Pith. sign in

REVIEW 3 major objections 4 minor 18 cited by

Chain-of-Agents: End-to-End Agent Foundation Models via Multi-Agent Distillation and Agentic RL

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read This paper shows that multi-agent collaboration can be distilled end-to-end into a single LLM's weights, producing a model that outperforms prompt-engineered multi-agent systems and tool-integrated-reasoning baselines on web and code benchm

desk verdict The code/math half of this paper is a genuinely useful training recipe, but the web-agent SOTA headline leans on an uncalibrated LLM judge used in both RL and evaluation, so the central claim is only half-supported. read the letter →

arxiv 2508.13167 v1 pith:VKQR6QF5 submitted 2025-08-06 cs.AI cs.CL

classification cs.AIcs.CL
keywords chain-of-agentsmulti-agentdistillationagenticreinforcementlearningagentfoundationmodelstool-integratedreasoningwebagentscodeLLM-as-judge
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces Chain-of-Agents, a way to make one LLM behave like an entire multi-agent system during a single decoding pass: the model dynamically activates role-playing and tool agents, and its output trajectory is trained end-to-end. To create training data, the authors run a state-of-the-art multi-agent system on thousands of tasks and transcribe its successful executions into chain-of-agents trajectories, then fine-tune on those and apply reinforcement learning with binary rewards (an LLM judge for open-ended web tasks, test-case pass/fail for code and math). The resulting Agent Foundation Models set new state-of-the-art numbers at the 32B scale on GAIA (55.3%), BrowseComp (11.1%), HLE (18.0%), AIME2025 (59.8%), LiveCodeBench v5 (47.9%), and CodeContests (32.7%), while cutting inference token cost by roughly 85% compared with a traditional multi-agent framework. A sympathetic reader would take away that the coordination policy of a multi-agent system can live in weights rather than in prompts, making agentic ability trainable and scalable.

What carries the argument

The chain-of-agents trajectory is the load-bearing object: a single model emits interleaved <think>, <plan>, <tool>, <observation>, <reflection>, and <answer> tokens, with role and tool selection expressed as token choices. Multi-agent distillation supplies the training signal by transcribing a working multi-agent system's execution into that format; progressive filtering keeps only long (5-20 hop), reflective, correct trajectories; and agentic RL sharpens the policy with sparse binary rewards. Observation masking during SFT keeps the model from memorizing noisy tool outputs.

What would settle it

Take a random sample of AFM's predictions on GAIA and HLE (say 100 each), grade them by human annotators, and compare pass/fail agreement against the 72B judge model's scores; if agreement is near chance or biased toward the judge's own output style, the headline web-agent numbers would not survive independent verification. For the code/math side, the falsifier is a counter-example question where the model's trajectory plans and reflects correctly but its final code fails all hidden test cases, checking whether the RL reward actually tracks execution success rather than format.

Watch

Extended reading notes

Core claim

The central claim is that the reasoning pattern of a multi-agent system—planning, searching, crawling, reflecting, verifying, then answering—can be re-expressed as a single autoregressive trajectory, and that this trajectory can be taught to a base LLM in two stages. First, multi-agent distillation turns successful executions of an open-source multi-agent system into sequences of the form <think><plan><tools><observation><reflection>...<answer>, filtered for complexity, quality, reflection presence, and error-correction value, and trained with observation masking. Second, agentic RL on verifiable tasks uses outcome rewards—an LLM judge's binary verdict for web agents, sandbox-executed test c

Load-bearing premise

The web-agent results rest on the assumption that a 72B language-model judge's binary 'correct/incorrect' verdict is a valid oracle for open-ended task success; no human-graded calibration is reported, so part of the GAIA/HLE/BrowseComp gains could be alignment with the judge rather than with ground truth.

