Pith. sign in

REVIEW 4 major objections 5 minor 40 references

To keep language-model agents from gaming tool retrieval by recycling sub-tasks, PCTD rewards only the marginal retrieval gain of decomposing a request, measured against the raw query as a paired baseline.

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 ↗

A counterfactual difference reward plus a preference reward reduces reward hacking in LLM task decomposition and improves tool retrieval, especially for unseen tools.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Solid empirical package with a useful new benchmark, but the counterfactual causal claim doesn't survive the implementation: Rcf is a margin reward, not a do-intervention. the 4 major comments →

arxiv 2607.15696 v1 pith:SO2PEAMV submitted 2026-07-17 cs.IR

PCTD: Preference-Guided Counterfactual Task Decomposition for Agent Tool Retrieval

Chu Zhao , Lei Tang , Minghang Li , Jianzhe Zhao , Guibing Guo , Zhengzong Chen , Yuanyuan Zhao , Fei Huang This is my paper
classification cs.IR
keywords task decompositiontool retrievalcounterfactual rewardpreference rewardreward hackingreinforcement learningmulti-turn dialogueLLM agents
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.

The reading

The paper claims that the standard reinforcement-learning practice of rewarding an agent with tool-retrieval scores (like NDCG) for its task decompositions is self-defeating: models learn to pump up the score by repeating or padding sub-tasks, a shallow pattern that collapses on unseen tools. The proposed fix is a counterfactual reward that compares the retrieval list obtained from the decomposed request with the list obtained from the raw, undecomposed request, and rewards only the difference. Because both branches share the same retriever, tool library, and ground-truth set, the authors argue that confounding factors such as keyword priors and tool popularity are subtracted out, leaving the true marginal contribution of decomposition. This reward is combined with a preference reward that scores decompositions against human references on completeness, accuracy, coreference resolution, standardized expression, and contextual consistency. If the approach holds, agents that decompose user requests will retrieve more relevant tools and, crucially, keep their advantage when the set of tools is new.

Core claim

The central claim is that the marginal retrieval gain of task decomposition can be isolated by differencing retrieval quality between a branch that retrieves with atomic sub-tasks and a branch that retrieves with the raw context. Concretely, for a given multi-turn input, the paper computes NDCG@K over the merged list obtained by retrieving with each atomic sub-task, and subtracts NDCG@K over the list obtained by retrieving with the raw context. The positive part of this difference, scaled by a rank term and a coverage term, forms the counterfactual reward Rcf = λ_rank max(0, Δ_rank) + λ_cov Δ_cov. The paper argues that this reward 'cuts off spurious correlations at their source,' so a policy

What carries the argument

The central object is the paired counterfactual reward Rcf. It treats the raw query as the control branch and the decomposition branch as the treatment: Δ_rank = NDCG@K(decomposition) − NDCG@K(raw) and Δ_cov = indicator(full tool set covered by decomposition) − indicator(covered by raw). The reward is λ_rank max(0, Δ_rank) + λ_cov Δ_cov, which gives the policy credit only for retrieval gains that the decomposition adds over the raw query. The design rests on a structural-causal-model argument that confounders (keyword priors, tool popularity, parametric memory) are shared by the two branches and therefore cancel in the difference.

Load-bearing premise

The load-bearing premise is that the raw-query branch and the decomposition branch share all nuisance factors (keyword priors, tool popularity, parametric memory), so subtracting their retrieval scores isolates the causal effect of decomposition; but since the decomposition is generated by the policy from the same input, the 'do-intervention' is never actually performed, and the observed difference may still be influenced by how the policy's generation correlates with those n

What would settle it

At preference weight zero, the model is trained only on the counterfactual reward, and the paper reports the duplication rate rises to 5.24% while out-of-domain NDCG@10 drops to 78.6. This is a direct probe: if Rcf truly eliminated spurious correlations, a policy optimizing it alone should have no incentive to repeat sub-tasks. A reader can check whether duplicating an atomic sub-task actually increases Δ_rank (e.g., because truncation of the merged list preserves a useful tool); if it does, the differencing fails to remove the spurious path.

