Pith. sign in

REVIEW 3 major objections 5 minor 50 references

The harness that mediates an LLM agent can rewrite its multi-step beliefs even when the task, environment, and model stay fixed.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

T0 review · grok-4.5

2026-07-11 17:53 UTC pith:32PNJDZV

load-bearing objection Clean isolation of harness effects on multi-step beliefs; the real soft spot is that the JSON rollout is still a prompted self-report, not free-form control state. the 3 major comments →

arxiv 2607.04528 v1 pith:32PNJDZV submitted 2026-07-05 cs.AI

Measuring Harness-Induced Belief Divergence in Multi-Step LLM Agents

classification cs.AI
keywords LLM agentsagent harnessbelief trajectoriesbelief divergencemulti-step reasoningagent evaluationsoftware agentsBIWM
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Software-agent benchmarks usually score only whether a task was solved, treating the surrounding execution system as neutral plumbing. This paper shows that the harness—what the agent is allowed to see, which actions run, which failures are repaired, which states are verified, and which evidence is logged—can change the agent’s multi-step beliefs about progress, risk, recoverability, failure mode, uncertainty, future success, repair cost, and next action, even when task, environment, and base model are held fixed. The authors introduce a structured K-step belief-rollout diagnostic and a cross-harness divergence score split into an immediate interface (arrival) term and a horizon-dependent growth term. On controlled coding tasks and public stress tests, blocked actions, compressed repairs, selective verification, and cost-aware pruning often leave terminal success intact while changing the beliefs that drive later decisions. They also offer BIWM, a no-training protocol that makes censored and compressed evidence visible and aligns belief trajectories across harness views, arguing that harness design is an experimental variable in agent evaluation, not an implementation detail.

Core claim

With task, environment, and base LLM fixed, the harness can change multi-step belief trajectories over progress, risk, recoverability, constraints, failure mode, uncertainty, future success, repair cost, and next action. Blocked actions, compressed repairs, selective verification, and cost-aware evidence pruning often preserve terminal success while changing the beliefs that drive later decisions. Harness design is therefore an experimental variable in agent evaluation, not an implementation detail.

What carries the argument

The belief-rollout diagnostic: a structured K-step trajectory of progress, risk, recoverability, constraints, failure mode, uncertainty, success and cost forecasts, and next action, compared across harnesses by a belief divergence that decomposes into an arrival term (immediate interface shifts in constraints and actions) and a growth term (horizon-dependent changes in categorical state, failure mode, and numeric forecasts). BIWM is the companion no-training protocol that canonicalizes observations, logs blocked branches, expands repair traces, records verification masks, shadow-executes risky branches, and aligns trajectories.

Load-bearing premise

The method treats a forced structured JSON belief rollout as a faithful readout of the intermediate world model that actually drives later decisions, rather than a prompt-sensitive self-report that can diverge from the true control state.

What would settle it

Run the same base model under two harnesses that produce large growth divergence and different next-action categories, then continue free multi-step execution without the belief-elicitation prompt: if subsequent tool sequences and intermediate environment states match despite the reported belief gap, the claim that harness-shaped beliefs drive later decisions is falsified.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Final success alone is insufficient for comparing agents when harnesses differ, because intermediate beliefs and next-action categories can diverge.
  • Evaluations should report or control observation filtering, action gating, repair compression, verification policy, and logging as experimental factors.
  • Blocking an action also censors evidence; without blocked-branch logs, an agent may treat a prohibited path as nonexistent rather than forbidden.
  • Scalar divergence can plateau while forecast components keep moving, so arrival/growth and component-level reporting are required.
  • Instrumentation and cross-harness alignment can reduce dependence on any single harness-conditioned belief trajectory.

Where Pith is reading between the lines

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

  • Leaderboards that mix frameworks with different default harnesses may partly rank interface effects rather than model capability.
  • Safety gates that only stop execution without shadow or blocked-action evidence can leave unsafe retry tendencies alive in the belief state.
  • Belief-divergence audits could become a standard disclosure layer for production agent platforms, parallel to logging protocol transparency.
  • The same measurement protocol extends naturally to web and desktop agents, where observation abstraction and risk gates are equally strong.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper argues that software-agent harnesses (observation maps, action gates, repair policies, verifiers, and logging) reshape multi-step LLM belief trajectories even when task, environment, and base model are fixed, so harness design is an experimental variable rather than neutral infrastructure. It defines a structured K-step belief rollout over progress, risk, recoverability, constraints, failure mode, numeric forecasts, and next action; introduces a weighted divergence Dbelief with an arrival/growth decomposition; and reports controlled coding results plus SWE-bench Verified and Terminal-Bench stress tests showing that blocked actions, compressed repairs, selective verification, and cost-aware pruning often preserve terminal success while shifting intermediate beliefs and coarse next-action categories. It further proposes BIWM, a no-training instrumentation protocol (canonicalisation, blocked-action logs, repair unrolling, verification masks, shadow execution, cross-harness alignment) that exposes censored evidence and reduces trajectory lock-in.

