Pith. sign in

REVIEW 3 major objections 6 minor 54 references

A ranking of tools is not an acquisition decision: stop where continuing no longer pays for its cost.

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 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-30 11:38 UTC pith:BWRDLWCA

load-bearing objection Clean, usable control layer for cost-aware tool depth; theory is standard but well applied, and the Gx proxy is the real soft spot—not a fake SOTA. the 3 major comments →

arxiv 2607.27083 v1 pith:BWRDLWCA submitted 2026-07-29 cs.LG cs.AI

Scores Are Not Decisions: Cost-Aware Stopping for Tool Acquisition in LLM Agents

Yicheng Feng , Yan Zhang , Yan Cheng , Wei Qi This is my paper
classification cs.LG cs.AI
keywords LLM agentstool acquisitioncost-aware stoppingdecision-focused learningranked list depthheterogeneous costsregret-weighted classification
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.

LLM agents face a practical control problem before any tool runs: how many of the ranked candidates should the harness actually expose? Too few leaves the task under-informed; too many burns cost, context, and privacy. This paper argues that relevance scores alone cannot settle that depth once tool costs differ, and that the right object is a stop-versus-continue decision along the ranked prefix. It trains a lightweight classifier on the offline payoff gap between stopping now and the best longer prefix, weighting each mistake by how much payoff is at stake, and proves that this target is Bayes-aligned with optimal stopping. On 1,343 tasks across five domains, the method leads deployable baselines under heterogeneous costs and high cost pressure, and in live Retail runs it cuts read-tool exposure by about 37% while keeping task success comparable. The point for a reader is modular: existing rankers stay frozen, the LLM stays frozen, and a small pre-execution plugin turns scores into cost-aware depth.

Core claim

Tool acquisition depth after ranking is a cost-aware stopping problem, not a score-threshold problem. Training on the signed payoff gap between the current prefix and its best continuation—weighted by the gap’s magnitude—yields a Bayes-aligned stopping rule that outperforms score-only and predict-then-threshold policies when costs are heterogeneous, without changing the upstream ranker or the underlying model.

What carries the argument

CAM-DF (cost-aware marginal decision-focused stopping): a regret-weighted logistic stopper over ranked prefixes whose labels and weights come from Δ = stop payoff − best-continuation payoff; Theorem 1 decomposes path regret into weighted mistakes, and Theorem 2 shows the Bayes rule tracks sign(E[Δ|features]).

Load-bearing premise

Training labels and the main offline payoff rest on knowing each task’s minimal required tool set and treating exact coverage of that set as the value of success.

What would settle it

If, under verified required-set labels and heterogeneous costs, a tuned score-per-cost or sufficiency-threshold rule matched or beat CAM-DF’s mean payoff with confidence intervals covering zero across the Retail cost grid and the live agent no longer retained success while cutting tool exposure, the central claim would fail.

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

If this is right

  • Agent harnesses can add a frozen pre-execution stop layer on top of any existing router or retriever without fine-tuning the LLM.
  • Under heterogeneous tool prices, score-monotone acquisition is structurally inconsistent with optimal cost-aware subsets, so cost features must enter the stop decision.
  • Gains should widen when rankings are weaker, costs more dispersed, or cost pressure higher—the regimes the theory flags.
  • A compact ten-feature variant (CAM-DF-lite) can retain most of the full policy’s gains when interpretability or catalog stability matters.
  • Live deployment can cut exposed read tools by roughly a third while holding observed task success near full-access levels.

Where Pith is reading between the lines

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

  • The same stop-versus-continue gap could later be recomputed after each tool return, turning the one-shot prefix plugin into multi-round adaptive acquisition the paper flags as open.
  • If production systems only have noisy trajectory-mined required sets, the method’s edge may shrink toward simple cost heuristics unless labels are audited.
  • Harnesses that already log tool costs and outcomes could treat λ as an operator dial for latency or budget SLOs rather than a fixed research parameter.

