Pith. sign in

REVIEW 4 major objections 5 minor 5 cited by

An 8B-parameter agent, trained on 460 tools, can select correct tools from a 1,346-tool pool that includes 886 tools never seen in training, outperforming larger reasoning models and fixed-tool systems across math, search, code, and multimo

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 · deepseek-v4-flash

2026-08-03 16:25 UTC pith:LNFII3KN

load-bearing objection Useful setup and a real dataset, but the headline claim about unseen-tool generalization is not actually tested by the experiments as designed. the 4 major comments →

arxiv 2512.13278 v2 pith:LNFII3KN submitted 2025-12-15 cs.CL cs.LG

AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning

Jiaru Zou , Ling Yang , Yunzhe Qi , Sirui Chen , Mengting Ai , Ke Shen , Jingrui He , Mengdi Wang This is my paper
classification cs.CL cs.LG
keywords dynamic tool selectionagentic reasoningPlackett–Luce rankingtool integrationreinforcement learningembedding-anchored selectionevolving toolsetsmultimodal reasoning
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 paper sets out to show that tool use in LLM agents does not have to be frozen at training time. It introduces AutoTool, a two-phase training recipe: first stabilize long reasoning trajectories with supervised fine-tuning and reinforcement learning, then refine tool choices with a Plackett–Luce ranking objective. The trained agent selects tools by comparing its predicted anchor embedding against embeddings of all candidate tools, so tools introduced after training can be chosen if their name-and-description embeddings align. On ten benchmarks, an 8B-parameter model trained this way outperforms larger reasoning models and fixed-tool systems, with average gains of 6.4% in math and science, 4.5% in search QA, 7.7% in code, and 6.9% in multimodal understanding. The paper's central generalization claim is that the same mechanism works when the inference toolset includes 886 tools the model never saw.

Core claim

AutoTool's central claim is that dynamic tool selection can be trained as an explicit, ranked decision inside the reasoning trajectory, and that the resulting policy generalizes to tools it has never seen. The model first learns stable reasoning and tool-integration trajectories through supervised fine-tuning and reinforcement learning, then optimizes only the tool-selection steps with a KL-regularized Plackett–Luce ranking objective. Selection is performed in embedding space: the model emits an anchor embedding, and the chosen tool is sampled from a softmax over distances to the embeddings of every candidate tool, which are computed from each tool's name and description. Trained on 460 tool

What carries the argument

The load-bearing mechanism is embedding-anchored tool selection: at each tool-selection step the model generates a rationale followed by an anchor token; the anchor's embedding is compared to embeddings of all candidate tools (computed by the model's own embedding layer from each tool's name and description), and the tool is sampled from a softmax over negative squared distances. The second piece is Phase II training, which casts the N sampled trajectory rollouts as a ranked list under the Plackett–Luce model, ties trajectory rewards to tool-selection quality, and trains the policy with a cross-entropy loss derived from the KL-regularized reinforcement-learning objective. The ranking objecti

Load-bearing premise

The whole generalization story rests on the assumption that embeddings of tools the model never saw in training—computed only from their name and description at inference—land close enough to the anchor-embedding distribution learned from 460 seen tools that softmax over distances picks the right one.

What would settle it

Take the 886 unseen tools, build a probe set where the correct tool for each query is known, and measure how often the embedding-distance selection rule picks it. If unseen-tool selection accuracy is at or near chance—or substantially below accuracy on the 460 seen tools—the central generalization claim fails. As a sharper test, add synthetic tools whose descriptions paraphrase seen tools with novel wording; if selection accuracy collapses despite semantic equivalence, the mechanism is keyed to surface vocabulary rather than function.

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

If this is right

  • An agent trained on a fixed 460-tool pool can be evaluated on a 1,346-tool pool with 886 unseen tools and still improve over fixed-tool baselines, implying the mechanism transfers rather than memorizing tool identifiers.
  • The Phase II Plackett–Luce ranking step adds consistent gains over trajectory stabilization alone, so explicit tool-selection optimization is complementary to reasoning stabilization.
  • A single 8B-parameter agent can be competitive with, or better than, much larger reasoning models and domain-specialized tool agents across math, search, code, and multimodal tasks, without task-specific retraining.
  • Dynamic tool selection nearly matches an oracle that is handed the correct tool in advance, suggesting the trained selection policy is close to the ceiling set by perfect selection.

Where Pith is reading between the lines

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

  • A natural extension is to use the same anchor-embedding selection for actions beyond external tools, e.g., choosing among internal computation strategies or memory retrievers, since the ranking objective is agnostic to what the candidates are.
  • The approach suggests tool-selection accuracy could be improved without additional trajectory data by rewriting tool descriptions, because embeddings are computed only from name and description; changing description wording should measurably shift selection performance.
  • The KL-regularized Plackett–Luce step could be applied to streaming tool registries: as new tools are added, only their embeddings need to be computed, and the policy can be re-run without gradient updates. Whether this works is a testable consequence, not something the paper measures.