Significance. If the measurement holds, the paper identifies a first-order confound in agent evaluation that final-success metrics systematically miss, with direct implications for coding, terminal, tool-use, and safety-gated agents. Strengths include a clean isolation design (fixed T, E, M; vary H only), explicit formal definitions of harness, belief space, and component distances, appendix proofs of basic properties and a censoring-growth theorem, a useful Darrival/Dgrowth split that explains scalar plateauing, weight-sensitivity analysis, action-level and UnsafeRetryRate checks, BIWM ablations across benchmarks, and a public code release with deterministic recomputation notes. These make the contribution more than a conceptual critique: it ships an auditable diagnostic protocol that other labs can re-run.

major comments (3)
  1. Definitions 2–3 and Algorithm 1 treat forced JSON K-step belief rollouts as the intermediate world-model whose cross-harness divergence explains later control. Section 10’s action-divergence analysis (0.28→0.595 by Dgrowth quartile; UnsafeRetryRate=0.70) still largely uses the same schema’s next_action_recommendation field rather than an independent free-form agent loop that acts without the belief prompt. This is the load-bearing bridge for the claim that harness design is a first-order experimental variable for agent evaluation, not only for schema-compliant self-reports. Please add at least one experiment where actions are taken in a standard ReAct-style loop without the belief schema, then compare (i) free-form action-category mismatch across harnesses and (ii) whether Dgrowth from a parallel, non-controlling elicitation still predicts those free-form mismatches. If the free-form con
  2. Tables 1–6 and related figures report mean divergences without seed-level variance, confidence intervals, or paired tests, despite three seeds and small task counts (HIBENCH-CODE: 8 tasks; SWE-bench slice limited; Terminal-Bench groups n=15). Given that middle ranks shift under weight variants (Table 10 / Figure 9) and several growth profiles are non-monotone, the mechanism-specific claims (e.g., structured largest growth at K=5; risk-gated failure-mode rise on SWE-bench) need uncertainty quantification. Report per-comparison standard errors or bootstrap CIs over task–seed pairs, and mark which horizon/harness differences remain significant under the same fixed weights used in the main text.
  3. Theorem 1’s growth monotonicity rests on condition (C3), a non-negative expected growth-component increment that is assumed rather than checked. The long-horizon results (Tables 2–3, Figure 5) show non-monotone Dgrowth and a transient failure-mode convergence at K=5, which already indicates that (C3) does not hold unconditionally. Either empirically test the sign of the growth increments on the logged rollouts and state where the theorem applies, or reframe Theorem 1 clearly as a sufficient-condition formalisation rather than explanatory support for the observed trajectories.
minor comments (5)
  1. Figure 1’s steak/chopstick analogy is memorable but slightly under-specified relative to the software harness six-tuple; a one-line mapping from utensils to OH/GH would help non-specialist readers.
  2. Notation: Dbelief is sometimes written with and without the horizon argument; state once that omitted K means the table’s fixed horizon.
  3. Appendix C/E self-consistency AUROCs are correctly caveated as non-environment-grounded, but the main text still risks over-reading BIWM-full AUROC gains; keep them exploratory and out of the abstract-level claims.
  4. Clarify whether η8 action truncation (Definition 4) can misclassify distinct long tool strings that share an 8-token prefix; a short ablation with full-string equality would strengthen Dact.
  5. The GitHub URL appears with a line break in the abstract/PDF; fix the link formatting for reproducibility.

