pith. sign in

arxiv: 2606.20662 · v1 · pith:37ZCEIGWnew · submitted 2026-06-09 · 💻 cs.AI · cs.MA

Confidence Laundering in Agent Systems: Why Uncertainty Needs a Latent Carrier

Kaiwen Shi , Zheyuan Zhang , Han Bao , Colby Nelson , Yanfang Ye This is my paper

Pith reviewed 2026-06-27 12:51 UTC · model grok-4.3

classification 💻 cs.AI cs.MA
keywords agent systemsuncertainty propagationconfidence launderinglatent uncertaintydecision handoffsinterface designerror amplificationAI agents
0
0 comments X

The pith

Agent systems lose uncertainty at component handoffs, turning local fragility into downstream overconfidence.

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

Modern agent systems present fragile upstream decisions to later components as clean intermediate results, so that uncertainty disappears at the transfer points. The paper claims this interface loss, not the mere presence of uncertain steps, is what allows local ambiguity to amplify into system errors. It defines uncertain decision handoff as the transfer of an intermediate decision made under uncertainty and names the resulting failure mode confidence laundering. To keep fragility usable, the authors propose attaching a latent uncertainty carrier directly to those handoffs. The position moves the problem from estimating uncertainty at each step to designing interfaces that preserve it for downstream components.

Core claim

Uncertainty does not propagate through an agent trajectory simply because some steps are uncertain; it propagates only when it survives the handoff between components. Fragile upstream states are often repackaged as procedurally valid artifacts that downstream agents over-trust, a failure the authors call confidence laundering. They propose latent uncertainty as an uncertainty-bearing carrier attached to decision handoffs that preserves pre-commitment fragility in a form downstream components can use without replacing text outputs with hidden states.

What carries the argument

latent uncertainty, an uncertainty-bearing carrier attached to decision handoffs that preserves upstream fragility for downstream use

If this is right

  • Uncertainty propagates only through surviving handoffs, not through the mere existence of uncertain steps in a trajectory.
  • Local ambiguity can become system-level error amplification when fragile states are presented as clean artifacts.
  • Attaching a latent uncertainty carrier to handoffs preserves pre-commitment fragility for downstream components.
  • Agent uncertainty handling should shift from step-wise estimation to uncertainty-preserving interface design.

Where Pith is reading between the lines

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

  • Existing modular agent frameworks could be extended with explicit handoff protocols that carry uncertainty markers alongside text outputs.
  • The same interface bottleneck may appear in human-AI collaboration when intermediate outputs hide the uncertainty behind them.
  • Testing whether current agent benchmarks already contain hidden confidence laundering would require logging uncertainty at each handoff point.
  • Recoverable agent systems might need standardized carrier formats so that different component types can interpret upstream fragility without custom engineering.

Load-bearing premise

Downstream components will actually read and act on a latent uncertainty carrier attached to handoffs instead of ignoring it or demanding replacement of text outputs.

What would settle it

A controlled comparison of agent chains with and without explicit latent uncertainty carriers at handoffs, measuring whether the carrier version shows measurably lower rates of error amplification from upstream uncertain decisions.

Figures

Figures reproduced from arXiv: 2606.20662 by Colby Nelson, Han Bao, Kaiwen Shi, Yanfang Ye, Zheyuan Zhang.

Figure 1
Figure 1. Figure 1: Uncertainty Quantification Methods: from LLMs era to Agent era. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of uncertainty interface collapse and uncertainty-preserving handoff. Without an [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Interface uncertainty loss with￾out an uncertainty carrier. The many-to-one projection described above creates an in￾terface bottleneck in agent uncertainty propagation. Once an uncertainty-bearing state is compressed into a commit￾ted artifact, uncertainty may still influence the trajectory, but it no longer travels as an accessible state. It survives only indirectly, through weak evidence, brittle plans,… view at source ↗
Figure 4
Figure 4. Figure 4: Impact of uncertainty signals on overconfidence across different language models. For [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of information loss across different uncertainty carriers. Scalar confidence [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Empirical evidence for latent uncertainty as an uncertainty carrier. (a) Semantic cluster [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
read the original abstract