Editorial extensions

If this is right

  • Multi-agent frameworks can be replaced by a single trained model, eliminating per-task prompt/workflow engineering and inter-agent token overhead.
  • Agentic capability becomes data-centric: collecting more and better multi-agent trajectories can improve the model without redesigning the framework.
  • RL on verifiable agentic tasks (test cases, exact answers) transfers to open-ended web tasks when paired with an LLM-judge reward, extending TIR-style training to multi-tool coordination.
  • Test-time scaling composes with the paradigm: Pass@3 with best-of-N selection raises GAIA from 55.3 to 69.9, implying the model's trajectory distribution is useful as a proposal sampler.
  • The paper's generalization experiment indicates that a code-trained CoA model can operate unseen tools from prompt schemas, suggesting tool-use skill transfers across agent domains.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the web-agent reward and evaluation both come from the same judge family, a separate human-graded pass on GAIA/HLE outputs would be the cleanest test of whether the reported gains are genuine task competence or alignment with the judge's preferences; the code/math numbers are immune to this concern since they are test-case based.
  • The recipe implies a rising-ceiling property: whenever a better multi-agent teacher appears, its successful trajectories can be distilled into the same backbone, so AFM performance can track the state of the art in multi-agent design without a new architecture.
  • The reflection-enrichment and error-correction upsampling steps suggest that what matters most in the curriculum is exposure to self-corrected failures; ablating these filters would isolate how much of the SFT gain comes from reflection tokens versus trajectory length.
  • One can test whether CoA is a general inductive bias rather than a distillation artifact by training from scratch on synthesized trajectories without a teacher system; if performance holds, the paradigm itself, not the teacher, is doing the work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes Chain-of-Agents (CoA), a single-model paradigm that simulates multi-agent collaboration through dynamic role and tool agents, and trains 'Agent Foundation Models' (AFMs) by (i) distilling OAgents trajectories into CoA-format supervised fine-tuning data and (ii) applying DAPO-style reinforcement learning with verifiable rewards for code/math and an LLM-as-judge reward for web tasks. On MHQA, GAIA, BrowseComp, HLE, LiveCodeBench, CodeContests, and math benchmarks, the authors report state-of-the-art results at the 7B/32B scale, together with an efficiency analysis and an open-source release of models, code, and data.

Significance. The training recipe is a plausible and timely way to turn multi-agent collaboration into an end-to-end trainable policy, and the code/math results are supported by test-case and exact-answer verification. The temporal split on LiveCodeBench v4-v5, the avg@16 protocol for small math sets, and the explicit data-filtering stages are good methodological choices. The claimed web-agent superiority, however, rests on an LLM-as-judge signal that is used both as the RL reward and as the evaluation metric, with no human calibration; this is the main barrier to accepting the web-side headline claims. If the judge concern is resolved with objective scoring or calibration, the contribution is substantial and the open-sourced assets would be valuable to the community.