Circularity Check

1 steps flagged

No load-bearing circularity: main claims are controlled empirical measurements; only a mild tautological formalization in Theorem 1.

specific steps
  1. self definitional [Theorem 1 (Sec. 3.5) and proof A.5; condition (C3) / Eq. (42)–(43)]
    "(C3) Non-negative expected growth increment. For every t≥1, ∑_{i∈Igro} w̃_i E[D_i(B^{Hraw}_{t+1}, B^H_{t+1}) − D_i(B^{Hraw}_t, B^H_t)] ≥ 0. Then, for every K≥1, E[Dgrowth(Hraw, H; K+1)] ≥ E[Dgrowth(Hraw, H; K)]. … Condition (42), evaluated at t=K, implies that the right-hand side is non-negative. Therefore, E[G_{K+1}] ≥ E[G_K]."

    The stated conclusion is that expected growth divergence is non-decreasing in K. The decisive assumption (C3) is exactly that the expected weighted growth-component increments are non-negative at each step. The proof only sums those increments and invokes (C3); no independent mechanism is derived. The paper itself notes this formalizes persistence under an extra condition rather than proving unconditional censoring-induced growth. Minor and not load-bearing for the empirical harness-divergence results.

full rationale

The paper’s central claim is not a first-principles derivation that reduces to its inputs. It holds (T, E, M) fixed, varies harnesses, elicits structured K-step belief rollouts under a fixed schema, and reports measured cross-harness Dbelief / Darrival / Dgrowth plus action-category co-occurrence. Component weights are fixed a priori and sensitivity-checked; they change scale and middle ranks but do not force the existence of harness-conditioned shifts. BIWM is a no-training instrumentation protocol whose “exposure” vs “robustness” interpretations are definitional bookkeeping of what was logged, not fitted predictions renamed as theory. Self-consistency AUROCs are explicitly internal (same-model early vs late imagined state), not environment-grounded calibration sold as external prediction. There is no uniqueness theorem imported from overlapping authors, no ansatz smuggled via self-citation, and no parameter fit re-labeled as a prediction of a closely related target. The only mild self-definitional step is Theorem 1: monotonicity of expected Dgrowth is assumed via non-negative expected growth increments (C3) and then restated as the conclusion—honest conditional formalization, not load-bearing for the empirical claim. Score 1 reflects that minor tautology only; the measurement chain is self-contained against the paper’s own controlled and public-benchmark logs.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 4 invented entities

The paper is primarily an empirical measurement paper. Load-bearing structure comes from a hand-designed belief schema and weighted divergence, a canonical-embedding assumption that makes cross-harness comparison well-defined, and the premise that elicited JSON rollouts track decision-relevant beliefs. Free parameters are metric weights and schema discretizations rather than physics-style fitted constants. Invented entities are measurement constructs and a logging protocol, not new physical mediators.

free parameters (5)
  • belief component weights w=(0.30,0.15,0.25,0.25,0.05)
    Hand-chosen fixed weights for Dcat/Dfail/Dset/Dnum/Dact that determine Dbelief and the arrival/growth split; sensitivity analysis shows scale and middle ranks move with these weights.
  • action comparison truncation η8 (first 8 normalized tokens)
    Ad hoc discretization for Dact; changes when two recommended actions count as divergent.
  • categorical belief discretization (progress 1–5; risk/recoverability 1–3)
    Schema design choice that defines the belief space and Dcat normalization ranges.
  • numeric clip range κ=5.0 for risk/cost fields
    Normalization bound used in Dnum; chosen by protocol rather than derived.
  • rollout horizons K ∈ {1,3,5,8,...20}
    Evaluation grid chosen by authors; growth claims depend on these horizons.
axioms (6)
  • domain assumption Canonical Embedding: observations/actions from all harnesses inject into shared schema spaces O*, A* independent of harness instance (Assumption 1).
    Required for cross-harness distances to be comparable rather than syntax artifacts; implemented by BIWM canonicalisation.
  • domain assumption Elicited structured JSON belief states under a fixed template are valid samples from the model’s harness-conditioned belief distribution P^B_M.
    Definition 3 treats forced schema outputs as the world-model trajectory being measured.
  • domain assumption First-order property of the augmented process (B_t, h_t) under each harness (Definition 3 / condition C2).
    Used in the censoring-induced growth monotonicity theorem.
  • ad hoc to paper Non-negative expected growth-component increment across steps (condition C3 of Theorem 1).
    Sufficient condition for E[Dgrowth] monotonicity; not generally established empirically for all harnesses.
  • ad hoc to paper Component distances in Definition 4 are appropriate comparable coordinates for belief disagreement.
    Jaccard-style set distance, 8-token action equality, and normalized ordinal gaps are design choices, not uniquely forced.
  • standard math Standard agent evaluation separation of model, environment, and interface (Liu et al., Ma et al., Yang et al.).
    Background formalization used to isolate harness as the sole experimental variable.
invented entities (4)
  • K-step LLM belief rollout / belief space B no independent evidence
    purpose: Make intermediate multi-step agent beliefs auditable and comparable across harnesses.
    A paper-specific structured state (progress, risk, constraints, forecasts, next action) used as the measured object.
  • Dbelief with Darrival/Dgrowth decomposition no independent evidence
    purpose: Quantify and separate immediate interface mismatch from horizon-dependent belief drift.
    New diagnostic scores defined from weighted component distances; sensitivity depends on chosen weights.
  • BIWM (Belief-Invariant World-Modeling) protocol no independent evidence
    purpose: Expose censored/compressed evidence and align beliefs across harness views without training.
    No-training instrumentation stack: canonicalize, log blocked branches, unroll repairs, masks, shadow exec, align.
  • HIBENCH-CODE-v0 controlled coding benchmark no independent evidence
    purpose: Expose harness mechanisms on small controlled software-failure tasks.
    Author-constructed mechanism benchmark rather than an externally standardized suite.