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

If this is right

  • If the counterfactual reward truly removes spurious correlations, RL-based task decomposition should generalize to unseen tools without the usual out-of-domain collapse, because the policy is rewarded for causal gain rather than lexical overlap.
  • Combining the counterfactual reward with a preference reward should reduce repetitive decomposition (the paper reports duplication rate dropping from 5.2% to 0.7%) without sacrificing retrieval quality.
  • The state-machine-generated benchmark provides process-level annotations (dialogue state evolution, intent rewriting, atomic task sequences) that enable fine-grained evaluation of decomposition quality in multi-turn mobile interactions.
  • The method should improve retrieval across different embedding retriever backbones, since the decomposition is optimized at the input end and carries high signal-to-noise ratio.
  • A proper balance between the two rewards is necessary: pure counterfactual rewards still allow reward hacking (repetition rate 5.24% at zero preference weight), while pure preference rewards over-regularize and drop retrieval accuracy.

Where Pith is reading between the lines

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

  • The paired-differencing logic is not a true do-intervention: the decomposition is generated by the policy from the same input, so nuisance factors may influence the decomposition itself, and the computed difference is an observational contrast between two query transformations rather than a demonstrated causal estimate.
  • The same differencing idea (rewarding only the delta over a control transformation) could apply to other agent sub-tasks beyond decomposition — query rewriting, summarization for retrieval, or planning steps — wherever an absolute metric can be gamed.
  • A testable extension: compare against a 'superficial transformation' baseline (e.g., the raw query rephrased without decomposition) to disentangle whether gains come from decomposition structure or simply from longer queries; if gains persist, the causal story is strengthened.
  • The benchmark pipeline could be reused to generate process-level annotations for other agent benchmarks, potentially standardizing decomposition-quality measurement across multi-turn interactions.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 5 minor

Summary. The paper proposes PCTD, a GRPO-based framework for task decomposition in multi-turn agent tool retrieval. The method trains a policy to rewrite/decompose user queries into atomic subtasks; rewards combine a preference reward from a trained PRM comparing candidate decompositions to human references, and a 'counterfactual' reward Rcf defined as a paired comparison of NDCG@K/coverage between the decomposition branch and a raw-query branch. The authors also introduce MTDTool, a state-machine-generated benchmark for mobile multi-turn interactions, and report consistent improvements over closed-source, prompting, and RL baselines on ToolRet and MTDTool, with ablations and hyperparameter analysis.

Significance. If the empirical results hold, the paper contributes a useful reward-shaping recipe: using the raw-query retrieval score as a per-prompt reference within a thresholded-margin reward, combined with a PRM-based preference reward, appears to improve retrieval and reduce repetitive decomposition in the reported settings. The paper is refreshingly transparent: it releases code, reports failure modes in Table 8/Section 4.6, and includes multi-retriever and dataset-difficulty analyses. However, the significance of the central conceptual contribution is currently overstated. The evidence supports a pragmatic reward-design claim, not the causal claim that the reward 'cuts off spurious correlations at their source' via a do-intervention. The paper's own Table 8 shows that the pure Rcf variant still reward-hacks, so the causal mechanism is not demonstrated.