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 AutoTool, a training framework for LLM agents that select tools dynamically from large, evolving toolsets. The method has two phases: (i) SFT plus RL for trajectory stabilization, and (ii) a KL-regularized Plackett-Luce ranking objective that is claimed to refine multi-step tool selection. The authors curate a 200k-instance dataset with explicit tool-selection rationales, spanning 1,346 tools of which 460 are seen during training and 886 are held out for inference. Experiments with Qwen3-8B and Qwen2.5-VL-7B report gains over SFT and GRPO across math, search, code, and multimodal benchmarks, and the abstract claims stronger generalization by leveraging unseen tools at inference.

Significance. If the central generalization claim were established, this would be a useful contribution: an 8B model selecting from a 1,346-tool pool, including tools never seen in training, would be practically valuable, and the large curated dataset with tool-selection rationales could be a resource for the community. The empirical results in Table 2 are consistent with the proposed pipeline being effective relative to SFT and GRPO. However, the headline claim about unseen tools is not actually measured, and the theoretical framing of PL ranking is overstated. These issues materially affect the paper's contribution as written.

major comments (4)
  1. [Appendix A.1/A.2, Table 2] The paper's central claim—that AutoTool 'exhibits stronger generalization by dynamically leveraging unseen tools'—is not supported by any isolated measurement. Inference uses a 1,346-tool candidate pool including 886 unseen tools (Appendix A.1), but all benchmark tasks are solvable with tools that are among the 460 seen tools (code interpreter, search API, OCR). Table 2 reports only end-to-end accuracy; there is no per-item tool-usage analysis, no seen/unseen split of tool-selection accuracy, and no task where an unseen tool is the unique correct choice. The 'Takeaway' box in §3.2 cites a word2vec-style analogy rather than a measurement. As designed, the evaluation cannot distinguish a model that ignores all 886 unseen tools and selects among the 460 seen tools from one that truly generalizes to unseen tools. This is a load-bearing gap for the abstract's main claim.
  2. [§3.3–3.4, Eqs. 5–8, Appendix B.1] The 'PL-ranking' objective is theoretically a restatement of KL-regularized RL / DPO. Eq. 8 directly optimizes πθ toward the closed-form optimal policy π*(τ) ∝ πold(τ) exp(Rtool(τ)/β). Proposition 3.1 states that πθ = π* iff the induced PL distributions coincide; the proof in Appendix B.1 relies on softmax shift-invariance (Lemma B.2) and normalization. This equivalence holds for any pair of policies and does not use any additional ranking information beyond the reward already encoded in Rtool. Thus the PL construction adds no constraint beyond the CE loss in Eq. 8. The authors should either demonstrate a concrete use of permutation-level PL information or reframe the contribution as DPO with a trajectory-level tool-selection reward.
  3. [Table 2] The claimed consistent gains over SFT and GRPO are reported without error bars, repeated seeds, or significance tests. Several differences are small relative to expected binomial variability on 500-instance sets (e.g., Qwen2.5-VL-7B V-Math 44.3 vs. 42.7; V-Code 52.5 vs. 49.6; Qwen3-8B Bamboogle 56.8 vs. 52.8). Without variance estimates or significance testing, the 'consistently outperforms' claim is not fully established. Please provide repeated runs or confidence intervals, at least for the main comparisons.
  4. [Table 3] The 'pre-given tool assignment' oracle baseline is underspecified. It is not stated whether the pre-given tools are drawn from the 460 seen tools or include unseen tools. If they are all seen tools, the comparison cannot speak to the unseen-tool generalization claim; if they include unseen tools, the construction and labeling procedure should be described. The near-tie between AutoTool and the oracle on some benchmarks is therefore difficult to interpret without this information.
minor comments (5)
  1. [Title/header] The title on page 1 lacks spacing: 'AUTOTOOL: DYNAMICTOOLSELECTION AND INTEGRATION FORAGENTICREASONING'.
  2. [§3.2, Eq. 4] Using the Frobenius norm for vector embeddings is unusual; the Euclidean norm is standard. Also clarify how the 'anchor token' embedding e'_i is obtained and how gradients flow from Eq. 8 through the selection mechanism in Eq. 4.
  3. [Table 2] Typos: 'GQPA' should be 'GPQA', and 'Bamboogl' should be 'Bamboogle'. Also the table header 'Math↑Search↑Multimodal↑' would benefit from spacing.
  4. [Appendix C] The 'Data Example' blocks for V-Chart, V-Math, and V-Code all show the same cryptarithmetic example prompt; this appears to be a copy-paste error and should be corrected to reflect each dataset.
  5. [Figure 1] The figure contains garbled text ('%#*&￥……') which should be cleaned up for a final version.