Modern agent systems can turn uncertainty into overconfidence. Fragile upstream decisions are often exposed to downstream components as clean intermediate artifacts, while the uncertainty behind those decisions is lost at the interface. As a result, local ambiguity can become system-level error amplification. We argue that this reveals an interface bottleneck in agent uncertainty propagation: uncertainty does not propagate simply because a trajectory contains uncertain steps; it propagates only when it survives the handoff between components. We define uncertain decision handoff as the transfer of an intermediate decision made under uncertainty, and identify confidence laundering as a failure mode in which fragile upstream states are repackaged as procedurally valid artifacts that downstream agents over-trust. To address this bottleneck, we propose latent uncertainty as an uncertainty-bearing carrier attached to decision handoffs. Rather than replacing text with hidden states, latent uncertainty aims to preserve pre-commitment fragility in a form that downstream components can use. This position shifts agent uncertainty propagation from step-wise uncertainty estimation toward uncertainty-preserving interface design for more recoverable agent systems.

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 is a position paper arguing that agent systems lose uncertainty at component interfaces, causing 'confidence laundering' where fragile upstream decisions are repackaged as clean artifacts that downstream agents over-trust. It defines 'uncertain decision handoff' and claims that uncertainty propagates only when it survives handoffs (not merely because uncertain steps exist). To address this, it proposes 'latent uncertainty' as a carrier attached to handoffs that preserves pre-commitment fragility in a usable form without replacing text outputs with hidden states, shifting focus from step-wise estimation to uncertainty-preserving interface design.

Significance. If the latent uncertainty carrier can be made operational, the interface-bottleneck diagnosis could meaningfully inform design of recoverable multi-agent systems by highlighting a propagation failure mode distinct from local uncertainty estimation. The paper offers no empirical support, derivations, or examples, so significance remains prospective and depends on future formalization.

major comments (2)
  1. [Abstract] Abstract: the claim that 'uncertainty does not propagate simply because a trajectory contains uncertain steps; it propagates only when it survives the handoff' is presented as the central diagnosis but receives no supporting argument, example trajectory, or reduction to existing agent architectures, rendering the interface-bottleneck thesis unevaluable as stated.
  2. [Abstract] Abstract: the proposed solution ('latent uncertainty as an uncertainty-bearing carrier attached to decision handoffs' that 'downstream components can use' without hidden states) supplies neither attachment protocol, syntax, consumption rule, nor pseudocode; this leaves the proposal at the level of a negative definition and does not demonstrate that the carrier would survive typical text/API interfaces or avoid being ignored.
minor comments (2)
  1. The abstract introduces 'confidence laundering' and 'uncertain decision handoff' as new terms; explicit comparison to related concepts such as epistemic uncertainty propagation or belief tracking in POMDPs would clarify novelty.
  2. The manuscript is short and contains no figures, tables, or worked examples; adding even one concrete agent pipeline (e.g., planner-to-executor handoff) would make the laundering failure mode concrete.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive comments on our position paper. We address each major comment below and indicate the revisions we will make to strengthen the manuscript while preserving its conceptual focus.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that 'uncertainty does not propagate simply because a trajectory contains uncertain steps; it propagates only when it survives the handoff' is presented as the central diagnosis but receives no supporting argument, example trajectory, or reduction to existing agent architectures, rendering the interface-bottleneck thesis unevaluable as stated.

    Authors: We agree that the central claim would be more readily evaluable with concrete support. In revision we will insert a short illustrative trajectory (a planning agent producing a fragile intermediate plan that is passed as a clean artifact to an execution agent) together with a brief reduction to standard tool-calling and multi-agent orchestration patterns. This addition will clarify the handoff failure mode without converting the paper into an empirical study. revision: yes

  2. Referee: [Abstract] Abstract: the proposed solution ('latent uncertainty as an uncertainty-bearing carrier attached to decision handoffs' that 'downstream components can use' without hidden states) supplies neither attachment protocol, syntax, consumption rule, nor pseudocode; this leaves the proposal at the level of a negative definition and does not demonstrate that the carrier would survive typical text/API interfaces or avoid being ignored.