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 / 6 minor

Summary. The paper formulates pre-execution tool acquisition for LLM agents as cost-aware stopping over ranked tool prefixes. Given an upstream ranking and heterogeneous tool costs, CAM-DF learns a stop/continue classifier from the offline payoff gap Δ(St)=U(At,x)−maxk>t U(Ak,x), using the sign as the label and |Δ| as a regret weight; CAM-DF-lite is a ten-feature interpretable variant. Theorems 1–2 give a pathwise prefix-regret decomposition and Bayes alignment of the weighted risk with sign(E[Δ|Xt]); Propositions 1–3 characterize the marginal value-per-cost rule and show score-monotone acquisition can be inconsistent under heterogeneous costs. Empirically, on τ-bench Retail and four additional domains (1,343 tasks), CAM-DF attains the highest mean payoff among deployable methods in the main Retail setting, with consistent gains over predict-then-threshold across ranking sources and cost regimes, and live execution reduces read-tool exposure by ~37% at comparable observed success.

Significance. If the results hold under the stated proxy, this is a useful modular control layer: it converts existing rankings into cost-aware acquisition depth without fine-tuning the LLM or changing the harness interface. The theory is standard cost-sensitive / decision-focused learning applied carefully rather than novel machinery, but the pathwise regret identity, Bayes alignment, and score-monotone failure propositions cleanly motivate the method and features. Strengths include extensive paired-bootstrap evaluation, multi-scorer robustness, ablations, cross-domain checks, live Retail execution with an independent simulator, full algorithmic appendices (A1–A2), and an interpretable lite variant with reported coefficients. The contribution is practical systems/ML for agent harnesses rather than a foundational theoretical breakthrough, but the problem framing (scores ≠ decisions under heterogeneous costs) is timely and well executed.

major comments (3)
  1. [Section 3, Eq. (2); Tables A5–A6, A19] Sec. 3, Eq. (2), and Algorithms A1–A2: both training targets and the primary offline payoff rest on annotated minimal required sets Gx under exact sufficiency U=1{Gx⊆A}−λ∑cj. Theorems 1–2 align the learner to this clairvoyant gap, so the headline payoff/SOTA numbers are only as meaningful as Gx as a proxy for real task value. Live Retail (Tables A5–A6) shows a strong directional signal (oracle 0.67 vs insufficient 0.24, p<0.001) and CAM-DF success 0.63–0.67, but n=67 is underpowered for non-inferiority (McNemar vs full p=0.21/0.73), some CAM-DF sets miss annotated required tools yet succeed, and insufficient still succeeds ~21–24%. Cross-domain labels are weaker (WorkBench trajectory consensus; Telecom dual-model shadow labels). Please (i) state explicitly in the abstract and Sec. 6 that offline SOTA is under the Gx-sufficiency proxy, (ii) quantify proxy–success disagreement more systema
  2. [Abstract; Section 6.1; Table 2; Figure 3; Tables A18, A22–A24] Abstract and Sec. 6.1 claim CAM-DF is “state-of-the-art under heterogeneous costs and high cost pressure” and attains the “highest payoff among deployable methods” on Retail. Table 2 and the dense grid (Fig. 3, Tables A18, A24) support consistent gains over predict-then-threshold and fixed-k, but the tuned score-per-cost threshold remains competitive at λ=0.12 (CIs often cover zero) and can lead or tie in parts of the grid and on Airline/MCP-Atlas partial-payoff cells (Tables A8, A22–A23). Please qualify the SOTA language to the baselines and regimes where paired CIs exclude zero, and avoid implying uniform dominance over all simple cost-aware heuristics.
  3. [Abstract; Section 6.3; Tables A5–A6; Conclusion] Live evaluation language (“maintaining comparable task success,” 37% fewer tools): Tables A5–A6 support reduced exposure (4.4 vs 7) with point estimates near full/oracle, but the paper correctly notes non-inferiority is not established. Soften abstract/conclusion wording to “no detectable drop in success in an underpowered n=67 check” (or similar), and flag that write/action tools remain unrestricted so the experiment isolates read-tool exposure only. One scorer/cost setting and ranked-prefix (not output-adaptive) scope should be stated next to the 37% figure.
minor comments (6)
  1. [Section 6; Table A2; Appendix D] Retail base costs (Table A2) are author-specified sensitivity weights, not measured telemetry. This is disclosed in the appendix; a one-sentence main-text note near the cost-regime discussion would help readers interpret heterogeneous-cost gains (the permutation check in App. D is reassuring and could be cited briefly).
  2. [Table A4; Section 6] Pipeline reconciliation (Table A4) shows nominally identical baselines differing slightly across released pipelines (e.g., predict-then-threshold 0.352 vs 0.361). Main-text tables should name the pipeline or state that all inference is within-pipeline paired, to avoid confusion when comparing across appendix tables.
  3. [Sections 3–4] Notation: U(A,x), v(A,Gx), and the exact/partial variants are clear, but Qstop/Qcontinue vs Ft and Δ(St) are introduced in close succession (Sec. 3–4); a small notation table or tighter cross-reference would help.
  4. [Figure 2] Figure 2 is effective; ensure the caption states λ=0.12, d=0 and that markers are mean tools vs mean payoff over tasks so readers do not read it as a causal frontier.
  5. [Section 2; Table A1] Related Work Table A1 is helpful; consider promoting a shortened version to the main text so positioning vs Stop-RAG / Adaptive-RAG / ranked-list truncation is visible without the supplement.
  6. [Abstract; Introduction] Typos/style: occasional missing spaces after commas in compound phrases in the abstract/intro (“agentharnessesface”, “rankingalonedoesnot”) appear to be PDF line-break artifacts; clean in the camera-ready.