pith-pipeline@v1.1.0-grok45 · 35261 in / 3896 out tokens · 43180 ms · 2026-07-11T17:53:32.367774+00:00 · methodology

0 comments
read the original abstract

Software-agent benchmarks usually report whether an agent solves a task, but the agent reaches that outcome through a harness that controls what it sees, which actions it can take, which failures are repaired, which states are verified, and which evidence is logged. We show that this harness can change the agent's multi-step beliefs even when the task, environment, and base LLM are fixed. We introduce a belief-rollout diagnostic that elicits structured K-step trajectories over progress, risk, recoverability, constraints, failure mode, uncertainty, future success, repair cost, and next action under alternative harnesses. We define a cross-harness belief divergence and decompose it into an arrival term for immediate interface shifts and a growth term for horizon-dependent belief changes. On controlled coding tasks and public-benchmark stress tests, blocked actions, compressed repairs, selective verification, and cost-aware evidence pruning often preserve terminal success while changing the beliefs that drive later decisions. We further introduce BIWM, a no-training protocol that canonicalizes observations, logs censored branches, expands repair traces, records verification masks, executes risky branches in shadow, and aligns belief trajectories across harness views. The results suggest that harness design is an experimental variable in agent evaluation, not an implementation detail. Our code is available at https://github.com/Hik289/Harness-induce-bias.git.

Figures

Figures reproduced from arXiv: 2607.04528 by Haiwen Yi, Xinyuan Song.