major comments (4)
  1. [§3.2.1, Eq. (9), Algorithm 1] Eq. (9) defines Δ(D;X) = E[Y|do(D=δ),X] − E[Y|do(D=δ0),X] and the text states this is estimated by a paired-counterfactual comparison. In the actual implementation (Algorithm 1, lines 2–7), D is sampled from π_θold(·|x_t); no intervention or do-calculus is applied to any variable, and no identification argument (e.g., back-door/front-door adjustment using the SCM in Fig. 3) links the observed NDCG values to the do-contrast. At best, Rcf is an observational conditional contrast between two query transformations. The causal language should be removed or justified with a genuine intervention and an explicit identification proof.
  2. [§3.2.1, Eq. (16) and GRPO (Eq. 2)] Per prompt x_t, the raw-query quantities N_raw and I_raw in Eq. (16) are constants across the B GRPO candidates. Group-wise standardization in Eq. (2)/Algorithm 1 line 11 subtracts the group mean, so the additive constant −N_raw (and −I_raw) cancels; the only remaining effect of the baseline is the nonlinear threshold max(0,·) and the λ_cov coverage-difference term. Consequently, Rcf is best described as a thresholded retrieval-gain reward, not a controlled causal contrast. The distinction matters because the paper's wording in §3.2.1 and Fig. 3 claims a 'do-intervention' and 'subtracted out' confounders; neither is achieved by the implemented equations.
  3. [§4.6, Table 8, Appendix E.1] The paper's own failure analysis contradicts the central 'cutting off spurious correlations at their source' claim. At w_pref=0, where Rcf is used alone, the duplication rate is 5.24% and OOD N@10 drops to 78.59 (from 82.74 in the full model), which is exactly the repetitive-decomposition reward hacking the method claims to prevent. The text in §4.6 concedes that 'pure causal optimization without preference constraints still induces reward hacking.' This means the observed benefits should be attributed to the joint reward including the PRM preference reward and the coverage term, not to the counterfactual mechanism per se. A revised paper should reframe the contribution as an empirical reward-design study and temper the causal claims accordingly.
  4. [§4.1/§4.3, Tables 1–2] The causal interpretation is further weakened by the evaluation design. The OOD test set is generated by the same state-machine pipeline (unseen tool verticals in MTDTool, and held-out ToolRet domains), but no analysis separates 'true out-of-tool generalization' from distributional overlap in the synthetic generation. More importantly, since Rcf is defined on the same NDCG@K metric used for evaluation, improvements in N@10/C@10 are partly aligned with the reward by construction; this does not invalidate the empirical gains, but it does undermine the claim that Rcf provides an independent 'prediction' of causal benefit. Please provide an ablation comparing Rcf against a non-causal reward with the same marginal statistics (e.g., N_atom alone) to isolate the claimed effect.
minor comments (5)
  1. [Fig. 3] The figure labels the framework as 'Intervention Implementation' with a do-intervention, but the implementation contains no intervention. Please relabel to 'paired comparison' or provide actual intervention code.
  2. [Algorithm 1] The algorithm notation mixes PPO and GRPO terms (e.g., 'PPO mini-batch' in Appendix B vs. GRPO in Eq. (2)). Please harmonize notation and clarify the optimizer details.
  3. [Appendix E.2, Table 9] The 'Fine-tuned Qwen3-Embedding-4B' row shows In-Domain N@10 dropping by 0.06% (93.37 → 93.31). The text describes this as 'remains competitive,' which is fair, but the slight drop should be explicitly noted in the main text if this row is used to claim retriever-agnostic robustness.
  4. [References] Several citations appear with inconsistent page ranges or venue formatting (e.g., Verma et al. 2024a has no venue; some arXiv references lack IDs). Please standardize.
  5. [Table 6] The header lists 'APIBench' and 'API-Bank' as two columns while the text refers to 'API-Bank'; please reconcile the naming and check whether the checkmark columns align with the text.

Circularity Check

2 steps flagged

The counterfactual reward is defined directly from the NDCG@K evaluation metric; the raw-query 'baseline' is a per-prompt threshold that does not estimate any confounder, so the claimed causal mechanism reduces by construction to a thresholded retrieval reward—and the paper's own Table 8 shows Rcf alone reproduces the reward hacking it claims to prevent.

specific steps
  1. self definitional [Sec. 3.2.1, Eqs. (10)-(16); Algorithm 1 lines 4-11]
    "In practice, we instantiate the potential outcome Y with the retrieval ranking metric NDCG@K... L^K_raw = R_K(x), N_raw = NDCG@K(L^K_raw,G_t) ... N_atom = NDCG@K(L^K_atom,G_t) ... Δ_rank = N_atom − N_raw ... R_cf = λ_rank max(0,Δ_rank) + λ_cov Δ_cov. ... GRPO advantage: normalize within the n-candidate group, Â(j) ← R(j) − 1/n Σ R(k) / std({R(k)}) + ε_std."