Circularity Check

0 steps flagged

No load-bearing circular derivation: Bayes alignment and regret weights are standard cost-sensitive learning applied to an offline-constructed target; deployment features exclude Gx.

full rationale

The paper’s derivation chain is self-contained and does not reduce the claimed stopping rule or SOTA payoff to its inputs by construction. Theorem 1 is a pathwise identity decomposing prefix regret into weighted mistakes; Theorem 2 is the usual cost-sensitive Bayes boundary (sign E[Δ|X]) for regret-weighted 0–1 risk, citing Elkan and Elmachtoub–Grigas rather than a self-authored uniqueness theorem. Labels y*_t = 1{Δ(St)≥0} and weights |Δ| are built offline from annotated Gx and U (Eq. 2, Alg. A1), while the feature map ϕ and runtime gate (Alg. A2) explicitly exclude Gx, |Gx|, and missing-required indicators—ordinary supervised construction, not a self-definitional loop. Propositions 1–3 are structural characterizations (marginal value-per-cost; additive closed form; score-monotone failure under heterogeneous costs) with explicit counterexamples, not fitted parameters renamed as predictions. No load-bearing self-citation chain, imported uniqueness, or ansatz-via-citation appears. Train/val/test splits and live Retail execution (external success, not U) further separate learning from evaluation. The only mild proximity to circularity is that the primary offline metric is the same U used to build labels—standard for decision-focused work, not a forced identity between input and claimed result. Score 1 reflects that proximity without elevating it to a circular step.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 2 invented entities

The central claim rests on a decision model (exact or partial sufficiency minus λ-weighted costs), restriction to upstream score order, offline availability of required-set labels for training, and several operator-chosen scalars (λ, cost dispersion, regularization). Theory then uses standard probabilistic decision analysis rather than new physics-like entities. The largest external burdens are the Gx labeling model and author-specified heterogeneous costs, not hidden free constants inside a closed-form law.