    Authors: As a position paper the proposal is deliberately conceptual, emphasizing the shift toward interface design rather than a ready-to-implement mechanism. Nevertheless, we accept that a high-level attachment protocol and consumption rules would make the idea more tangible. In the revised manuscript we will add a concise description of an attachment protocol (including how a compact uncertainty tag can be appended to text outputs or API responses) together with pseudocode sketches showing consumption by a downstream component. These additions will illustrate survival across typical text interfaces while remaining within the scope of a position paper. revision: partial

Circularity Check

0 steps flagged

No circularity: conceptual position paper with definitional claims only

full rationale

The manuscript is a position paper that introduces terminology ('uncertain decision handoff', 'confidence laundering', 'latent uncertainty') and argues for an interface-level design shift. No equations, parameter fits, or derivations appear in the provided text. No self-citations are used to justify uniqueness theorems or to close a derivation chain. The central claim—that uncertainty propagates only when it survives handoffs—is presented as an observation leading to a proposed carrier concept, not as a quantity derived from or equivalent to its own inputs. The proposal remains at the level of high-level definition without reducing to a fitted input renamed as prediction or any of the enumerated circular patterns. This is the expected non-finding for a non-mathematical conceptual work.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Review based solely on abstract; full text unavailable. The ledger reflects the conceptual nature with minimal free parameters but one invented entity and a domain assumption.

axioms (1)
  • domain assumption Uncertainty is lost at component interfaces in agent systems, turning local ambiguity into system-level error amplification.
    Core premise stated directly in the abstract as the identified bottleneck.
invented entities (1)
  • latent uncertainty no independent evidence
    purpose: An uncertainty-bearing carrier attached to decision handoffs to preserve pre-commitment fragility.
    Newly proposed construct intended to address the interface bottleneck; no independent evidence supplied.

pith-pipeline@v0.9.1-grok · 5712 in / 1148 out tokens · 20511 ms · 2026-06-27T12:51:32.214514+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

68 extracted references · 16 linked inside Pith

  1. [1]

    Llms4all: A review of large language models across academic disciplines.arXiv preprint arXiv:2509.19580, 2025

    Yanfang Ye, Zheyuan Zhang, Tianyi Ma, Zehong Wang, Yiyang Li, Shifu Hou, Weixiang Sun, Kaiwen Shi, Yijun Ma, Wei Song, et al. Llms4all: A review of large language models across academic disciplines.arXiv preprint arXiv:2509.19580, 2025

    arXiv 2025

  2. [2]

    The obvious invisible threat: Llm-powered gui agents’ vulnerability to fine-print injections

    Chaoran Chen, Zhiping Zhang, Bingcan Guo, Shang Ma, Ibrahim Khalilov, Simret Gebreegzi- abher, Yanfang Ye, Ziang Xiao, Yaxing Yao, Tianshi Li, et al. The obvious invisible threat: Llm-powered gui agents’ vulnerability to fine-print injections. InSoups. The Twenty-First Symposium on Usable Privacy and Security (SOUPS), 2025

    2025

  3. [3]

    Clear: Towards contextual llm-empowered privacy policy analysis and risk generation for large language model applications

    Chaoran Chen, Daodao Zhou, Yanfang Ye, Toby Jia-jun Li, and Yaxing Yao. Clear: Towards contextual llm-empowered privacy policy analysis and risk generation for large language model applications. InProceedings of the 30th International Conference on Intelligent User Interfaces, pages 277–297, 2025

    2025

  4. [4]

    Agentic confidence calibration.arXiv preprint arXiv:2601.15778, 2026

    Jiaxin Zhang, Caiming Xiong, and Chien-Sheng Wu. Agentic confidence calibration.arXiv preprint arXiv:2601.15778, 2026

    arXiv 2026

  5. [5]

    Toolformer: Language models can teach them- selves to use tools.ArXiv, abs/2302.04761, 2023

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettle- moyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach them- selves to use tools.ArXiv, abs/2302.04761, 2023. URL https://api.semanticscholar. org/CorpusID:256697342

    Pith/arXiv arXiv 2023

  6. [6]

    Uncertainty quantification in llm agents: Foundations, emerging challenges, and opportunities

    Changdae Oh, Seongheon Park, To Eun Kim, Jiatong Li, Wendi Li, Samuel Yeh, Xuefeng Du, Hamed Hassani, Paul Bogdan, Dawn Song, et al. Uncertainty quantification in llm agents: Foundations, emerging challenges, and opportunities. InAgentic AI in the Wild: From Hallucinations to Reliable Autonomy

  7. [7]

    Agentrouter: A knowledge-graph-guided llm router for collaborative multi-agent question answering.arXiv preprint arXiv:2510.05445, 2025