Circularity Check

1 steps flagged

Prop. 3.1's PL-ranking 'theoretical bridge' is definitional: the appendix defines the PL model via policy ratios, making the equivalence a tautology, while main-text Eq. 7 defines it via rewards alone.

specific steps
  1. self definitional [Section 3.4, Eq. 7–8 and Proposition 3.1; Appendix B.1, Eq. 9]
    "Pπθ (σ|T) = ∏_{j=1}^N exp(Rtool(τσ(j))) / Σ_{l=j}^N exp(Rtool(τσ(l))) ... πθ = π∗ ⇐⇒ Pπθ(σ|T) = Pπ∗(σ|T), ∀σ. ... [Appendix B.1] the Plackett-Luce (PL) ranking model induced by a policy π ... will be defined as: Pπ(σ|T,x) = ∏_{i=1}^{|T|} exp(β log π(τσ(i)|x)/πold(τσ(i)|x)) / ..."

    The appendix's PL distribution is defined with scores β log(π(τ|x)/πold(τ|x)). Under that definition, equality of Pπθ and Pπ* is equivalent, by softmax shift-invariance and normalization, to πθ=π*; no property of Rtool or of the optimal policy is used. The 'bridge' is therefore the definition, not a derived link. In the main text, Eq. 7 instead defines Pπθ directly from Rtool, which does not depend on πθ, so Proposition 3.1 as stated is a different (and false) claim. Eq. 8 is just the closed-form KL-regularized policy πold exp(Rtool/β)/Z written as a CE loss; the PL ranking machinery adds no independent constraint.

full rationale

The empirical comparisons in Tables 1–3 are against external benchmarks and baseline methods, so the performance results are not self-referential. The one genuinely circular element is the theoretical bridge: Proposition 3.1 is made true by the appendix's construction of the PL model from policy ratios, and the main-text reward-based PL distribution does not even depend on the policy. The abstract's claim of 'leveraging unseen tools' is also under-measured (Appendix A.2 reports only end-to-end accuracy with 886 unseen tools added; no per-item seen-vs-unseen selection accuracy is given, and all benchmark answers are plausibly solvable with the 460 seen tools). That is an evidential gap rather than a circular reduction, so it does not raise the score beyond the definitional PL step. Overall: partial circularity in the derivation chain, with independent empirical substance remaining.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 1 invented entities

The method rests on standard RLHF/DPO identities, an unspecified PRM, equal-weight reward composition, and — most importantly — the unvalidated assumption that embedding-space alignment transfers from seen to unseen tools. The free parameters are mostly unreported hyperparameters.

free parameters (5)
  • gamma (γ) = not reported
    Eq. 4: softmax temperature controlling selection skewness over embedding distances; hand set, no sensitivity analysis.
  • beta (β) = not reported
    Eq. 5: KL regularization intensity for the tool-selection objective; hand set, no sensitivity analysis.
  • reward combination weights = implicit 1:1
    Eq. 6: rtool = PRM + Acc, with no scaling or weights reported; equal weighting is a modeling choice.
  • process reward model (PRM) = unspecified
    Eq. 6 uses PRM for dense step-level supervision, but its source, architecture, and training data are not given; the target π* depends on it.
  • rollout size N and Phase II epochs = N=8, 3 epochs
    Implementation details chosen by hand; no ablation is reported.
axioms (6)
  • standard math Softmax shift-invariance (Lemma B.2)
    Used in Prop B.1 to infer an additive constant from equal softmax outputs; correct, but the further conclusion that policies are equal requires normalization of probability distributions.
  • domain assumption Embedding space is generalizable to unseen tools
    Takeaway box in §3.2: 'anchoring selection in the embedding space which is generalizable' — no evidence is provided that unseen-tool embeddings align with the trained anchor-embedding distribution.
  • domain assumption DeepSeek-R1 rationales and LLM-as-judge filtering are reliable
    §2 data pipeline generates rationales with DeepSeek-R1 and filters with LLM-as-judge; no human validation or inter-annotator agreement is reported.
  • domain assumption PRM provides valid dense supervision
    Eq. 6 assumes process reward model scores tool-selection steps correctly, but the PRM is unspecified and unvalidated.
  • domain assumption Per-step attribution of final-answer correctness to the selected tool
    Eq. 6 assigns Acc(x, tk) to the trajectory's tool-selection step; if multiple tools would yield the same correct answer, the reward attribution is confounded.
  • standard math Closed-form optimal policy for KL-regularized RL (Rafailov et al. 2023)
    Eq. 8 relies on the standard DPO/RLHF closed-form solution π*(τ) ∝ πold(τ) exp(Rtool/β), which holds under the stated KL regularization.
invented entities (1)
  • anchor token embedding e'_i no independent evidence
    purpose: Predicted embedding matched against candidate tool embeddings via softmax distance (Eq. 4) to select tools, including unseen ones.
    Introduced in §3.2 and not separately ablated or measured; no evidence that it generalizes to unseen tools beyond the overall benchmark numbers.