major comments (3)
  1. [§3.3.2 Eq. (10), §4.1.1 Metrics, §D.4] The web-agent reward and the final evaluation both use Qwen-2.5-72B-Instruct as an LLM-as-judge, but no human agreement, no cross-judge consistency, and no comparison to the official GAIA/BrowseComp/HLE answer keys is reported. GAIA has known answers, BrowseComp uses verifiable short answers, and HLE has deterministic grading; these could be scored objectively. As written, the GAIA 55.3, BrowseComp 11.1, and HLE 18.0 results may partly reflect alignment with the judge's biases (including family bias, since the trained model and judge belong to the same Qwen family) rather than genuine task success. The code/math results are unaffected, but this issue is load-bearing for the abstract's 'new state-of-the-art' claim in web-agent settings.
  2. [§4.1.2, Table 7] GAIA is evaluated on only 103 text-only examples. The claimed SOTA margins are small: AFM-RL at 55.3 vs. WebSailor at 53.2 and WebShaper at 53.3 is a difference of about two examples, and even the 3.8-point margin vs. WebDancer is about four examples. No confidence intervals, bootstraps, or repeated-run variance are reported. The 'new state-of-the-art' phrasing on GAIA is therefore fragile and should be qualified with uncertainty estimates or additional evaluation trials.
  3. [§4.1.1 Metrics, Table 6] The MHQA results in Table 6 are also evaluated with the Qwen-2.5-72B judge. It is not clear which baseline numbers were re-scored with this judge and which were taken from original papers using possibly different metrics. This can make the comparisons inconsistent, especially against methods whose published scores use exact match or other graders. Please state explicitly which baselines were reproduced, and report exact-match accuracy alongside judge-based accuracy for at least the main MHQA tables.
minor comments (4)
  1. [Table 12] The AFM-SFT and AFM-RL rows appear twice without 7B/32B labels, so the reader cannot tell which row corresponds to which scale. Please add explicit model-size columns or row labels.
  2. [§5.2] The 'Generalization on Unseen Agents' claim is supported only by a single case study. This is interesting as an illustration, but it should be framed as anecdotal or supplemented with quantitative success rates on unseen-tool invocations.
  3. [§5.1] The efficiency comparison uses only 10 GAIA instances, and the metric 'prompt engineering cost' is not precisely defined. Please state the measurement procedure and, ideally, report a confidence interval or a larger evaluation set.
  4. [Various] Typos and unclear phrasing: 'progressive progressive filtering', 'sophiscated', 'contentional TIR', and the Figure 1 caption says 'Chain-of-Action' instead of 'Chain-of-Agents'. Also, Table 6's column headings are inconsistent with the dataset descriptions in §4.1.1 (e.g., TQ appears in the text but not in the table, while TriviaQA is placed under 'Single-Hop QA' in the table but described as multi-hop elsewhere).

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: benchmark results are empirically grounded; the shared LLM-judge for web reward and evaluation is a validity risk, not a derivation loop.

full rationale

The paper is an empirical systems paper; its central claim is a benchmark comparison, not a derived quantity. The code/math results are grounded in independent, rule-based verification: Rcode uses unit tests and Math-Verify (Sec. 3.3.2, Sec. 4.2.1), and LiveCodeBench/CodeContests/AIME results are scored by test-case execution or exact-match answer verification. These results do not reduce to the training reward by construction. The web-agent results do use an LLM-as-judge both in the RL reward (Eq. 10: score_answer is 1 if judge model Mj deems the prediction correct) and in the reported metric (Sec. 4.1.1: Qwen-2.5-72B provides binary correctness assessments). This is a legitimate measurement-validity concern: the model is optimized to satisfy an LLM judge's verdicts and is then measured by the same kind of verdict, so part of the GAIA/BrowseComp/HLE numbers could reflect judge alignment rather than task competence. However, this is not a circular derivation in the strict sense: the benchmark questions are held out from RL training, the judge is not a fitted parameter, and the paper does not define the target quantity as the training reward. The teacher OAgents [82] and the TaskCraft data-generation pipeline [49] are self-citations with substantial author overlap, but OAgents is independently published, open-source, and directly compared in Table 7, so the self-citation is not load-bearing in a way that forces the outcome. The abstract's unqualified 'new state-of-the-art' wording overstates Table 7 (OAgents/DeepResearch score higher on GAIA), but that is a correctness/overclaim issue, not circularity. Overall: no significant circularity; score 2 reflects the minor, non-load-bearing self-citations in the training-data pipeline.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The central claim depends on hand-set data filters, an uncalibrated LLM-judge reward, the OAgents teacher (which shares authors with this paper), and the lossy flattening of multi-agent state into one trajectory. These are choices made in the paper, not externally forced constraints.

