Benchmarking large language models in retrieval-augmented generation

    Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun. Benchmarking large language models in retrieval-augmented generation. Proceedings of the AAAI Conference on Artificial Intelligence, 38 0 (16): 0 17754--17762, 2024. doi:10.1609/aaai.v38i16.29728. URL https://ojs.aaai.org/index.php/AAAI/article/view/29728

    work page doi:10.1609/aaai.v38i16.29728 2024

  23. [36]

    Training verifiers to solve math word problems

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. URL https://arxiv.org/abs/2110.14168

    Pith/arXiv arXiv 2021

  24. [37]

    Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =

    Florin Cuconasu, Giovanni Trappolini, Federico Siciliano, Simone Filice, Cesare Campagnano, Yoelle Maarek, Nicola Tonellotto, and Fabrizio Silvestri. The power of noise: Redefining retrieval for RAG systems. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR '24, pages 719--729. Associ...

    work page doi:10.1145/3626772.3657834 2024

  25. [38]

    Not what you’ve signed up for: Compromising real- world LLM-integrated applications with indirect prompt injection,

    Kai Greshake, Sahar Abdelnabi, Shailesh Mishra, Christoph Endres, Thorsten Holz, and Mario Fritz. Not what you've signed up for: Compromising real-world LLM -integrated applications with indirect prompt injection. In Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security, AISec '23, pages 79--90, New York, NY, USA, 2023. Association ...

    work page doi:10.1145/3605764.3623985 2023

  26. [39]

    REALM : Retrieval augmented language model pre-training

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. REALM : Retrieval augmented language model pre-training. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 3929--3938. PMLR, 2020. URL https://proceedings.mlr.press/v119/guu20a.html

    2020

  27. [40]

    Large language models cannot self-correct reasoning yet

    Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. Large language models cannot self-correct reasoning yet. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=IkmD3fKBPQ

    2024

  28. [41]

    MRKL systems: A modular, neuro-symbolic architecture that combines large language models, external knowledge sources and discrete reasoning

    Ehud Karpas, Omri Abend, Yonatan Belinkov, Barak Lenz, Opher Lieber, Nir Ratner, Yoav Shoham, Hofit Bata, Yoav Levine, Kevin Leyton-Brown, Dor Muhlgay, Noam Rozen, Erez Schwartz, Gal Shachaf, Shai Shalev-Shwartz, Amnon Shashua, and Moshe Tenenholtz. MRKL systems: A modular, neuro-symbolic architecture that combines large language models, external knowledg...

    Pith/arXiv arXiv 2022

  29. [42]

    ToolHaystack : Stress-testing tool-augmented language models in realistic long-term interactions

    Beong-woo Kwak, Minju Kim, Dongha Lim, Hyungjoo Chae, Dongjin Kang, Sunghwan Kim, Dongil Yang, and Jinyoung Yeo. ToolHaystack : Stress-testing tool-augmented language models in realistic long-term interactions. In Findings of the Association for Computational Linguistics: EMNLP 2025, pages 24696--24727, Suzhou, China, 2025. Association for Computational L...

    work page doi:10.18653/v1/2025.findings-emnlp.1344 2025

  30. [43]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems, volume 33, pages 9459--9474, 2020. URL ...

    2020

  31. [44]

    When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories

    Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Daniel Khashabi, and Hannaneh Hajishirzi. When not to trust language models: Investigating effectiveness of parametric and non-parametric memories. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 9802--9822, Toronto, Canada, 2023....

    work page doi:10.18653/v1/2023.acl-long.546 2023

  32. [45]

    WebGPT : Browser-assisted question-answering with human feedback

    Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, Xu Jiang, Karl Cobbe, Tyna Eloundou, Gretchen Krueger, Kevin Button, Matthew Knight, Benjamin Chess, and John Schulman. WebGPT : Browser-assisted question-answering with human feedback. arXiv preprint arX...

    Pith/arXiv arXiv 2021

  33. [46]

    Toolformer : Language models can teach themselves to use tools

    Timo Schick, Jane Dwivedi-Yu, Roberto Dess \`i , Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer : Language models can teach themselves to use tools. In Advances in Neural Information Processing Systems, volume 36, 2023. URL https://proceedings.neurips.cc/paper_files/paper/2023/hash/d842425e4b...

    2023

  34. [47]

    The synthetic web: Adversarially-curated mini-internets for diagnosing epistemic weaknesses of language agents

    Shrey Shah and Levent Ozgur. The synthetic web: Adversarially-curated mini-internets for diagnosing epistemic weaknesses of language agents. arXiv preprint arXiv:2603.00801, 2026. URL https://arxiv.org/abs/2603.00801

    arXiv 2026

  35. [48]

    Chi, Nathanael Sch \"a rli, and Denny Zhou

    Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H. Chi, Nathanael Sch \"a rli, and Denny Zhou. Large language models can be easily distracted by irrelevant context. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 31210--31227. PMLR, 2023. URL http...