    Zheyuan Zhang, Kaiwen Shi, Zhengqing Yuan, Zehong Wang, Tianyi Ma, Keerthiram Muruge- san, Vincent Galassi, Chuxu Zhang, and Yanfang Ye. Agentrouter: A knowledge-graph-guided llm router for collaborative multi-agent question answering.arXiv preprint arXiv:2510.05445, 2025

    arXiv 2025

  8. [8]

    Uprop: Investigating the uncertainty propagation of llms in multi-step agentic decision-making.arXiv preprint arXiv:2506.17419, 2025

    Jinhao Duan, James Diffenderfer, Sandeep Madireddy, Tianlong Chen, Bhavya Kailkhura, and Kaidi Xu. Uprop: Investigating the uncertainty propagation of llms in multi-step agentic decision-making.arXiv preprint arXiv:2506.17419, 2025

    arXiv 2025

  9. [9]

    Saup: Situation awareness uncertainty propagation on llm agent.arXiv preprint arXiv:2412.01033, 2024

    Qiwei Zhao, Xujiang Zhao, Yanchi Liu, Wei Cheng, Yiyou Sun, Mika Oishi, Takao Osaki, Katsushi Matsuda, Huaxiu Yao, and Haifeng Chen. Saup: Situation awareness uncertainty propagation on llm agent.arXiv preprint arXiv:2412.01033, 2024

    arXiv 2024

  10. [10]

    How language model hallucinations can snowball.arXiv preprint arXiv:2305.13534, 2023

    Muru Zhang, Ofir Press, William Merrill, Alisa Liu, and Noah A Smith. How language model hallucinations can snowball.arXiv preprint arXiv:2305.13534, 2023

    arXiv 2023

  11. [11]

    Where llm agents fail and how they can learn from failures.arXiv preprint arXiv:2509.25370, 2025

    Kunlun Zhu, Zijia Liu, Bingxuan Li, Muxin Tian, Yingxuan Yang, Jiaxun Zhang, Pengrui Han, Qipeng Xie, Fuyang Cui, Weijia Zhang, et al. Where llm agents fail and how they can learn from failures.arXiv preprint arXiv:2509.25370, 2025

    arXiv 2025

  12. [12]

    From spark to fire: Can situational reading interest lead to long-term reading motivation? Literacy Research and Instruction, 45(2):91–117, 2005

    John T Guthrie, Laurel W Hoa, Allan Wigfield, Stephen M Tonks, and Kathleen C Perencevich. From spark to fire: Can situational reading interest lead to long-term reading motivation? Literacy Research and Instruction, 45(2):91–117, 2005

    2005

  13. [13]

    Graphtracer: Graph-guided failure tracing in llm agents for robust multi-turn deep search.arXiv preprint arXiv:2510.10581, 2025

    Heng Zhang, Yuling Shi, Xiaodong Gu, Haochen You, Zijian Zhang, Lubin Gan, Yilei Yuan, and Jin Huang. Graphtracer: Graph-guided failure tracing in llm agents for robust multi-turn deep search.arXiv preprint arXiv:2510.10581, 2025

    arXiv 2025

  14. [14]

    Uncertainty calibration for tool-using language agents

    Hao Liu, Zi-Yi Dou, Yixin Wang, Nanyun Peng, and Yisong Yue. Uncertainty calibration for tool-using language agents. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 16781–16805, 2024

    2024

  15. [15]

    Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation.arXiv preprint arXiv:2302.09664, 2023

    Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation.arXiv preprint arXiv:2302.09664, 2023. 11

    Pith/arXiv arXiv 2023

  16. [16]

    Uncertainty propagation on llm agent

    Qiwei Zhao, Dong Li, Yanchi Liu, Wei Cheng, Yiyou Sun, Mika Oishi, Takao Osaki, Katsushi Matsuda, Huaxiu Yao, Chen Zhao, et al. Uncertainty propagation on llm agent. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6064–6073, 2025

    2025

  17. [17]

    Tracer: Trajectory risk aggregation for critical episodes in agentic reasoning.arXiv preprint arXiv:2602.11409, 2026

    Sina Tayebati, Divake Kumar, Nastaran Darabi, Davide Ettori, Ranganath Krishnan, and Amit Ranjan Trivedi. Tracer: Trajectory risk aggregation for critical episodes in agentic reasoning.arXiv preprint arXiv:2602.11409, 2026

    arXiv 2026

  18. [18]

    Selaur: Self evolving llm agent via uncertainty-aware rewards.arXiv preprint arXiv:2602.21158, 2026