Figure 1
Figure 1. Figure 1: The same task and the same LLM can induce different beliefs under different harnesses. This schematic analogy keeps the task fixed while changing the interface. Under the raw reference harness H0, which provides a knife and fork, the model judges the steak task to be doable and assigns a high probability of success. Under the risk-gated harness H2, the chopstick interface is relabelled as a policy-violatio… view at source ↗
Figure 2
Figure 2. Figure 2: Measurement overview. The task, environment, base LLM, schema, and decoding rule are held fixed while the harness changes the evidence stream through observation abstraction, action gating, repair compression, selective verification, or logging. The diagnostic compares the resulting K-step belief trajectories with Dbelief and then separates the score into the arrival readout Darrival, which captures immedi… view at source ↗
Figure 3
Figure 3. Figure 3: Harness-to-belief instrumentation pipeline. The harness layer instantiates six execution interfaces, from the raw reference harness H0 to structured, risk-gated, repair-heavy, verification-selective, and cost-aware variants H1–H5. Each harness filters the evidence presented to the fixed LLM and induces a structured K-step belief rollout over progress, risk, success probability, and failure mode. The no-tra… view at source ↗
Figure 4
Figure 4. Figure 4: Growth divergence across rollout horizons. The figure reports Dgrowth for the structured (H1), risk￾gated (H2), and verification-selective (H4) harnesses rel￾ative to the raw reference H0. Structured parsing shows sustained growth through K = 5, whereas risk gating and selective verification exhibit mechanism-specific non-monotone trajectories. These differences are largely hidden in the overall divergence… view at source ↗
Figure 5
Figure 5. Figure 5: Overall divergence stabilises before all com￾ponent distances do. The figure reports Dbelief from K = 1 to K = 20 for each mediated harness rela￾tive to H0 on the long-horizon HIBENCH-CODE sup￾plement. The y-axis spans 0.2 to 0.8, leaving visible margin around all measured trajectories while preserv￾ing the later-horizon variation. The dashed line marks K = 5, where the averaged failure-mode distance ex￾hi… view at source ↗
Figure 6
Figure 6. Figure 6: shows that the aligned view has lower growth divergence than the mean individual har￾ness view across the evaluated horizons. The gap increases with rollout depth, while the arrival read￾out remains nearly unchanged. Alignment there￾fore acts mainly on the evolving belief components rather than on the fixed interface differences cap￾tured by Darrival. The wrapper and alignment results address two distinct … view at source ↗
Figure 7
Figure 7. Figure 7: Harness effects persist across grouped pub￾lic benchmarks. Dbelief by harness across grouped benchmarks. Left: SWE-bench Verified (A_bug / B_feature / C_refactor). Right: Terminal-Bench (X_risky / Y_timeout / Z_verif, 15 tasks each). The grouped Terminal-Bench slice removes the under￾populated buckets from the earlier stress test and shows a recurring harness effect, with the leading mechanism changing by … view at source ↗
Figure 9
Figure 9. Figure 9: Metric-weight sensitivity changes scale but not the diagnostic story. Dbelief at K = 5 under three metric weight settings. The structured view remains the largest-divergence case, while middle ranks vary with the arrival/growth emphasis; this is why the main experiments report Darrival, Dgrowth, and component distances alongside the scalar. while the middle ranks shift. This is the right behaviour for a we… view at source ↗
Figure 10
Figure 10. Figure 10: BIWM has separate exposure and robust￾ness effects. Baseline (grey dashed), BIWM-wrapper (orange solid), and BIWM-6 aligned (blue solid) Dbelief across the five mediated harnesses. Large upward shifts for repair-heavy and verification-selective exe￾cution indicate exposure of suppressed channels; down￾ward shifts from baseline to aligned indicate robustness against trajectory lock-in. 13 Cross-Harness Bel… view at source ↗

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

50 extracted references · 9 linked inside Pith

  1. [1]

    International Conference on Learning Representations , year=

    ReAct: Synergizing Reasoning and Acting in Language Models , author=. International Conference on Learning Representations , year=

  2. [2]

    Advances in Neural Information Processing Systems , volume =

    Toolformer: Language Models Can Teach Themselves to Use Tools , author =. Advances in Neural Information Processing Systems , volume =

  3. [3]

    Liu, Xiao and Yu, Hao and Zhang, Hanchen and Xu, Yifan and Lei, Xuanyu and Lai, Hanyu and Gu, Yu and Ding, Hangliang and Men, Kaiwen and Yang, Kejuan and Zhang, Shudan and Deng, Xiang and Zeng, Aohan and Du, Zhengxiao and Zhang, Chenhui and Shen, Sheng and Zhang, Tianjun and Su, Yu and Sun, Huan and Huang, Minlie and Dong, Yuxiao and Tang, Jie , journal =

  4. [4]

    Ma, Chang and Zhang, Junlei and Zhu, Zhihao and Yang, Cheng and Yang, Yujiu and Jin, Yaohui and Lan, Zhenzhong and Kong, Lingpeng and He, Junxian , journal =

  5. [5]

    and Yang, John and Wettig, Alexander and Yao, Shunyu and Pei, Kexin and Press, Ofir and Narasimhan, Karthik , booktitle =

    Jimenez, Carlos E. and Yang, John and Wettig, Alexander and Yao, Shunyu and Pei, Kexin and Press, Ofir and Narasimhan, Karthik , booktitle =

  6. [6]

    2024 , month = aug, howpublished =

    Introducing. 2024 , month = aug, howpublished =

  7. [7]

    International Conference on Learning Representations , year =

    Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces , author =. International Conference on Learning Representations , year =

  8. [8]

    2023 , eprint=

    Reflexion: Language Agents with Verbal Reinforcement Learning , author=. 2023 , eprint=

  9. [9]

    arXiv preprint arXiv:2405.15793 , year=

    SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering , author=. arXiv preprint arXiv:2405.15793 , year=

  10. [10]

    International Conference on Learning Representations , year=

    AgentSquare: Automatic LLM Agent Search in Modular Design Space , author=. International Conference on Learning Representations , year=

  11. [11]

    arXiv preprint arXiv:2605.22166 , year=

    Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents , author=. arXiv preprint arXiv:2605.22166 , year=