    For a fixed prompt x_t, the raw branch values N_raw and I_raw (Eqs. 10, 15) are identical for every candidate decomposition. Thus Eq. 16 is by construction λ_rank·max(0,N_atom−c)+λ_cov·I_atom−c′, with c=N_raw and c′=λ_cov·I_raw fixed. The 'counterfactual subtraction' does not measure or adjust Z (keywords, tool priors, parametric memory); it only applies a per-prompt threshold to the same NDCG@K that is later used as the evaluation metric. No do(D=δ) vs do(D=δ0) intervention is executed in Algorithm 1—D is sampled from π_θold(·|x_t)—so Eq. 9's causal contrast is never realized; the implementable reward is a thresholded outcome-metric reward, exactly the class of reward the paper says induces spurious correlations.

  2. other [Sec. 4.6, 'When might PCTD fail?'; Table 8 / Figure 4]
    "When the preference reward weight w_pref = 0, the model relies solely on counterfactual signals for optimization. Although it maintains high retrieval accuracy In-Domain (N@10 = 90.6), the duplication rate surges to 5.24%, and OOD performance drops sharply to 78.6. This confirms that pure causal optimization without preference constraints still induces reward hacking, resulting in generalization failure."

    This is the paper's own controlled test of the advertised mechanism. The counterfactual-only setting (w_pref=0) is precisely Rcf from Eq. 16; it exhibits the same repetitive-decomposition reward hacking and OOD degradation that the paper attributes to directly using NDCG/Recall as rewards. The claimed 'cutting off spurious correlations at their source' therefore has no observable content beyond the thresholded NDCG term: the anti-reward-hacking and OOD gains are actually carried by the separately anchored preference reward R_pref, not by the counterfactual subtraction. The central causal claim is thus equivalent to—and contradicted by—the paper's own ablation.

full rationale

The score is not higher because the paper contains independent, non-circular content: the preference reward R_pref (Eq. 17) is anchored to a separately trained PRM and reference decomposition; the coverage term Δ_cov adds signal beyond NDCG; and ToolRet is an external benchmark whose OOD results provide a check outside the self-generated MTDTool pipeline. No load-bearing self-citation chain was found, and no uniqueness theorem is imported from the authors' prior work. The circular core is the counterfactual mechanism itself: the paper defines its causal reward as a difference of the same NDCG@K used for evaluation, and the raw-query baseline is a fixed per-prompt threshold rather than an estimated confounder. Algorithm 1 samples D from the policy rather than intervening, so Eq. 9's do-contrast is a renaming of an observational NDCG contrast. The paper's own Table 8 confirms that Rcf alone does not cut off spurious correlations; the reported OOD robustness is largely attributable to the preference reward. The MTDTool benchmark's auto-generated references and OOD splits are also self-referential, but the external ToolRet results prevent this from being the dominant issue.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The method rests on hand-chosen reward weights (w_pref, lambdas), dataset-generation parameters (p0, alpha, pmin), and a trained-but-unspecified PRM. No new physical or structural entities are introduced. The main epistemological load is carried by the paired-differencing causal assumption, which is asserted rather than derived.