pith-pipeline@v1.3.0-alltime-deepseek · 20845 in / 13474 out tokens · 113098 ms · 2026-08-03T16:25:50.087553+00:00 · methodology

0 comments
read the original abstract

Agentic reinforcement learning has advanced large language models (LLMs) to reason through long chain-of-thought trajectories while interleaving external tool use. Existing approaches assume a fixed inventory of tools, which limits the adaptability of LLM agents to new or evolving toolsets. We present AutoTool, a training framework that equips LLM agents with dynamic tool-selection capabilities throughout their reasoning trajectories. AutoTool employs a dual-phase optimization pipeline: (i) SFT and RL-based trajectory stabilization for coherent reasoning, and (ii) KL-regularized Plackett-Luce Ranking to refine consistent multi-step tool selection. We further build a 200k dataset with explicit tool-selection rationales across 1,000+ tools and 100+ tasks spanning mathematics, science, code generation, and multimodal reasoning. Across ten diverse benchmarks, we train two base models, Qwen3-8B and Qwen2.5-VL-7B, with AutoTool. With fewer parameters, AutoTool consistently outperforms advanced LLM agents and tool-integration methods, yielding average gains of 6.4% in math & science reasoning, 4.5% in search-based QA, 7.7% in code generation, and 6.9% in multimodal understanding. In addition, AutoTool exhibits stronger generalization by dynamically leveraging unseen tools from evolving toolsets during inference.

Figures

Figures reproduced from arXiv: 2512.13278 by Jiaru Zou, Jingrui He, Ke Shen, Ling Yang, Mengdi Wang, Mengting Ai, Sirui Chen, Yunzhe Qi.

Figure 1
Figure 1. Figure 1: Illustration of fixed vs. (challenging) evolving tool environments. AutoTool enables LLM [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Data curation pipeline for AutoTool (Detailed in Section 2). The overall pipeline has three stages: (i) Toolset & Task Collection, assembling 1,000+ tools with metadata across 100+ tasks in math, science, code, and multimodal reasoning; (ii) Tool-Selection Rationale Generation, producing explicit justifications for tool choices; and (iii) Trajectory Augmentation, combining CoT reasoning, tool-selection, an… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration on AutoTool’s Tool-Selection Refinement Phase. We cast tool-selection as a [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Case Study of AutoTool on V-Math. 9 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Template to generate tool-selection rationales during AutoTool’s data curation process. [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. ToolAnchor: Anchoring Counterfactual Context to Boost Agentic Tool-use Capability

    cs.AI 2026-07 conditional novelty 6.0

    A three-stage teacher-hypothesize, student-verify, then-train pipeline lets a post-trained tool-using agent adopt new visual tools without retraining from scratch.

  2. Security Considerations for Multi-agent Systems

    cs.CR 2026-03 unverdicted novelty 6.0

    No existing AI security framework covers a majority of the 193 identified multi-agent system threats in any category, with OWASP Agentic Security Initiative achieving the highest overall coverage at 65.3%.

  3. Code as Agent Harness

    cs.CL 2026-05 accept novelty 5.0

    A survey that organizes existing work on LLM-based agents around code as the central harness, structured in three layers of interfaces, mechanisms, and multi-agent scaling, with applications across domains and listed ...

  4. Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering

    cs.SE 2026-04 accept novelty 5.0

    LLM agent progress depends on externalizing cognitive functions into memory, skills, protocols, and harness engineering that coordinates them reliably.

  5. Agentic Reasoning for Large Language Models

    cs.AI 2026-01 unverdicted novelty 4.0

    The survey structures agentic reasoning for LLMs into foundational, self-evolving, and collective multi-agent layers while distinguishing in-context orchestration from post-training optimization and reviewing applicat...

Reference graph

Works this paper leans on

79 extracted references · 47 linked inside Pith · cited by 5 Pith papers

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Vqa: Visual question answering

    Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision, pp.\ 2425--2433, 2015

    2015

  3. [3]

    Program synthesis with large language models

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, et al. Program synthesis with large language models. arXiv preprint arXiv:2108.07732, 2021

    Pith/arXiv arXiv 2021

  4. [4]

    Qwen2.5-vl technical report, 2025

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report, 2...

    Pith/arXiv arXiv 2025

  5. [5]

    Large language models as tool makers

    Tianle Cai, Xuezhi Wang, Tengyu Ma, Xinyun Chen, and Denny Zhou. Large language models as tool makers. ArXiv preprint, abs/2305.17126, 2023. URL https://arxiv.org/abs/2305.17126

    Pith/arXiv arXiv 2023

  6. [6]

    Evaluating large language models trained on code

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021

    Pith/arXiv arXiv 2021

  7. [7]

    Pan, Wen Zhang, Huajun Chen, Fan Yang, Zenan Zhou, and Weipeng Chen

    Mingyang Chen, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng Zhu, Haofen Wang, Jeff Z. Pan, Wen Zhang, Huajun Chen, Fan Yang, Zenan Zhou, and Weipeng Chen. Research: Learning to reason with search for llms via reinforcement learning, 2025. URL https://arxiv.org/abs/2503.19470