free parameters (6)
  • cost pressure λ = main grids use 0.05, 0.12, 0.20
    Converts tool cost into task-value units and directly changes stop/continue labels; swept and operator-chosen rather than derived.
  • Retail base tool costs c̄_j = e.g. calculate 0.50 … get_product_details 2.00 before normalization
    Author-specified sensitivity vector, not measured production telemetry; heterogeneous-cost claims depend on it (with permutation checks).
  • cost dispersion d = main d ∈ {0,1.0,1.5}; dense grid to 2.0
    Controls spread around mean costs via c_j(d)=max{0.10,1+d(c̄_j/c̄_mean−1)}; experimental factor chosen by authors.
  • logistic regularization γ / C = C=1.0 in implementation notes
    ℓ2-regularized logistic CAM-DF fit hyperparameter.
  • weight floor ε = 1e-4
    Keeps zero-gap examples in training; small but ad hoc.
  • CAM-DF-lite / baseline validation thresholds = grids of ~19–30 thresholds; lite medians often <0.5
    Several competitors and CAM-DF-lite stop thresholds are tuned on validation payoff; full CAM-DF uses fixed 0.5.
axioms (6)
  • domain assumption Task value is exact set-sufficiency 1{Gx⊆A} (or partial coverage |A∩Gx|/|Gx|) with linear cost penalty λ∑cj.
    Section 3 payoff definition; main tables use exact sufficiency.
  • domain assumption Acquisition is restricted to prefixes of a fixed upstream score ranking; arbitrary subsets are out of scope at decision time.
    Section 3 stopping-along-ranking formulation; reduces 2^m subsets to m+1 nested decisions.
  • domain assumption Required set Gx is latent at deployment but available offline on logged/benchmark tasks to build Δ labels and metrics.
    Section 3 observable-state paragraph and training recipe.
  • standard math Bayes-optimal rule for |Δ|-weighted 0-1 loss is sign(E[Δ|X]); logistic surrogate preserves the boundary under standard calibration conditions.
    Theorem 2 and Lemma 2; classical cost-sensitive / surrogate classification results.
  • standard math Under additive expected values, optimal unconstrained acquisition keeps tools with pj/cj≥λ; score-monotone rules can fail when value order and value-per-cost order disagree.
    Propositions 2–3; used to motivate cost features and score-only suboptimality.
  • domain assumption Frozen LLM necessity scores are a valid public ranking interface even if uncalibrated as membership probabilities.
    Section 3 scoring interface and experimental ranking sources.
invented entities (2)
  • CAM-DF / CAM-DF-lite stopping policies independent evidence
    purpose: Deployable classifiers that map public prefix features to stop/continue along a tool ranking.
    Method objects, not physical entities; empirically tested plugins rather than postulated hidden mechanisms.
  • Decision gap Δ(St)=Qstop−Qcontinue over ranked prefixes no independent evidence
    purpose: Defines offline stop labels and regret weights for decision-focused training.
    Constructed from the paper’s payoff model; standard in stopping/regret form, specialized here to tool prefixes.

pith-pipeline@v1.2.0-daily-grok45 · 39421 in / 3881 out tokens · 75203 ms · 2026-07-30T11:38:50.178068+00:00 · methodology

0 comments
read the original abstract

As LLM agents increasingly depend on diverse external services such as search engines, databases, and connectors, agent harnesses face a fundamental tool-selection challenge: acquiring too few tools leaves the task under-informed, while too many adds cost, context load, and privacy exposure. Routers and retrievers can rank candidate tools by relevance, but a ranking alone does not determine how many are worth selecting. Existing approaches leave acquisition under heterogeneous costs unaddressed. We formulate this decision as cost-aware marginal decision-focused stopping (CAM-DF) over ranked tool prefixes, with CAM-DF-lite as a compact interpretable variant. We train directly on the offline gap between stopping now and the best continuation: its sign labels the decision, its magnitude weights each error by the payoff at stake. We prove this objective is Bayes-aligned with the stopping target and that score-only rules are suboptimal under heterogeneous costs. We evaluate on 1,343 tasks across five tool-use domains. On $\tau$-bench Retail, CAM-DF attains the highest payoff among deployable methods, with gains over a predict-then-threshold baseline across all five ranking sources and two cost regimes. Our approach is state-of-the-art under heterogeneous costs and high cost pressure, with larger gains under weaker rankings. In live execution, CAM-DF exposes the agent to 37\% fewer tools than full access while maintaining comparable task success. The CAM-DF family is a lightweight pre-execution plugin that turns existing tool rankings into lower-cost acquisition decisions without fine-tuning the underlying LLM.