free parameters (5)
  • w_pref = 0.5
    Fusion weight between preference and counterfactual rewards in Eq. 18; chosen by hand after a hyperparameter sweep (Table 8).
  • lambda_rank = 1.0
    Coefficient for the NDCG-gain term in Rcf (Eq. 16); set to 1.0 unless otherwise specified.
  • lambda_cov = 1.0
    Coefficient for the coverage-gain term in Rcf (Eq. 16); set to 1.0 unless otherwise specified.
  • p0, alpha, pmin = not reported numerically
    Topic-switch probability parameters in Eq. 7 of the dataset state machine; they shape the MTDTool distribution and therefore affect both training and evaluation difficulty.
  • PRM parameters (s_phi) = fitted, unspecified
    The Process Reward Model used in Eq. 17 is a trained Qwen2.5-1.5B model, but its training data, labels, and procedure are not described.
axioms (5)
  • standard math GRPO provides a valid policy optimization update for the decomposition policy.
    Used in Eq. 2 and Algorithm 1; standard RL machinery with clipping and KL penalty.
  • domain assumption Paired differencing of NDCG between decomposition and raw-query branches identifies the causal effect of decomposition.
    Invoked in Eq. 9 and §3.2.1; this is the paper's central causal claim and is not proven by the implemented difference-based reward.
  • domain assumption The retriever RK is deterministic and shared, so nuisance factors cancel by differencing.
    Assumed in §3.2.1; the same retriever is used for both branches, but the decomposition branch changes the query text, so the difference mixes query-rewriting effects with tool-prior effects.
  • domain assumption MTDTool's state-machine-generated dialogues are representative of real mobile multi-turn tool use.
    The benchmark is built by an LLM pipeline with hand-coded constraints (§3.1); no human-validated real-world distribution is provided.
  • domain assumption The LLM-generated reference decomposition y* is a valid gold standard for preference scoring.
    Used in Eq. 17 and in the GSB evaluation; the reference is generated by the same pipeline that creates the dialogues, so it may encode the generator's biases.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of PCTD: Preference-Guided Counterfactual Task Decomposition for Agent Tool Retrieval." pith.science (2026). https://pith.science/paper/SO2PEAMV

@misc{pith2026260715696,
  author       = {Pith},
  title        = {Pith review of: PCTD: Preference-Guided Counterfactual Task Decomposition for Agent Tool Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SO2PEAMV}},
  note         = {Machine review of arXiv:2607.15696}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Task decomposition aims to transform ambiguous instructions into executable atomic subtasks, thereby guiding high-precision tool retrieval. However, our analysis reveals that directly adopting tool retrieval metrics, i.e., Recall or NDCG, as rewards for task decomposition can easily induce reward hacking in reinforcement learning-based methods. Specifically, models tend to maximize retrieval matching through strategies such as repetitive decomposition. This spurious correlation between the shallow features of decomposition results and retrieval metric impairs generalization in Out-of-Domain (OOD) scenarios involving unseen tools. To address this issue, we propose PCTD, a Preference-guided Counterfactual Task Decomposition framework. PCTD quantifies the marginal causal gain of decomposition on retrieval ranking through a counterfactual reward, thereby cutting off spurious correlations at their source. Meanwhile, it introduces a preference reward to impose fine-grained structural supervision on logical coherence and atomicity, encouraging the model to generate high-quality decompositions. In addition, we construct MTDTool, the task decomposition benchmark specifically designed for mobile multi-turn interactions. Extensive experiments demonstrate that PCTD alleviates repetitive decomposition and surpasses SOTA methods in retrieval, decomposition quality, and OOD generalization.

Figures

Figures reproduced from arXiv: 2607.15696 by Chu Zhao, Fei Huang, Guibing Guo, Jianzhe Zhao, Lei Tang, Minghang Li, Yuanyuan Zhao, Zhengzong Chen.