    Dengjia Zhang, Xiaoou Liu, Lu Cheng, Yaqing Wang, Kenton Murray, and Hua Wei. Selaur: Self evolving llm agent via uncertainty-aware rewards.arXiv preprint arXiv:2602.21158, 2026

    arXiv 2026

  19. [19]

    Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474, 2020

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474, 2020

    2020

  20. [20]

    Teaching models to express their uncertainty in words.arXiv preprint arXiv:2205.14334, 2022

    Stephanie Lin, Jacob Hilton, and Owain Evans. Teaching models to express their uncertainty in words.arXiv preprint arXiv:2205.14334, 2022

    Pith/arXiv arXiv 2022

  21. [21]

    A survey on llm-based multi-agent systems: workflow, infrastructure, and challenges.Vicinagearth, 1(1):9, 2024

    Xinyi Li, Sai Wang, Siqi Zeng, Yu Wu, and Yi Yang. A survey on llm-based multi-agent systems: workflow, infrastructure, and challenges.Vicinagearth, 1(1):9, 2024

    2024

  22. [22]

    Do not abstain! identify and solve the uncertainty

    Jingyu Liu, JingquanPeng JingquanPeng, Xiaopeng Wu, Xubin Li, Tiezheng Ge, Bo Zheng, and Yong Liu. Do not abstain! identify and solve the uncertainty. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 17177–17197, 2025

    2025

  23. [23]

    A survey on uncertainty quantification of large language models: Taxonomy, open research challenges, and future directions.ACM Computing Surveys, 58(3):1–38, 2025

    Ola Shorinwa, Zhiting Mei, Justin Lidard, Allen Z Ren, and Anirudha Majumdar. A survey on uncertainty quantification of large language models: Taxonomy, open research challenges, and future directions.ACM Computing Surveys, 58(3):1–38, 2025

    2025

  24. [24]

    Search wisely: Mitigating sub-optimal agentic searches by reducing uncertainty

    Peilin Wu, Mian Zhang, Xinlu Zhang, Xinya Du, and Zhiyu Chen. Search wisely: Mitigating sub-optimal agentic searches by reducing uncertainty. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 19734–19745, 2025

    2025

  25. [25]

    Lm-polygraph: Uncertainty estimation for language models

    Ekaterina Fadeeva, Roman Vashurin, Akim Tsvigun, Artem Vazhentsev, Sergey Petrakov, Kirill Fedyanin, Daniil Vasilev, Elizaveta Goncharova, Alexander Panchenko, Maxim Panov, et al. Lm-polygraph: Uncertainty estimation for language models. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages ...

    2023

  26. [26]

    Language models (mostly) know what they know.arXiv preprint arXiv:2207.05221, 2022

    Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. Language models (mostly) know what they know.arXiv preprint arXiv:2207.05221, 2022

    Pith/arXiv arXiv 2022

  27. [27]

    Detecting hallucinations in large language model generation: A token probability approach

    Ernesto Quevedo, Jorge Yero Salazar, Rachel Koerner, Pablo Rivas, and Tomas Cerny. Detecting hallucinations in large language model generation: A token probability approach. InWorld Congress in Computer Science, Computer Engineering & Applied Computing, pages 154–173. Springer, 2024

    2024

  28. [28]

    Learning to reason without external rewards.arXiv preprint arXiv:2505.19590, 2025

    Xuandong Zhao, Zhewei Kang, Aosong Feng, Sergey Levine, and Dawn Song. Learning to reason without external rewards.arXiv preprint arXiv:2505.19590, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    Unlocking exploration in rlvr: Uncertainty-aware advantage shaping for deeper reasoning.arXiv preprint arXiv:2510.10649, 2025

    Can Xie, Ruotong Pan, Xiangyu Wu, Yunfei Zhang, Jiayi Fu, Tingting Gao, and Guorui Zhou. Unlocking exploration in rlvr: Uncertainty-aware advantage shaping for deeper reasoning.arXiv preprint arXiv:2510.10649, 2025

    Pith/arXiv arXiv 2025

  30. [30]

    Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms.arXiv preprint arXiv:2306.13063, 2023

    Miao Xiong, Zhiyuan Hu, Xinyang Lu, Yifei Li, Jie Fu, Junxian He, and Bryan Hooi. Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms.arXiv preprint arXiv:2306.13063, 2023. 12