Figures

Figures reproduced from arXiv: 2607.27083 by Wei Qi, Yan Cheng, Yan Zhang, Yicheng Feng.

Figure 1
Figure 1. Figure 1: CAM-DF as a pre-execution stop plugin inside one agent turn. The agent flow is unchanged: a router, retriever, scoring [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Payoff versus tool acquisition on Retail ( [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Retail payoff gains of CAM-DF over (a) Tuned [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] 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

54 extracted references · 1 canonical work pages · 1 internal anchor

  1. [1]

    International Conference on Learning Representations (ICLR) , year =

    -bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains , author =. International Conference on Learning Representations (ICLR) , year =

  2. [2]

    arXiv preprint arXiv:2602.00933 , year =

    MCP-Atlas: A Large-Scale Benchmark for Tool-Use Competency with Real MCP Servers , author =. arXiv preprint arXiv:2602.00933 , year =

    Pith/arXiv arXiv

  3. [3]

    arXiv preprint arXiv:2511.02734 , year =

    CostBench: Evaluating Multi-Turn Cost-Optimal Planning and Adaptation in Dynamic Environments for LLM Tool-Use Agents , author =. arXiv preprint arXiv:2511.02734 , year =

    Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2405.00823 , year =

    WorkBench: a Benchmark Dataset for Agents in a Realistic Workplace Setting , author =. arXiv preprint arXiv:2405.00823 , year =

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2506.07982 , year =

    ^2 -Bench: Evaluating Conversational Agents in a Dual-Control Environment , author =. arXiv preprint arXiv:2506.07982 , year =

    Pith/arXiv arXiv

  6. [6]

    International Conference on Learning Representations (ICLR) , year =

    MetaTool Benchmark for Large Language Models: Deciding Whether to Use Tools and Which to Use , author =. International Conference on Learning Representations (ICLR) , year =

  7. [7]

    arXiv preprint arXiv:2407.01502 , year =

    AI Agents That Matter , author =. arXiv preprint arXiv:2407.01502 , year =

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2411.01643 , year =

    EcoAct: Economic Agent Determines When to Register What Action , author =. arXiv preprint arXiv:2411.01643 , year =

    Pith/arXiv arXiv

  9. [9]

    arXiv preprint arXiv:2605.00737 , year =

    To Call or Not to Call: A Framework to Assess and Optimize LLM Tool Calling , author =. arXiv preprint arXiv:2605.00737 , year =

    Pith/arXiv arXiv

  10. [10]

    arXiv preprint arXiv:2605.09252 , year =

    LLM Agents Already Know When to Call Tools Even Without Reasoning , author =. arXiv preprint arXiv:2605.09252 , year =

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2603.19896 , year =

    Utility-Guided Agent Orchestration for Efficient LLM Tool Use , author =. arXiv preprint arXiv:2603.19896 , year =

    arXiv

  12. [12]

    arXiv preprint arXiv:2511.17006 , year =

    Budget-Aware Tool-Use Enables Effective Agent Scaling , author =. arXiv preprint arXiv:2511.17006 , year =

    arXiv

  13. [13]

    Proceedings of NAACL , year =

    When2Call: When (not) to Call Tools , author =. Proceedings of NAACL , year =

  14. [14]

    Findings of ACL , year =

    Budget-Constrained Tool Learning with Planning , author =. Findings of ACL , year =