    Pith/arXiv arXiv 2025

  8. [8]

    Label ranking methods based on the plackett-luce model

    Weiwei Cheng, Eyke H \"u llermeier, and Krzysztof J Dembczynski. Label ranking methods based on the plackett-luce model. In Proceedings of the 27th International Conference on Machine Learning (ICML-10), pp.\ 215--222, 2010

    2010

  9. [9]

    Training verifiers to solve math word problems

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021

    Pith/arXiv arXiv 2021

  10. [10]

    Toolregistry: A protocol-agnostic tool management library for function-calling llms

    Peng Ding. Toolregistry: A protocol-agnostic tool management library for function-calling llms. arXiv preprint arXiv:2507.10593, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    Image super-resolution using deep convolutional networks

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38 0 (2): 0 295--307, 2015

    2015

  12. [12]

    Agentic reinforced policy optimization

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

    Pith/arXiv arXiv 2025

  13. [13]

    Anytool: Self-reflective, hierarchical agents for large-scale api calls

    Yu Du, Fangyun Wei, and Hongyang Zhang. Anytool: Self-reflective, hierarchical agents for large-scale api calls. In Proceedings of the 41st International Conference on Machine Learning (ICML 2024), volume 235 of Proceedings of Machine Learning Research, pp.\ 6757--6775. PMLR, 2024. URL https://arxiv.org/abs/2402.04253. ArXiv preprint arXiv:2402.04253

    Pith/arXiv arXiv 2024

  14. [14]

    A comprehensive survey of self-evolving ai agents: A new paradigm bridging foundation models and lifelong agentic systems

    Jinyuan Fang, Yanwen Peng, Xi Zhang, Yingxu Wang, Xinhao Yi, Guibin Zhang, Yi Xu, Bin Wu, Siwei Liu, Zihao Li, et al. A comprehensive survey of self-evolving ai agents: A new paradigm bridging foundation models and lifelong agentic systems. arXiv preprint arXiv:2508.07407, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    Retool: Reinforcement learning for strategic tool use in llms, 2025 a

    Jiazhan Feng, Shijue Huang, Xingwei Qu, Ge Zhang, Yujia Qin, Baoquan Zhong, Chengquan Jiang, Jinxin Chi, and Wanjun Zhong. Retool: Reinforcement learning for strategic tool use in llms, 2025 a . URL https://arxiv.org/abs/2504.11536

    Pith/arXiv arXiv 2025

  16. [16]

    Group-in-group policy optimization for llm agent training

    Lang Feng, Zhenghai Xue, Tingcong Liu, and Bo An. Group-in-group policy optimization for llm agent training. arXiv preprint arXiv:2505.10978, 2025 b

    Pith/arXiv arXiv 2025

  17. [17]

    Word embedding based generalized language model for information retrieval

    Debasis Ganguly, Dwaipayan Roy, Mandar Mitra, and Gareth JF Jones. Word embedding based generalized language model for information retrieval. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval, pp.\ 795--798, 2015

    2015

  18. [18]

    A survey of self-evolving agents: On path to artificial super intelligence

    Huan-ang Gao, Jiayi Geng, Wenyue Hua, Mengkang Hu, Xinzhe Juan, Hongzhang Liu, Shilong Liu, Jiahao Qiu, Xuan Qi, Yiran Wu, et al. A survey of self-evolving agents: On path to artificial super intelligence. arXiv preprint arXiv:2507.21046, 2025 a

    Pith/arXiv arXiv 2025

  19. [19]

    Multi-modal agent tuning: Building a vlm-driven agent for efficient tool usage

    Zhi Gao, Bofei Zhang, Pengxiang Li, Xiaojian Ma, Tao Yuan, Yue Fan, Yuwei Wu, Yunde Jia, Song-Chun Zhu, and Qing Li. Multi-modal agent tuning: Building a vlm-driven agent for efficient tool usage. In International Conference on Learning Representations (ICLR), 2025 b

    2025

  20. [20]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps

    Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. arXiv preprint arXiv:2011.01060, 2020

    Pith/arXiv arXiv 2011

  22. [22]

    Visual sketchpad: Sketching as a visual chain of thought for multimodal language models

    Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. Advances in Neural Information Processing Systems, 37: 0 139348--139379, 2024

    2024

  23. [23]

    Gpt-4o system card

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Livecodebench: Holistic and contamination free evaluation of large language models for code

    Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint arXiv:2403.07974, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    Mmsearch: Benchmarking the potential of large models as multi-modal search engines

    Dongzhi Jiang, Renrui Zhang, Ziyu Guo, Yanmin Wu, Jiayi Lei, Pengshuo Qiu, Pan Lu, Zehui Chen, Chaoyou Fu, Guanglu Song, et al. Mmsearch: Benchmarking the potential of large models as multi-modal search engines. arXiv preprint arXiv:2409.12959, 2024