    Pith/arXiv arXiv 2023

  31. [31]

    Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171, 2022

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171, 2022

    Pith/arXiv arXiv 2022

  32. [32]

    Can llms learn uncertainty on their own? expressing uncertainty effectively in a self-training manner

    Shudong Liu, Zhaocong Li, Xuebo Liu, Runzhe Zhan, Derek F Wong, Lidia S Chao, and Min Zhang. Can llms learn uncertainty on their own? expressing uncertainty effectively in a self-training manner. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 21635–21645, 2024

    2024

  33. [33]

    First return, entropy-eliciting explore

    Tianyu Zheng, Tianshun Xing, Qingshui Gu, Taoran Liang, Xingwei Qu, Xin Zhou, Yizhi Li, Zhoufutu Wen, Chenghua Lin, Wenhao Huang, et al. First return, entropy-eliciting explore. arXiv preprint arXiv:2507.07017, 2025

    arXiv 2025

  34. [34]

    Agentic reinforced policy optimization

    Guanting Dong, Hangyu Mao, Kai Ma, Licheng Bao, Yifei Chen, Zhongyuan Wang, Zhongxia Chen, Jiazhen Du, Huiyang Wang, Fuzheng Zhang, et al. Agentic reinforced policy optimization. arXiv preprint arXiv:2507.19849, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Towards uncertainty-aware language agent

    Jiuzhou Han, Wray Buntine, and Ehsan Shareghi. Towards uncertainty-aware language agent. InFindings of the Association for Computational Linguistics: ACL 2024, pages 6662–6685, 2024

    2024

  36. [36]

    Rethinking llm uncertainty: A multi-agent approach to estimating black-box model uncertainty.arXiv preprint arXiv:2412.09572, 2024

    Yu Feng, Phu Mon Htut, Zheng Qi, Wei Xiao, Manuel Mager, Nikolaos Pappas, Kishaloy Halder, Yang Li, Yassine Benajiba, and Dan Roth. Rethinking llm uncertainty: A multi-agent approach to estimating black-box model uncertainty.arXiv preprint arXiv:2412.09572, 2024

    arXiv 2024

  37. [37]

    Towards reliable alignment: Uncertainty-aware rlhf

    Debangshu Banerjee and Aditya Gopalan. Towards reliable alignment: Uncertainty-aware rlhf. arXiv preprint arXiv:2410.23726, 2024

    arXiv 2024

  38. [38]

    Uncertainty-penalized direct preference optimization.arXiv preprint arXiv:2410.20187, 2024

    Sam Houliston, Alizée Pace, Alexander Immer, and Gunnar Rätsch. Uncertainty-penalized direct preference optimization.arXiv preprint arXiv:2410.20187, 2024

    arXiv 2024

  39. [39]

    Vcrl: Variance-based curriculum reinforcement learning for large language models

    Guochao Jiang, Wenfeng Feng, Guofeng Quan, Chuzhan Hao, Yuewei Zhang, Guohua Liu, and Hao Wang. Vcrl: Variance-based curriculum reinforcement learning for large language models. arXiv preprint arXiv:2509.19803, 2025

    arXiv 2025

  40. [40]

    Uncertainty and influence aware reward model refinement for reinforcement learning from human feedback

    Zexu Sun, Yiju Guo, Yankai Lin, Xu Chen, Qi Qi, Xing Tang, Ji-Rong Wen, et al. Uncertainty and influence aware reward model refinement for reinforcement learning from human feedback. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  41. [41]

    Towards agents that know when they don’t know: Uncertainty as a control signal for structured reasoning

    Josefa Lia Stoisser, Marc Boubnovski Martell, Lawrence Phillips, Gianluca Mazzoni, Lea Mørch Harder, Philip Torr, Jesper Ferkinghoff-Borg, Kaspar Martens, and Julien Fauqueur. Towards agents that know when they don’t know: Uncertainty as a control signal for structured reasoning. arXiv preprint arXiv:2509.02401, 2025

    arXiv 2025

  42. [42]

    Soft self-consistency improves language models agents

    Han Wang, Archiki Prasad, Elias Stengel-Eskin, and Mohit Bansal. Soft self-consistency improves language models agents. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 287–301, 2024

    2024

  43. [43]

    When agents disagree with themselves: Measuring behavioral consistency in llm-based agents.arXiv preprint arXiv:2602.11619, 2026