  12. [12]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , year=

    Reasoning with Language Model is Planning with World Model , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , year=

  13. [13]

    International Conference on Machine Learning , year=

    Learning Latent Dynamics for Planning from Pixels , author=. International Conference on Machine Learning , year=

  14. [14]

    2024 , eprint=

    Evaluating the World Model Implicit in a Generative Model , author=. 2024 , eprint=

  15. [15]

    Wang, Kangrui and Zhang, Pingyue and Wang, Zihan and Gao, Yaning and Li, Linjie and Wang, Qineng and Chen, Hanyang and Wan, Chi and Lu, Yiping and Yang, Zhengyuan and Wang, Lijuan and Krishna, Ranjay and Wu, Jiajun and Fei-Fei, Li and Choi, Yejin and Li, Manling , journal=

  16. [16]

    2024 , eprint=

    WebArena: A Realistic Web Environment for Building Autonomous Agents , author=. 2024 , eprint=

  17. [17]

    2024 , eprint=

    -bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains , author=. 2024 , eprint=

  18. [18]

    2024 , eprint=

    RedCode: Risky Code Execution and Generation Benchmark for Code Agents , author=. 2024 , eprint=

  19. [19]

    What Twelve

    Moghadasi, Mahdi Naser and Ghaderi, Faezeh , journal=. What Twelve

  20. [20]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining , doi=

    Uncertainty Quantification and Confidence Calibration in Large Language Models: A Survey , author=. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining , doi=

  21. [21]

    arXiv preprint arXiv:2510.16492 , year=

    Check Yourself Before You Wreck Yourself: Selectively Quitting Improves LLM Agent Safety , author=. arXiv preprint arXiv:2510.16492 , year=

  22. [22]

    ToolSafe: Enhancing Tool Invocation Safety of

    Mou, Yutao and Xue, Zhangchi and Li, Lijun and Liu, Peiyang and Zhang, Shikun and Ye, Wei and Shao, Jing , journal=. ToolSafe: Enhancing Tool Invocation Safety of

  23. [23]

    arXiv preprint arXiv:2605.04785 , year=

    AgentTrust: Runtime Safety Evaluation and Interception for AI Agent Tool Use , author=. arXiv preprint arXiv:2605.04785 , year=

  24. [24]

    A survey on

    Li, Xinyi and Wang, Sai and Zeng, Siqi and Wu, Yu and Yang, Yi , journal=. A survey on

  25. [25]

    Understanding the planning of

    Huang, Xu and Liu, Weiwen and Chen, Xiaolong and Wang, Xingmei and Wang, Hao and Lian, Defu and Wang, Yasheng and Tang, Ruiming and Chen, Enhong , journal=. Understanding the planning of

  26. [26]

    arXiv preprint arXiv:2503.21460 , year=

    Large Language Model Agent: A Survey on Methodology, Applications and Challenges , author=. arXiv preprint arXiv:2503.21460 , year=

  27. [27]

    and others , booktitle=

    Wang, Xingyao and Li, Boxuan and Song, Yufan and Xu, Frank F. and others , booktitle=. 2025 , url=

  28. [28]

    Stein, Merlin , howpublished=. How are

  29. [29]

    Wang, Xingyao and Rosenberg, Simon and Michelini, Juan and Smith, Calvin and others , journal=. The

  30. [30]

    2025 , url=

    Bouzenia, Islem and Devanbu, Premkumar and Pradel, Michael , booktitle=. 2025 , url=

  31. [31]

    Jin, Haolin and Huang, Linghan and Cai, Haipeng and Yan, Jun and Li, Bo and Chen, Huaming , journal=. From

  32. [32]

    Gu, Yu and Zheng, Boyuan and Gou, Boyu and Zhang, Kai and Chang, Cheng and Srivastava, Sanjari and Xie, Yanan and Qi, Peng and Sun, Huan and Su, Yu , journal=. Is Your

  33. [33]

    arXiv preprint arXiv:2301.04104 , year=

    Mastering Diverse Domains through World Models , author=. arXiv preprint arXiv:2301.04104 , year=

  34. [34]

    arXiv preprint arXiv:2507.23773 , year=

    General Agentic Planning Through Simulative Reasoning with World Models , author=. arXiv preprint arXiv:2507.23773 , year=

  35. [35]