    2023

  36. [49]

    FEVER: a large-scale dataset for Fact Extraction and VERification

    James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. FEVER : A large-scale dataset for fact extraction and VER ification. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 809--819, New Orleans, Louisiana,...

    work page internal anchor Pith review doi:10.18653/v1/n18-1074 2018

  37. [50]

    LLM s cannot find reasoning errors, but can correct them given the error location

    Gladys Tyen, Hassan Mansoor, Victor Carbune, Peter Chen, and Tony Mak. LLM s cannot find reasoning errors, but can correct them given the error location. In Findings of the Association for Computational Linguistics: ACL 2024, pages 13894--13908, Bangkok, Thailand, 2024. Association for Computational Linguistics. doi:10.18653/v1/2024.findings-acl.826. URL ...

    work page doi:10.18653/v1/2024.findings-acl.826 2024

  38. [51]

    The instruction hierarchy: Training LLM s to prioritize privileged instructions

    Eric Wallace, Kai Xiao, Reimar Leike, Lilian Weng, Johannes Heidecke, and Alex Beutel. The instruction hierarchy: Training LLM s to prioritize privileged instructions. arXiv preprint arXiv:2404.13208, 2024. URL https://arxiv.org/abs/2404.13208

    Pith/arXiv arXiv 2024

  39. [52]

    AgentNoiseBench : Benchmarking robustness of tool-using LLM agents under noisy condition

    Ruipeng Wang, Yuxin Chen, Yukai Wang, Chang Wu, Junfeng Fang, Xiaodong Cai, Qi Gu, Hui Su, An Zhang, Xiang Wang, Xunliang Cai, and Tat-Seng Chua. AgentNoiseBench : Benchmarking robustness of tool-using LLM agents under noisy condition. arXiv preprint arXiv:2602.11348, 2026. URL https://arxiv.org/abs/2602.11348

    arXiv 2026

  40. [53]

    How easily do irrelevant inputs skew the responses of large language models? In First Conference on Language Modeling, 2024

    Siye Wu, Jian Xie, Jiangjie Chen, Tinghui Zhu, Kai Zhang, and Yanghua Xiao. How easily do irrelevant inputs skew the responses of large language models? In First Conference on Language Modeling, 2024. URL https://openreview.net/forum?id=S7NVVfuRv8

    2024

  41. [54]

    Adaptive chameleon or stubborn sloth: Revealing the behavior of large language models in knowledge conflicts

    Jian Xie, Kai Zhang, Jiangjie Chen, Renze Lou, and Yu Su. Adaptive chameleon or stubborn sloth: Revealing the behavior of large language models in knowledge conflicts. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=auKAUJZMO6

    2024

  42. [55]

    The confidence dichotomy: Analyzing and mitigating miscalibration in tool-use agents

    Weihao Xuan, Qingcheng Zeng, Heli Qi, Yunze Xiao, Junjue Wang, and Naoto Yokoya. The confidence dichotomy: Analyzing and mitigating miscalibration in tool-use agents. arXiv preprint arXiv:2601.07264, 2026. URL https://arxiv.org/abs/2601.07264

    arXiv 2026

  43. [56]

    H otpot QA : A Dataset for Diverse, Explainable Multi-hop Question Answering

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. HotpotQA : A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2369--2380, Brussels, Belgium, 2018. Association for Computatio...

    work page doi:10.18653/v1/d18-1259 2018

  44. [57]

    Narasimhan, and Yuan Cao

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. ReAct : Synergizing reasoning and acting in language models. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=WE_vluYUL-X

    2023

  45. [58]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1 , pages =

    Jingwei Yi, Yueqi Xie, Bin Zhu, Emre Kiciman, Guangzhong Sun, Xing Xie, and Fangzhao Wu. Benchmarking and defending against indirect prompt injection attacks on large language models. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1, KDD '25, pages 1809--1820. Association for Computing Machinery, 2025. doi:10.114...

    work page doi:10.1145/3690624.3709179 2025

  46. [59]

    Making retrieval-augmented language models robust to irrelevant context

    Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan Berant. Making retrieval-augmented language models robust to irrelevant context. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=ZS4m74kZpH

    2024

  47. [60]

    Chain-of-note: Enhancing robustness in retrieval-augmented language models

    Wenhao Yu, Hongming Zhang, Xiaoman Pan, Peixin Cao, Kaixin Ma, Jian Li, Hongwei Wang, and Dong Yu. Chain-of-note: Enhancing robustness in retrieval-augmented language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 14672--14685, Miami, Florida, USA, 2024. Association for Computational Linguistics. d...

    work page doi:10.18653/v1/2024.emnlp-main.813 2024

  48. [61]

    doi:10.18653/v1/2024.findings-acl.624

    Qiusi Zhan, Zhixiang Liang, Zifan Ying, and Daniel Kang. InjecAgent : Benchmarking indirect prompt injections in tool-integrated large language model agents. In Findings of the Association for Computational Linguistics: ACL 2024, pages 10471--10506, Bangkok, Thailand, 2024. Association for Computational Linguistics. doi:10.18653/v1/2024.findings-acl.624. ...

    work page doi:10.18653/v1/2024.findings-acl.624 2024