    Aman Mehta. When agents disagree with themselves: Measuring behavioral consistency in llm-based agents.arXiv preprint arXiv:2602.11619, 2026

    arXiv 2026

  44. [44]

    Xiaoou Liu, Tiejin Chen, Longchao Da, Chacha Chen, Zhen-Yu Lin, and Hua Wei. Uncertainty quantification and confidence calibration in large language models: A survey.Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V .2, 2025. URL https://api.semanticscholar.org/CorpusID:277150701

    2025

  45. [45]

    Uncertainty estimation in autoregressive structured prediction

    Andrey Malinin and Mark Gales. Uncertainty estimation in autoregressive structured prediction. arXiv preprint arXiv:2002.07650, 2020

    arXiv 2002

  46. [46]

    Detecting hallucinations in large language models using semantic entropy.Nature, 630(8017):625–630, 2024

    Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy.Nature, 630(8017):625–630, 2024

    2024

  47. [47]

    Kernel language entropy: Fine-grained uncertainty quantification for llms from semantic similarities.Advances in Neural Information Processing Systems, 37:8901–8929, 2024

    Alexander Nikitin, Jannik Kossen, Yarin Gal, and Pekka Marttinen. Kernel language entropy: Fine-grained uncertainty quantification for llms from semantic similarities.Advances in Neural Information Processing Systems, 37:8901–8929, 2024. 13

    2024

  48. [48]

    Why semantic entropy fails: Geometry-aware and calibrated uncertainty for policy optimization

    Zheyuan Zhang, Kaiwen Shi, Han Bao, Zehong Wang, Tianyi Ma, and Yanfang Ye. Why semantic entropy fails: Geometry-aware and calibrated uncertainty for policy optimization. arXiv preprint arXiv:2605.21801, 2026

    Pith/arXiv arXiv 2026

  49. [49]

    Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback

    Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher D Manning. Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages ...

    2023

  50. [50]

    Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models

    Potsawee Manakul, Adian Liusie, and Mark Gales. Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models. InProceedings of the 2023 conference on empirical methods in natural language processing, pages 9004–9017, 2023

    2023

  51. [51]

    Generating with confidence: Uncertainty quantification for black-box large language models.arXiv preprint arXiv:2305.19187, 2023

    Zhen Lin, Shubhendu Trivedi, and Jimeng Sun. Generating with confidence: Uncertainty quantification for black-box large language models.arXiv preprint arXiv:2305.19187, 2023

    arXiv 2023

  52. [52]

    The internal state of an llm knows when it’s lying

    Amos Azaria and Tom Mitchell. The internal state of an llm knows when it’s lying. InFindings of the Association for Computational Linguistics: EMNLP 2023, pages 967–976, 2023

    2023

  53. [53]

    Semantic entropy probes: Robust and cheap hallucination detection in llms.arXiv preprint arXiv:2406.15927, 2024

    Jannik Kossen, Jiatong Han, Muhammed Razzak, Lisa Schut, Shreshth Malik, and Yarin Gal. Semantic entropy probes: Robust and cheap hallucination detection in llms.arXiv preprint arXiv:2406.15927, 2024

    Pith/arXiv arXiv 2024

  54. [54]

    Drift-bench: Diagnosing cooperative breakdowns in llm agents under input faults via multi-turn interaction.arXiv preprint arXiv:2602.02455, 2026

    Han Bao, Zheyuan Zhang, Pengcheng Jing, Zhengqing Yuan, Kaiwen Shi, and Yanfang Ye. Drift-bench: Diagnosing cooperative breakdowns in llm agents under input faults via multi-turn interaction.arXiv preprint arXiv:2602.02455, 2026

    arXiv 2026

  55. [55]

    Active retrieval augmented generation

    Zhengbao Jiang, Frank F Xu, Luyu Gao, Zhiqing Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang, Jamie Callan, and Graham Neubig. Active retrieval augmented generation. InProceedings of the 2023 conference on empirical methods in natural language processing, pages 7969–7992, 2023

    2023

  56. [56]

    Ren, Anushri Dixit, Alexandra Bodrova, Sumeet Singh, Stephen Tu, Noah Brown, Peng Xu, Leila Takayama, F