free parameters (4)
  • RL query pass-rate threshold r_q = 0.3
    Queries with an 8-to-32 sample pass rate above 0.3 are excluded from RL training (Eq. 8-9); this hand-set threshold determines which tasks the RL stage sees and shapes the claimed web benchmark gains.
  • Complexity filter minimum agent-tool interactions = 5
    Trajectories with fewer than 5 agent-tool interactions are discarded in Section 3.2.1; this threshold defines the difficulty curriculum of the SFT corpus.
  • Code RL easy-query filter = solved in all 8 trials
    Queries a fine-tuned 7B model solves in all 8 samples are discarded (Section 3.3.1), biasing RL toward harder problems and potentially inflating pass@1 gains on contest benchmarks.
  • Reflection enrichment and error-correction upsampling weights = unspecified
    Downsampling of trajectories without reflection and upsampling of double-check recovery trajectories (Section 3.2.1) are data-balance decisions that affect SFT quality, but their exact rates are not reported.
assumptions (5)
  • domain assumption Flattened CoA trajectories preserve the essential coordination of the multi-agent teacher
    Section 3.2.1 converts OAgents' separate agent states and inter-agent messages (Eq. 2) into a single <think>/<tools>/<observation>/<reflection> sequence (Eq. 5); the student can only learn what this lossy projection retains.
  • ad hoc to paper LLM-as-judge (Qwen-2.5-72B) provides valid binary correctness signals for open-ended web tasks
    The web-agent RL reward is exactly the judge's score (Eq. 10) and MHQA evaluation uses the same judge type; no agreement rate, calibration, or human correlation is reported (Sections 3.3.2, 4.1.1).
  • domain assumption Teacher OAgents trajectories are high-quality SOTA demonstrations
    The whole SFT corpus is distilled from OAgents [82], a system built by an overlapping author group; the paper assumes its outputs are reliable enough to teach the student.
  • domain assumption Reported baseline numbers are directly comparable without identical evaluation harnesses
    Tables 6 and 7 mix published numbers with a few gray-highlighted reproduced runs; differences in judge prompts, tool backends, or GAIA subsets can shift score comparisons.
  • ad hoc to paper Observation masking is a correct inductive bias
    Loss over external tool outputs is masked (Eq. 7) claiming prevention of 'environmental noise propagation'; the paper provides no ablation showing this helps.
invented entities (2)
  • Chain-of-Agents role markers (Thinking/Plan/Reflection/Verification and tool agents)
    purpose: Simulate multi-agent collaboration inside one model by tagging each trajectory step as a different agent role
    These are new conceptual and prompt-level constructs; their benefit is evidenced only by the paper's own benchmark comparisons, with no external falsifiable handle.
  • Agent Foundation Models (AFMs)
    purpose: Name for the trained CoA models, positioning them as a new foundation class
    No new architectural component is introduced; it is a naming and positioning assertion, not an independently evidenced entity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Chain-of-Agents: End-to-End Agent Foundation Models via Multi-Agent Distillation and Agentic RL." pith.science (2026). https://pith.science/paper/VKQR6QF5

@misc{pith2026250813167,
  author       = {Pith},
  title        = {Pith review of: Chain-of-Agents: End-to-End Agent Foundation Models via Multi-Agent Distillation and Agentic RL},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VKQR6QF5}},
  note         = {Machine review of arXiv:2508.13167}
}
read the original abstract

Recent advances in large language models (LLMs) and multi-agent systems have demonstrated remarkable capabilities in complex problem-solving tasks such as deep research, vibe coding, and mathematical reasoning. However, most existing multi-agent systems are built upon manual prompt/workflow engineering with sophisticated agent frameworks, making them computationally inefficient, less capable, and can not benefit from data-centric learning. In this work, we introduce Chain-of-Agents (CoA), a novel paradigm of LLM reasoning that enables native end-to-end complex problem-solving in the same way as a multi-agent system (i.e., multi-turn problem solving with multiple tools and multiple agents) within one model. In chain-of-agents problem-solving, the model dynamically activates different tool agents and role-playing agents to simulate multi-agent collaboration in an end-to-end fashion. To elicit end-to-end chain-of-agents problem-solving abilities in LLMs, we introduce a multi-agent distillation framework to distill state-of-the-art multi-agent systems into chain-of-agents trajectories for agentic supervised fine-tuning. We then use agentic reinforcement learning on verifiable agentic tasks to further improve the models' capabilities on chain-of-agents problem solving. We call the resulting models Agent Foundation Models (AFMs). Our empirical studies demonstrate that AFM establishes new state-of-the-art performance across diverse benchmarks in both web agent and code agent settings. We make the entire research, including the model weights, code for training and evaluation, and the training data, fully open-sourced, which offers a solid starting point for future research on agent models and agentic RL.