    The Verifier Tax: Horizon Dependent Safety Success Tradeoffs in Tool Using

    Sah, Tanmay and Srivastava, Vishal and Sah, Dolly and Jordan, Kayden , journal=. The Verifier Tax: Horizon Dependent Safety Success Tradeoffs in Tool Using

  36. [36]

    When Refusals Fail: Unstable Safety Mechanisms in Long-Context

    Hadeliya, Tsimur and Jauhar, Mohammad Ali and Sakpal, Nidhi and Cruz, Diogo , journal=. When Refusals Fail: Unstable Safety Mechanisms in Long-Context

  37. [37]

    From Reward-Hack Activations to Agentic Risk States: Context-Calibrated Mechanistic Monitoring in

    Wilhelm, Patrick and Kao, Odej , journal=. From Reward-Hack Activations to Agentic Risk States: Context-Calibrated Mechanistic Monitoring in

  38. [38]

    Duan, Jinhao and Diffenderfer, James and Madireddy, Sandeep and Chen, Tianlong and Kailkhura, Bhavya and Xu, Kaidi , journal=

  39. [39]

    Wang, Hanlin and Wang, Jian and Leong, Chak Tou and Li, Wenjie , booktitle=

  40. [40]

    arXiv preprint arXiv:2606.11211 , year=

    Calibration Drift Under Reasoning: How Chain-of-Thought Budgets Induce Overconfidence in Large Language Models , author=. arXiv preprint arXiv:2606.11211 , year=

  41. [41]

    Proceedings of the 34th International Conference on Machine Learning , pages=

    On Calibration of Modern Neural Networks , author=. Proceedings of the 34th International Conference on Machine Learning , pages=

  42. [42]

    arXiv preprint arXiv:2207.05221 , year=

    Language Models (Mostly) Know What They Know , author=. arXiv preprint arXiv:2207.05221 , year=

  43. [43]

    Advances in Neural Information Processing Systems , year=

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models , author=. Advances in Neural Information Processing Systems , year=

  44. [44]

    and Hashimoto, Tatsunori , journal=

    Ruan, Yangjun and Dong, Honghua and Wang, Andrew and Pitis, Silviu and Zhou, Yongchao and Ba, Jimmy and Dubois, Yann and Maddison, Chris J. and Hashimoto, Tatsunori , journal=. Identifying the Risks of

  45. [45]

    Xie, Tianbao and Zhang, Danyang and Chen, Jixuan and Li, Xiaochuan and Zhao, Siheng and Cao, Ruisheng and Hua, Toh Jing and Cheng, Zhoujun and Shin, Dongchan and Lei, Fangyu and others , booktitle=

  46. [46]

    Advances in Neural Information Processing Systems Datasets and Benchmarks Track , year=

    Debenedetti, Edoardo and Zhang, Jie and Balunovic, Mislav and Beurer-Kellner, Luca and Fischer, Marc and Tram. Advances in Neural Information Processing Systems Datasets and Benchmarks Track , year=

  47. [47]

    2023 , doi=

    Li, Minghao and Zhao, Yingxiu and Yu, Bowen and Song, Feifan and Li, Hangyu and Yu, Haiyang and Li, Zhoujun and Huang, Fei and Li, Yongbin , booktitle=. 2023 , doi=

  48. [48]

    Qin, Yujia and Liang, Shihao and Ye, Yining and Zhu, Kunlun and Yan, Lan and Lu, Yaxi and Lin, Yankai and Cong, Xin and Tang, Xiangru and Qian, Bill and Zhao, Sihan and Hong, Lauren and Tian, Runchu and Xie, Ruobing and Zhou, Jie and Gerstein, Mark and Li, Dahai and Liu, Zhiyuan and Sun, Maosong , booktitle=

  49. [49]

    and Zhang, Tianjun and Wang, Xin and Gonzalez, Joseph E

    Patil, Shishir G. and Zhang, Tianjun and Wang, Xin and Gonzalez, Joseph E. , booktitle=. Gorilla: Large Language Model Connected with Massive

  50. [50]

    and Mao, Huanzhi and Ji, Charlie Cheng-Jie and Yan, Fanjia and Suresh, Vishnu and Stoica, Ion and Gonzalez, Joseph E

    Patil, Shishir G. and Mao, Huanzhi and Ji, Charlie Cheng-Jie and Yan, Fanjia and Suresh, Vishnu and Stoica, Ion and Gonzalez, Joseph E. , booktitle=. The Berkeley Function Calling Leaderboard (