Figure 1
Figure 1. Figure 1: Overall performance and empirical study results. Middle: Performance of different models on [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the MTDTool construction pipeline. The data generation follows a four-stage process. (1) Initialization: Tools and scenarios are sampled to establish initial constraints, and two LLM generators produce the initial user query q 0 and assistant output o 0 . (2) State Evolution & Context Decision: At each turn t ≥ 1, the system decides whether to inherit the previous context or switch to a new top… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the PCTD. 3.2 Joint Reward Mechanism In the scenario of contextual task atomization, the efficacy of model-generated decomposition results should not be determined solely by superficial textual rationality, but rather by their actual contribution to downstream tool retrieval performance and the semantic quality of the decomposition structure itself. To this end, we propose a joint reward mechan… view at source ↗
Figure 4
Figure 4. Figure 4: Hyperparameter analysis under different wpre f settings. without the guidance of marginal causal gain, the model degenerates into shallow matching that relies on keyword co-occurrence. Ultimately, the complete PCTD achieves optimal performance in both retrieval accuracy and generalization capability by synergizing the functional correction of counterfactual rewards with the structural constraints of prefer… view at source ↗
Figure 5
Figure 5. Figure 5: Training curves of different metrics. 4.6 Deeper Analysis When might PCTD fail? We further analyze that improper reward weight settings can lead to the failure of PCTD. As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Dataset distribution (grouped percentage). [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Overall dataset distribution [PITH_FULL_IMAGE:figures/full_fig_p019_7.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

40 extracted references · 8 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2402.02716 , year=

    Understanding the planning of llm agents: A survey , author=. arXiv preprint arXiv:2402.02716 , year=

    Pith/arXiv arXiv

  2. [2]

    arXiv preprint arXiv:2503.16416 , year=

    Survey on evaluation of llm-based agents , author=. arXiv preprint arXiv:2503.16416 , year=

    Pith/arXiv arXiv

  3. [3]

    arXiv preprint arXiv:2210.03629 , year=

    React: Synergizing reasoning and acting in language models , author=. arXiv preprint arXiv:2210.03629 , year=

    Pith/arXiv arXiv

  4. [4]

    Proceedings of the 3rd Workshop on Towards Knowledgeable Foundation Models (KnowFM) , pages=

    ToolReAGt: tool retrieval for LLM-based complex task solution via retrieval augmented generation , author=. Proceedings of the 3rd Workshop on Towards Knowledgeable Foundation Models (KnowFM) , pages=

  5. [5]

    Advances in neural information processing systems , volume=

    Toolformer: Language models can teach themselves to use tools , author=. Advances in neural information processing systems , volume=

  6. [6]

    Advances in Neural Information Processing Systems , volume=

    Gorilla: Large language model connected with massive apis , author=. Advances in Neural Information Processing Systems , volume=

  7. [7]

    International Conference on Learning Representations , volume=

    Toolllm: Facilitating large language models to master 16000+ real-world apis , author=. International Conference on Learning Representations , volume=

  8. [8]

    Data Science and Engineering , pages=

    Retrieval-augmented generation for ai-generated content: A survey , author=. Data Science and Engineering , pages=. 2026 , publisher=

    2026

  9. [9]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Llm collaboration with multi-agent reinforcement learning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  10. [10]

    Nature , volume=

    DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning , author=. Nature , volume=. 2025 , publisher=

    2025

  11. [11]

    Plan-RAG: Planning-guided Retrieval Augmented Generation , author=

  12. [12]

    arXiv preprint arXiv:2601.07782 , year=

    Beyond Single-Shot: Multi-step Tool Retrieval via Query Planning , author=. arXiv preprint arXiv:2601.07782 , year=

    arXiv

  13. [13]

    Advances in Neural Information Processing Systems , volume=

    Toolrl: Reward is all tool learning needs , author=. Advances in Neural Information Processing Systems , volume=

  14. [14]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track , pages=

    Encouraging good processes without the need for good answers: Reinforcement learning for llm agent planning , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track , pages=

    2025

  15. [15]

    Advances in Neural Information Processing Systems , volume=

    Group-in-group policy optimization for llm agent training , author=. Advances in Neural Information Processing Systems , volume=

  16. [16]