Discussion (0). Sign in to comment.

Forward citations

Cited by 18 Pith papers

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

  1. SCOUT: Active Information Foraging for Long-Text Understanding with Decoupled Epistemic States

    cs.CL 2026-05 unverdicted novelty 7.0 of 10

    SCOUT achieves state-of-the-art long-text understanding with up to 8x lower token use by actively foraging for sparse query-relevant information and updating a compact provenance-grounded epistemic state.

  2. MAD-OPD: Breaking the Ceiling in On-Policy Distillation via Multi-Agent Debate

    cs.CL 2026-05 unverdicted novelty 7.0 of 10

    MAD-OPD recasts on-policy distillation teachers as a debating collective to supply better supervision, lifting agentic and code performance over single-teacher OPD across multiple model sizes.

  3. Student-Centered Distillation Narrows the Agentic Gap Between Small and Large LLMs

    cs.CL 2025-09 conditional novelty 7.0 of 10

    A 7B LLM agent trained with student-led distillation and one-step teacher corrections nearly matches a 72B teacher on reasoning and tool-use benchmarks.

  4. Efficient Agentic Reasoning Through Self-Regulated Simulative Planning

    cs.AI 2026-05 unverdicted novelty 6.0 of 10

    SR²AM achieves competitive Pass@1 accuracy on diverse tasks with 25.8-95.3% fewer reasoning tokens than much larger models by using self-regulated simulative planning trained via supervised learning and RL.

  5. ICRL: Learning to Internalize Self-Critique with Reinforcement Learning

    cs.AI 2026-05 unverdicted novelty 6.0 of 10

    ICRL uses joint RL training of solver and critic with distribution-calibration re-weighting and role-wise advantage estimation to internalize critique into unassisted LLM performance, yielding 6.4-point gains on agent...

  6. SOD: Step-wise On-policy Distillation for Small Language Model Agents

    cs.CL 2026-05 unverdicted novelty 6.0 of 10

    SOD reweights on-policy distillation strength step-by-step using divergence to stabilize tool use in small language model agents, yielding up to 20.86% gains and 26.13% on AIME 2025 for a 0.6B model.

  7. Agent-World: Scaling Real-World Environment Synthesis for Evolving General Agent Intelligence

    cs.AI 2026-04 unverdicted novelty 6.0 of 10

    Agent-World autonomously synthesizes verifiable real-world tasks and uses continuous self-evolution to train 8B and 14B agents that outperform proprietary models on 23 benchmarks.

  8. Paper Circle: An Open-source Multi-agent Research Discovery and Analysis Framework

    cs.CL 2026-04 unverdicted novelty 6.0 of 10

    Paper Circle is an open-source multi-agent system that retrieves papers via offline and online sources, applies multi-criteria scoring and diversity ranking, and converts papers into typed knowledge graphs for structu...

  9. MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling

    cs.CL 2025-11 unverdicted novelty 6.0 of 10

    MiroThinker shows that scaling agent-environment interactions via reinforcement learning lets a 72B open-source model reach up to 81.9% on GAIA and approach commercial performance on research benchmarks.