    Allen Z. Ren, Anushri Dixit, Alexandra Bodrova, Sumeet Singh, Stephen Tu, Noah Brown, Peng Xu, Leila Takayama, F. Xia, Jacob Varley, Zhenjia Xu, Dorsa Sadigh, Andy Zeng, and Anirudha Majumdar. Robots that ask for help: Uncertainty alignment for large language model planners.ArXiv, abs/2307.01928, 2023. URL https://api.semanticscholar.org/ CorpusID:259342058

    arXiv 2023

  57. [57]

    React: Synergizing reasoning and acting in language models.ArXiv, abs/2210.03629,

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models.ArXiv, abs/2210.03629,

    Pith/arXiv arXiv

  58. [58]

    URLhttps://api.semanticscholar.org/CorpusID:252762395

  59. [59]

    Griffiths, Yuan Cao, and Karthik Narasimhan

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language mod- els.ArXiv, abs/2305.10601, 2023. URL https://api.semanticscholar.org/CorpusID: 258762525

    Pith/arXiv arXiv 2023

  60. [60]

    Plummer, Zhaoran Wang, and Hongxia Yang

    Chau Pham, Boyi Liu, Yingxiang Yang, Zhengyu Chen, Tianyi Liu, Jianbo Yuan, Bryan A. Plummer, Zhaoran Wang, and Hongxia Yang. Let models speak ciphers: Multiagent debate through embeddings.ArXiv, abs/2310.06272, 2023. URL https://api.semanticscholar. org/CorpusID:263830342

    arXiv 2023

  61. [61]

    Weston, and Yuandong Tian

    Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason E. Weston, and Yuandong Tian. Training large language models to reason in a continuous latent space. ArXiv, abs/2412.06769, 2024. URL https://api.semanticscholar.org/CorpusID: 274610816

    Pith/arXiv arXiv 2024

  62. [62]

    Enabling agents to communicate entirely in latent space.ArXiv, abs/2511.09149,

    Zhuoyun Du, Runze Wang, Huiyu Bai, Zouying Cao, Xiaoyong Zhu, Bo Zheng, Wei Chen, and Haochao Ying. Enabling agents to communicate entirely in latent space.ArXiv, abs/2511.09149,

    Pith/arXiv arXiv

  63. [63]

    URLhttps://api.semanticscholar.org/CorpusID:282939672. 14

  64. [64]

    Maguire, and Dejan Kosti ´c

    Xiangyu Shi, Marco Chiesa, Gerald Q. Maguire, and Dejan Kosti ´c. Kvcomm: Enabling efficient llm communication through selective kv sharing.ArXiv, abs/2510.03346, 2025. URL https://api.semanticscholar.org/CorpusID:281842742

    arXiv 2025

  65. [65]

    Ng-router: Graph-supervised multi-agent collaboration for nutrition question answering

    Kaiwen Shi, Zheyuan Zhang, Zhengqing Yuan, Keerthiram Murugesan, Vincent Galassi, Chuxu Zhang, and Yanfang Ye. Ng-router: Graph-supervised multi-agent collaboration for nutrition question answering. InProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7508–7527, 2026

    2026

  66. [66]

    Autodata: A multi-agent system for open web data collection.Advances in Neural Information Processing Systems, 38:173416– 173448, 2026

    Tianyi Ma, Yiyue Qian, Zheyuan Zhang, Zehong Wang, Xiaoye Qian, Feifan Bai, Yifan Ding, Xuwei Luo, Shinan Zhang, Keerthiram Murugesan, et al. Autodata: A multi-agent system for open web data collection.Advances in Neural Information Processing Systems, 38:173416– 173448, 2026

    2026

  67. [67]

    Mapro: Recasting multi-agent prompt optimization as maximum a posteriori inference

    Zheyuan Zhang, Lin Ge, Hongjiang Li, Weicheng Zhu, Chuxu Zhang, and Yanfang Ye. Mapro: Recasting multi-agent prompt optimization as maximum a posteriori inference. InFindings of the Association for Computational Linguistics: EACL 2026, pages 4458–4480, 2026

    2026

  68. [68]

    attended by

    Jiatan Huang, Zheyuan Zhang, Kaiwen Shi, Yanfang Ye, and Chuxu Zhang. Evolver- outer: Co-evolving routing and prompt for multi-agent question answering.arXiv preprint arXiv:2604.05149, 2026. 15 A Experimental Details A.1 Task, Tool-Use Setting, and Model We use HotpotQA as the evaluation setting because it contains multi-hop questions that often require g...

    Pith/arXiv arXiv 2026