    Advances in Neural Information Processing Systems , volume=

    Taskbench: Benchmarking large language models for task automation , author=. Advances in Neural Information Processing Systems , volume=

  17. [17]

    Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    Api-bank: A comprehensive benchmark for tool-augmented llms , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    2023

  18. [18]

    International Conference on Learning Representations , volume=

    Mint: Evaluating llms in multi-turn interaction with tools and language feedback , author=. International Conference on Learning Representations , volume=

  19. [19]

    arXiv preprint arXiv:2406.12045 , year=

    tau -bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains , author=. arXiv preprint arXiv:2406.12045 , year=

    Pith/arXiv arXiv

  20. [20]

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

    Retrieval models aren’t tool-savvy: Benchmarking tool retrieval for large language models , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  21. [21]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    Planning, creation, usage: Benchmarking llms for comprehensive tool utilization in real-world complex scenarios , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    2024

  22. [22]

    International Conference on Learning Representations , volume=

    Agentbench: Evaluating llms as agents , author=. International Conference on Learning Representations , volume=

  23. [23]

    International Conference on Learning Representations , volume=

    Gaia: a benchmark for general ai assistants , author=. International Conference on Learning Representations , volume=

  24. [24]

    International Conference on Learning Representations , volume=

    Webarena: A realistic web environment for building autonomous agents , author=. International Conference on Learning Representations , volume=

  25. [25]

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

    Hammerbench: Fine-grained function-calling evaluation in real mobile assistant scenarios , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  26. [26]

    arXiv preprint arXiv:2305.03653 , year=

    Query expansion by prompting large language models , author=. arXiv preprint arXiv:2305.03653 , year=

    Pith/arXiv arXiv

  27. [27]

    Frontiers of Computer Science , volume=

    A survey of large language models , author=. Frontiers of Computer Science , volume=. 2026 , publisher=

    2026

  28. [28]

    ACM Transactions on Intelligent Systems and Technology , volume=

    A comprehensive overview of large language models , author=. ACM Transactions on Intelligent Systems and Technology , volume=. 2025 , publisher=

    2025

  29. [29]

    Foundations and Trends in Optimization , volume=

    Causal machine learning: A survey and open problems , author=. Foundations and Trends in Optimization , volume=. 2025 , publisher=

    2025

  30. [30]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    Re-invoke: Tool invocation rewriting for zero-shot tool retrieval , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    2024

  31. [31]

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

    Query2doc: Query expansion with large language models , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  32. [32]

    Findings of the Association for Computational Linguistics: NAACL 2024 , pages=

    Planning and editing what you retrieve for enhanced tool learning , author=. Findings of the Association for Computational Linguistics: NAACL 2024 , pages=

    2024

  33. [33]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  34. [34]

    arXiv preprint arXiv:2506.05176 , year=

    Qwen3 embedding: Advancing text embedding and reranking through foundation models , author=. arXiv preprint arXiv:2506.05176 , year=

    Pith/arXiv arXiv

  35. [35]

    arXiv preprint arXiv:2308.03281 , year=

    Towards general text embeddings with multi-stage contrastive learning , author=. arXiv preprint arXiv:2308.03281 , year=

    Pith/arXiv arXiv

  36. [36]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Improving text embeddings with large language models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  37. [37]

    ACM Computing Surveys , volume=

    A survey on large language models for mathematical reasoning , author=. ACM Computing Surveys , volume=. 2026 , publisher=

    2026

  38. [38]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    Enhancing tool retrieval with iterative feedback from large language models , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    2024

  39. [39]

    ACM Computing Surveys , volume=

    Counterfactual explanations and algorithmic recourses for machine learning: A review , author=. ACM Computing Surveys , volume=. 2024 , publisher=

    2024

  40. [40]

    Machine Intelligence Research , volume=

    Counterfactual learning on graphs: A survey , author=. Machine Intelligence Research , volume=. 2025 , publisher=

    2025

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.