  10. SFR-DeepResearch: Towards Effective Reinforcement Learning for Autonomously Reasoning Single Agents

    cs.AI 2025-09 conditional novelty 6.0 of 10

    A 20B autonomously reasoning deep-research agent trained with synthetic-data RL reaches 28.7% on Humanity's Last Exam, exceeding several larger and proprietary baselines.

  11. Agentic Environment Engineering for Large Language Models: A Survey of Environment Modeling, Synthesis, Evaluation, and Application

    cs.CL 2026-06 unverdicted novelty 5.0 of 10

    This survey categorizes agentic environments for LLMs by eight attributes and domains, introduces symbolic and neural synthesis paradigms with evaluation, and outlines four agent evolution pathways plus three environm...

  12. SOD: Step-wise On-policy Distillation for Small Language Model Agents

    cs.CL 2026-05 conditional novelty 5.0 of 10

    A step-wise reweighting of on-policy distillation, based on per-step student-teacher divergence, improves tool-integrated reasoning in 0.6B and 1.7B language models.

  13. LiteResearcher: A Scalable Agentic RL Training Framework for Deep Research Agent

    cs.AI 2026-04 unverdicted novelty 5.0 of 10

    LiteResearcher uses a lite virtual world to make agentic RL training scalable and stable, enabling a 4B model to achieve 71.3% on GAIA and 78.0% on Xbench, outperforming larger open-source and commercial systems.

  14. LiteResearcher: A Scalable Agentic RL Training Framework for Deep Research Agent

    cs.AI 2026-04 conditional novelty 5.0 of 10

    Injecting 1% targeted synthetic data into GPT-2's pre-training substantially improves performance on 8 of 9 failing BLiMP grammatical paradigms, indicating data scarcity causes formal linguistic failures.

  15. Claw-R1: A Step-Level Data Middleware System for Agentic Reinforcement Learning

    cs.LG 2026-06 unverdicted novelty 4.0 of 10

    Claw-R1 provides a Gateway Server and Data Pool to manage step-level agent interaction traces as structured data assets for agentic RL training.

  16. Scaling Mobile Agent Systems: From Capability Density to Collective Intelligence

    cs.DC 2026-04 unverdicted novelty 3.0 of 10

    A vision paper outlining a two-pronged research agenda for scaling mobile agents from isolated devices to distributed intelligent systems.

  17. A Survey of Reinforcement Learning for Large Reasoning Models

    cs.CL 2025-09 accept novelty 3.0 of 10

    A survey compiling RL methods, challenges, data resources, and applications for enhancing reasoning in large language models and large reasoning models since DeepSeek-R1.

  18. Mathematical methods of reinforcement learning

    math.OC 2026-07 accept

    A survey unifying the operator-theoretic, probabilistic, and optimization-based mathematical structures underlying modern reinforcement learning algorithms.

Reference graph

Works this paper leans on

210 extracted references · 22 canonical work pages · cited by 16 Pith papers

  1. [1]

    Towards effective code-integrated reasoning

    Fei Bai, Yingqian Min, Beichen Zhang, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, Zheng Liu, Zhongyuan Wang, and Ji-Rong Wen. Towards effective code-integrated reasoning. arXiv preprint arXiv:2505.24480, 2025

  2. [2]

    Multi-agent reinforcement learning: A review of challenges and applications

    Lorenzo Canese, Gian Carlo Cardarilli, Luca Di Nunzio, Rocco Fazzolari, Daniele Giardino, Marco Re, and Sergio Spanò. Multi-agent reinforcement learning: A review of challenges and applications. Applied Sciences, 11(11):4948, 2021

  3. [3]

    Learning to reason with search for llms via reinforcement learning

    Mingyang Chen, Tianpeng Li, Haoze Sun, Yijie Zhou, Chenzheng Zhu, Haofen Wang, Jeff Z Pan, Wen Zhang, Huajun Chen, Fan Yang, et al. Learning to reason with search for llms via reinforcement learning. arXiv preprint arXiv:2503.19470, 2025

  4. [4]