  15. [15]

    IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    CATP-LLM: Empowering Large Language Models for Cost-Aware Tool Planning , author =. IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  16. [16]

    arXiv preprint arXiv:2605.24660 , year =

    How Many Tools Should an LLM Agent See? A Chance-Corrected Answer , author =. arXiv preprint arXiv:2605.24660 , year =

    Pith/arXiv arXiv

  17. [17]

    Proceedings of SIGIR , year =

    Choppy: Cut Transformer for Ranked List Truncation , author =. Proceedings of SIGIR , year =

  18. [18]

    Proceedings of SIGIR , year =

    Ranked List Truncation for Large Language Model-based Re-Ranking , author =. Proceedings of SIGIR , year =

  19. [19]

    International Conference on Learning Representations (ICLR) , year =

    ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs , author =. International Conference on Learning Representations (ICLR) , year =

  20. [20]

    arXiv preprint arXiv:2402.04253 , year =

    AnyTool: Self-Reflective, Hierarchical Agents for Large-Scale API Calls , author =. arXiv preprint arXiv:2402.04253 , year =

    Pith/arXiv arXiv

  21. [21]

    arXiv preprint arXiv:2305.06983 , year =

    Active Retrieval Augmented Generation , author =. arXiv preprint arXiv:2305.06983 , year =

    Pith/arXiv arXiv

  22. [22]

    International Conference on Learning Representations (ICLR) , year =

    Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection , author =. International Conference on Learning Representations (ICLR) , year =

  23. [23]

    NAACL , year =

    Adaptive-RAG: Learning to Adapt Retrieval-Augmented Large Language Models through Question Complexity , author =. NAACL , year =

  24. [24]

    ACL , year =

    DRAGIN: Dynamic Retrieval Augmented Generation based on the Real-time Information Needs of Large Language Models , author =. ACL , year =

  25. [25]

    arXiv preprint arXiv:2510.14337 , year =

    Stop-RAG: Value-Based Retrieval Control for Iterative RAG , author =. arXiv preprint arXiv:2510.14337 , year =

    arXiv

  26. [26]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    DynamicRAG: Leveraging Outputs of Large Language Model as Feedback for Dynamic Reranking in Retrieval-Augmented Generation , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  27. [27]

    Econometrica , volume =

    Optimal Search for the Best Alternative , author =. Econometrica , volume =

  28. [28]

    ACM SIGecom Exchanges , volume =

    Recent Developments in Pandora's Box Problem: Variants and Applications , author =. ACM SIGecom Exchanges , volume =. 2023 , note =

    2023

  29. [29]

    Proceedings of the AAAI Conference on Artificial Intelligence , year =

    Contextual Pandora's Box , author =. Proceedings of the AAAI Conference on Artificial Intelligence , year =

  30. [30]

    arXiv preprint arXiv:2510.01394 , year =

    Optimal Stopping vs Best-of- N for Inference Time Optimization , author =. arXiv preprint arXiv:2510.01394 , year =

    arXiv

  31. [31]

    arXiv preprint arXiv:2507.20501 , year =

    Post-estimation Adjustments in Data-driven Decision-making with Applications in Pricing , author =. arXiv preprint arXiv:2507.20501 , year =

    Pith/arXiv arXiv

  32. [32]

    Management Science , volume =

    Smart ``Predict, then Optimize'' , author =. Management Science , volume =. 2022 , doi =

    2022

  33. [33]

    Journal of Artificial Intelligence Research , volume =

    Decision-Focused Learning: Foundations, State of the Art, Benchmark and Future Opportunities , author =. Journal of Artificial Intelligence Research , volume =. 2024 , doi =

    2024

  34. [34]

    Proceedings of the AAAI Conference on Artificial Intelligence , year =

    Classification with Costly Features using Deep Reinforcement Learning , author =. Proceedings of the AAAI Conference on Artificial Intelligence , year =