    Pith/arXiv arXiv 2024

  26. [26]

    Search-r1: Training llms to reason and leverage search engines with reinforcement learning, 2025

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. Search-r1: Training llms to reason and leverage search engines with reinforcement learning, 2025. URL https://arxiv.org/abs/2503.09516

    Pith/arXiv arXiv 2025

  27. [27]

    Tptu-v2: Boosting task planning and tool usage of large language model-based agents in real-world systems

    Yilun Kong, Jingqing Ruan, Yihong Chen, Bin Zhang, Tianpeng Bao, Shiwei Shi, Guoqing Du, Xiaoru Hu, Hangyu Mao, Ziyue Li, et al. Tptu-v2: Boosting task planning and tool usage of large language model-based agents in real-world systems. arXiv preprint arXiv:2311.11315, 2023

    Pith/arXiv arXiv 2023

  28. [28]

    Repo: Replay-enhanced policy optimization

    Siheng Li, Zhanhui Zhou, Wai Lam, Chao Yang, and Chaochao Lu. Repo: Replay-enhanced policy optimization. arXiv preprint arXiv:2506.09340, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection

    Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European conference on computer vision, pp.\ 38--55. Springer, 2024

    2024

  30. [30]

    Understanding r1-zero-like training: A critical perspective

    Zichen Liu, Changyu Chen, Wenjun Li, Penghui Qi, Tianyu Pang, Chao Du, Wee Sun Lee, and Min Lin. Understanding r1-zero-like training: A critical perspective. arXiv preprint arXiv:2503.20783, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Arpo: End-to-end policy optimization for gui agents with experience replay

    Fanbin Lu, Zhisheng Zhong, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Arpo: End-to-end policy optimization for gui agents with experience replay. arXiv preprint arXiv:2505.16282, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Individual choice behavior, volume 4

    R Duncan Luce et al. Individual choice behavior, volume 4. Wiley New York, 1959

    1959

  33. [33]

    Agent rl scaling law: Agent rl with spontaneous code execution for mathematical problem solving

    Xinji Mai, Haotian Xu, Weinong Wang, Jian Hu, Yingying Zhang, Wenqiang Zhang, et al. Agent rl scaling law: Agent rl with spontaneous code execution for mathematical problem solving. arXiv preprint arXiv:2505.07773, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Chartgemma: Visual instruction-tuning for chart reasoning in the wild

    Ahmed Masry, Megh Thakkar, Aayush Bajaj, Aaryaman Kartha, Enamul Hoque, and Shafiq Joty. Chartgemma: Visual instruction-tuning for chart reasoning in the wild. arXiv preprint arXiv:2407.04172, 2024

    Pith/arXiv arXiv 2024

  35. [35]

    Making rest apis agent-ready: From openapi to mcp servers for tool-augmented llms

    Meriem Mastouri, Emna Ksontini, and Wael Kessentini. Making rest apis agent-ready: From openapi to mcp servers for tool-augmented llms. arXiv preprint arXiv:2507.16044, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    AIME 2025 dataset

    math ai. AIME 2025 dataset. https://huggingface.co/datasets/math-ai/aime25, 2025. Accessed: 2025-05-15

    2025

  37. [37]

    AIME 2024 dataset

    Maxwell-Jia. AIME 2024 dataset. https://huggingface.co/datasets/Maxwell-Jia/AIME_2024, 2024. Accessed: 2025-05-15

    2024

  38. [38]

    Distributed representations of words and phrases and their compositionality

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26, 2013

    2013

  39. [39]

    Webgpt: Browser-assisted question-answering with human feedback

    Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332, 2021

    Pith/arXiv arXiv 2021

  40. [40]

    Gorilla: Large language model connected with massive apis

    Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. Gorilla: Large language model connected with massive apis. Advances in Neural Information Processing Systems, 37: 0 126544--126565, 2024

    2024

  41. [41]

    Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl

    Yingzhe Peng, Gongrui Zhang, Miaosen Zhang, Zhiyuan You, Jie Liu, Qipeng Zhu, Kai Yang, Xingzhong Xu, Xin Geng, and Xu Yang. Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl. arXiv preprint arXiv:2503.07536, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    Spiqa: A dataset for multimodal question answering on scientific papers, 2025

    Shraman Pramanick, Rama Chellappa, and Subhashini Venugopalan. Spiqa: A dataset for multimodal question answering on scientific papers, 2025. URL https://arxiv.org/abs/2407.09413

    Pith/arXiv arXiv 2025

  43. [43]

    Measuring and narrowing the compositionality gap in language models

    Ofir Press, Muru Zhang, Sewon Min, Ludwig Schmidt, Noah A Smith, and Mike Lewis. Measuring and narrowing the compositionality gap in language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pp.\ 5687--5711, 2023

    2023

  44. [44]