  35. [35]

    International Conference on Machine Learning (ICML) , year =

    Learning to Maximize Mutual Information for Dynamic Feature Selection , author =. International Conference on Machine Learning (ICML) , year =

  36. [36]

    arXiv preprint arXiv:2502.08235 , year =

    The Danger of Overthinking: Examining the Reasoning-Action Dilemma in Agentic Tasks , author =. arXiv preprint arXiv:2502.08235 , year =

    Pith/arXiv arXiv

  37. [37]

    IJCAI , year =

    The Foundations of Cost-Sensitive Learning , author =. IJCAI , year =

  38. [38]

    Journal of the American Statistical Association , volume =

    Convexity, Classification, and Risk Bounds , author =. Journal of the American Statistical Association , volume =

  39. [39]

    The Econometrics Journal , volume =

    Double/Debiased Machine Learning for Treatment and Structural Parameters , author =. The Econometrics Journal , volume =

  40. [40]

    Electronic Journal of Statistics , volume =

    Calibrated Asymmetric Surrogate Losses , author =. Electronic Journal of Statistics , volume =

  41. [41]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

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

  42. [42]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Gorilla: Large Language Model Connected with Massive APIs , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  43. [43]

    Liu, Grace and Qu, Yuxiao and Schneider, Jeff and Singh, Aarti and Kumar, Aviral , journal =

  44. [44]

    Calibrate-Then-Act: Cost-Aware Exploration in

    Ding, Wenxuan and Tomlin, Nicholas and Durrett, Greg , journal =. Calibrate-Then-Act: Cost-Aware Exploration in

  45. [45]

    arXiv preprint arXiv:2602.11541 , year =

    Budget-Constrained Agentic Large Language Models: Intention-Based Planning for Costly Tool Use , author =. arXiv preprint arXiv:2602.11541 , year =

    arXiv

  46. [46]

    Acting Less is Reasoning More!

    Wang, Hongru and Qian, Cheng and Zhong, Wanjun and Chen, Xiusi and Qiu, Jiahao and Huang, Shijue and Jin, Bowen and Wang, Mengdi and Wong, Kam-Fai and Ji, Heng , journal =. Acting Less is Reasoning More!

  47. [47]

    Findings of the Association for Computational Linguistics: ACL 2025 , year =

    Qian, Cheng and Acikgoz, Emre Can and Wang, Hongru and Chen, Xiusi and Sil, Avirup and Hakkani-T. Findings of the Association for Computational Linguistics: ACL 2025 , year =

    2025

  48. [48]

    Kieback, Adrian and Amadasun, Uyiosa Philip and Chadha, Aman and Elkins, Aaron , journal =

  49. [49]

    Operations Research , volume =

    Ban, Gah-Yi and Rudin, Cynthia , title =. Operations Research , volume =. 2019 , doi =

    2019

  50. [50]

    Operations Research , volume =

    Bertsimas, Dimitris and Koduri, Nihal , title =. Operations Research , volume =. 2022 , doi =

    2022

  51. [51]

    Zico , title =

    Donti, Priya and Amos, Brandon and Kolter, J. Zico , title =. Advances in Neural Information Processing Systems , volume =. 2017 , url =

    2017

  52. [52]

    Mathematics of Operations Research , volume =

    Boodaghians, Shant and Fusco, Federico and Lazos, Philip and Leonardi, Stefano , title =. Mathematics of Operations Research , volume =. 2023 , doi =

    2023

  53. [53]

    Journal of Economic Theory , volume =

    Olszewski, Wojciech and Weber, Richard , title =. Journal of Economic Theory , volume =. 2015 , doi =

    2015

  54. [54]

    Online Pandora's Box for Contextual LLM Cascading

    Belloni, Alexandre and Chen, Yan and Wei, Yehua , title =. 2026 , eprint =. doi:10.48550/arXiv.2606.07392 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2606.07392 2026