    Fung, Yujia Qin, Zhiyuan Liu, and Heng Ji

    Cheng Qian, Chi Han, Yi R. Fung, Yujia Qin, Zhiyuan Liu, and Heng Ji. Creator: Tool creation for disentangling abstract and concrete reasoning of large language models. In Findings of the Association for Computational Linguistics: EMNLP 2023, pp.\ 462--477, 2023. URL https://aclanthology.org/2023.findings-emnlp.462/. EMNLP Findings 2023

    2023

  45. [45]

    Toolrl: Reward is all tool learning needs

    Cheng Qian, Emre Can Acikgoz, Qi He, Hongru Wang, Xiusi Chen, Dilek Hakkani-T \"u r, Gokhan Tur, and Heng Ji. Toolrl: Reward is all tool learning needs. arXiv preprint arXiv:2504.13958, 2025

    Pith/arXiv arXiv 2025

  46. [46]

    Tool LLM : Facilitating large language models to master 16000+ real-world API s

    Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Lauren Hong, Runchu Tian, Ruobing Xie, Jie Zhou, Mark Gerstein, dahai li, Zhiyuan Liu, and Maosong Sun. Tool LLM : Facilitating large language models to master 16000+ real-world API s. In The Twelfth International Conference on Lear...

    2024

  47. [47]

    Direct preference optimization: Your language model is secretly a reward model

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in neural information processing systems, 36: 0 53728--53741, 2023

    2023

  48. [48]

    Gpqa: A graduate-level google-proof q&a benchmark

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. Gpqa: A graduate-level google-proof q&a benchmark. In First Conference on Language Modeling, 2024

    2024

  49. [49]

    Toolformer: Language models can teach themselves to use tools

    Timo Schick, Jane Dwivedi-Yu, Roberto Dess \` , Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36: 0 68539--68551, 2023

    2023

  50. [50]

    Proximal policy optimization algorithms, 2017

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms, 2017. URL https://arxiv.org/abs/1707.06347

    Pith/arXiv arXiv 2017

  51. [51]

    rstar2-agent: Agentic reasoning technical report

    Ning Shang, Yifei Liu, Yi Zhu, Li Lyna Zhang, Weijiang Xu, Xinyu Guan, Buze Zhang, Bingcheng Dong, Xudong Zhou, Bowen Zhang, et al. rstar2-agent: Agentic reasoning technical report. arXiv preprint arXiv:2508.20722, 2025

    Pith/arXiv arXiv 2025

  52. [52]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    Pith/arXiv arXiv 2024

  53. [53]

    Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face

    Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. Advances in Neural Information Processing Systems, 36: 0 38154--38180, 2023

    2023

  54. [54]

    Hybridflow: A flexible and efficient rlhf framework

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. Hybridflow: A flexible and efficient rlhf framework. arXiv preprint arXiv: 2409.19256, 2024

    Pith/arXiv arXiv 2024

  55. [55]

    Agentic reasoning and tool integration for llms via reinforcement learning

    Joykirat Singh, Raghav Magazine, Yash Pandya, and Akshay Nambi. Agentic reasoning and tool integration for llms via reinforcement learning. arXiv preprint arXiv:2505.01441, 2025

    Pith/arXiv arXiv 2025

  56. [56]

    Restgpt: Connecting large language models with real-world restful apis

    Yifan Song, Weimin Xiong, Dawei Zhu, Wenhao Wu, Han Qian, Mingbo Song, Hailiang Huang, Cheng Li, Ke Wang, Rong Yao, et al. Restgpt: Connecting large language models with real-world restful apis. arXiv preprint arXiv:2306.06624, 2023

    Pith/arXiv arXiv 2023

  57. [57]

    Openthinkimg: Learning to think with images via visual tool reinforcement learning

    Zhaochen Su, Linjie Li, Mingyang Song, Yunzhuo Hao, Zhengyuan Yang, Jun Zhang, Guanjie Chen, Jiawei Gu, Juntao Li, Xiaoye Qu, et al. Openthinkimg: Learning to think with images via visual tool reinforcement learning. arXiv preprint arXiv:2505.08617, 2025

    Pith/arXiv arXiv 2025

  58. [58]

    Kimi k2: Open agentic intelligence

    Kimi Team, Yifan Bai, Yiping Bao, Guanduo Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, et al. Kimi k2: Open agentic intelligence. arXiv preprint arXiv:2507.20534, 2025

    Pith/arXiv arXiv 2025

  59. [59]

    QwQ-32B: Embracing the power of reinforcement learning

    Qwen Team. QwQ-32B: Embracing the power of reinforcement learning . https://qwenlm.github.io/blog/qwq-32b/, March 2025

    2025

  60. [60]

    Mllm-tool: A multimodal large language model for tool agent learning

    Chenyu Wang, Weixin Luo, Sixun Dong, Xiaohua Xuan, Zhengxin Li, Lin Ma, and Shenghua Gao. Mllm-tool: A multimodal large language model for tool agent learning. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp.\ 6678--6687. IEEE, 2025 a

    2025

  61. [61]

    Ragen: Understanding self-evolution in llm agents via multi-turn reinforcement learning

    Zihan Wang, Kangrui Wang, Qineng Wang, Pingyue Zhang, Linjie Li, Zhengyuan Yang, Xing Jin, Kefan Yu, Minh Nhat Nguyen, Licheng Liu, et al. Ragen: Understanding self-evolution in llm agents via multi-turn reinforcement learning. arXiv preprint arXiv:2504.20073, 2025 b

    Pith/arXiv arXiv 2025

  62. [62]

    V?: Guided visual search as a core mechanism in multimodal llms

    Penghao Wu and Saining Xie. V?: Guided visual search as a core mechanism in multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 13084--13094, 2024

    2024

  63. [63]

    Qwen3 technical report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025

    Pith/arXiv arXiv 2025

  64. [64]

    Hotpotqa: A dataset for diverse, explainable multi-hop question answering

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. arXiv preprint arXiv:1809.09600, 2018

    Pith/arXiv arXiv 2018

  65. [65]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR), 2023

    2023

  66. [66]

    Dapo: An open-source llm reinforcement learning system at scale

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, et al. Dapo: An open-source llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476, 2025 a

    Pith/arXiv arXiv 2025

  67. [67]

    Demystifying reinforcement learning in agentic reasoning

    Zhaochen Yu, Ling Yang, Jiaru Zou, Shuicheng Yan, and Mengdi Wang. Demystifying reinforcement learning in agentic reasoning. arXiv preprint arXiv:2510.11701, 2025 b

    arXiv 2025

  68. [68]

    Fung, Hao Peng, and Heng Ji

    Lifan Yuan, Yangyi Chen, Xingyao Wang, Yi R. Fung, Hao Peng, and Heng Ji. Craft: Customizing llms by creating and retrieving from specialized toolsets. In 12th International Conference on Learning Representations (ICLR 2024), 2024. URL https://arxiv.org/abs/2309.17428. ArXiv preprint arXiv:2309.17428

    Pith/arXiv arXiv 2024

  69. [69]

    Mammoth: Building math generalist models through hybrid instruction tuning

    Xiang Yue, Xingwei Qu, Ge Zhang, Yao Fu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mammoth: Building math generalist models through hybrid instruction tuning. arXiv preprint arXiv:2309.05653, 2023

    Pith/arXiv arXiv 2023

  70. [70]

    Gvpo: Group variance policy optimization for large language model post-training

    Kaichen Zhang, Yuzhong Hong, Junwei Bao, Hongfei Jiang, Yang Song, Dingqian Hong, and Hui Xiong. Gvpo: Group variance policy optimization for large language model post-training. arXiv preprint arXiv:2504.19599, 2025

    arXiv 2025

  71. [71]

    Cumulative reasoning with large language models

    Yifan Zhang, Jingqin Yang, Yang Yuan, and Andrew Chi-Chih Yao. Cumulative reasoning with large language models. arXiv preprint arXiv:2308.04371, 2023

    Pith/arXiv arXiv 2023

  72. [72]

    Geometric-mean policy optimization

    Yuzhong Zhao, Yue Liu, Junpeng Liu, Jingye Chen, Xun Wu, Yaru Hao, Tengchao Lv, Shaohan Huang, Lei Cui, Qixiang Ye, et al. Geometric-mean policy optimization. arXiv preprint arXiv:2507.20673, 2025

    arXiv 2025

  73. [73]

    Group sequence policy optimization

    Chujie Zheng, Shixuan Liu, Mingze Li, Xiong-Hui Chen, Bowen Yu, Chang Gao, Kai Dang, Yuqiong Liu, Rui Men, An Yang, et al. Group sequence policy optimization. arXiv preprint arXiv:2507.18071, 2025

    Pith/arXiv arXiv 2025

  74. [74]

    Judging llm-as-a-judge with mt-bench and chatbot arena

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in neural information processing systems, 36: 0 46595--46623, 2023

    2023

  75. [75]

    Llamafactory: Unified efficient fine-tuning of 100+ language models

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of 100+ language models. arXiv preprint arXiv:2403.13372, 2024

    Pith/arXiv arXiv 2024

  76. [76]

    Tattoo: Tool-grounded thinking prm for test-time scaling in tabular reasoning, 2025

    Jiaru Zou, Soumya Roy, Vinay Kumar Verma, Ziyi Wang, David Wipf, Pan Lu, Sumit Negi, James Zou, and Jingrui He. Tattoo: Tool-grounded thinking prm for test-time scaling in tabular reasoning, 2025. URL https://arxiv.org/abs/2510.06217

    arXiv 2025

  77. [77]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

  78. [78]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

  79. [79]

    BI PdOmĠddL uoZuZ Z:vr.vi d 0n o^[ ?J

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...

    arXiv 1966