    Process reinforcement through implicit rewards

    Ganqu Cui, Lifan Yuan, Zefan Wang, Hanbin Wang, Wendi Li, Bingxiang He, Yuchen Fan, Tianyu Yu, Qixin Xu, Weize Chen, et al. Process reinforcement through implicit rewards. arXiv preprint arXiv:2502.01456, 2025

  5. [5]

    Tool-star: Empowering llm-brained multi-tool reasoner via reinforcement learning.arXiv preprint arXiv:2505.16410, 2025

    Guanting Dong, Yifei Chen, Xiaoxi Li, Jiajie Jin, Hongjin Qian, Yutao Zhu, Hangyu Mao, Guorui Zhou, Zhicheng Dou, and Ji- Rong Wen. Tool-star: Empowering llm-brained multi-tool reasoner via reinforcement learning.arXiv preprint arXiv:2505.16410, 2025

  6. [6]

    Multi-agent systems: A survey

    Ali Dorri, Salil S Kanhere, and Raja Jurdak. Multi-agent systems: A survey. Ieee Access, 6:28573–28593, 2018

  7. [7]

    Retool: Reinforcement learning for strategic tool use in llms

    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. arXiv preprint arXiv:2504.11536, 2025

  8. [8]

    Airrag: Activating intrinsic reasoning for retrieval augmented generation via tree-based search

    Wenfeng Feng, Chuzhan Hao, Yuewei Zhang, Jingyi Song, and Hao Wang. Airrag: Activating intrinsic reasoning for retrieval augmented generation via tree-based search. arXiv preprint arXiv:2501.10053, 2025

Show all 210 references
  1. [9]

    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

  2. [10]

    How we built our multi-agent research system

    Jeremy Hadfield, Barry Zhang, Kenneth Lien, Florian Scholz, Jeremy Fox, and Daniel Ford. How we built our multi-agent research system. https://www.anthropic.com/engineering/built-multi-agent-research-system, 2025

  3. [12]

    Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems

    Chaoqun He, Renjie Luo, Yuzhuo Bai, Shengding Hu, Zhen Leng Thai, Junhao Shen, Jinyi Hu, Xu Han, Yujie Huang, Yuxiang Zhang, et al. Olympiadbench: A challenging benchmark for promoting agi with olympiad-level bilingual multimodal scientific problems. arXiv preprint arXiv:2402....

  4. [13]

    Skywork open reasoner 1 technical report

    Jujie He, Jiacai Liu, Chris Yuhao Liu, Rui Yan, Chaojie Wang, Peng Cheng, Xiaoyu Zhang, Fuxiang Zhang, Jiacheng Xu, Wei Shen, Siyuan Li, Liang Zeng, Tianwen Wei, Cheng Cheng, Bo An, Yang Liu, and Yahui Zhou. Skywork open reasoner 1 technical report. arXiv preprint arXiv:2505.2...

  5. [14]

    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

  6. [15]

    Owl: Optimized workforce learning for general multi-agent assistance in real-world task automation, 2025

    Mengkang Hu, Yuhang Zhou, Wendong Fan, Yuzhou Nie, Bowei Xia, Tao Sun, Ziyu Ye, Zhaoxuan Jin, Yingru Li, Qiguang Chen, Zeyu Zhang, Yifeng Wang, Qianshuo Ye, Bernard Ghanem, Ping Luo, and Guohao Li. Owl: Optimized workforce learning for general multi-agent assistance in real-wo...

  7. [16]

    Agentcoder: Multi-agent-based code generation with iterative testing and optimisation

    Dong Huang, Jie M.Zhang, Michael Luck, Qingwen Bu, Yuhao Qing, and Heming Cui. Agentcoder: Multi-agent-based code generation with iterative testing and optimisation. arXiv preprint arXiv:2312.13010, 2023

  8. [17]

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, et al. Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186, 2024

  9. [18]

    Ashraful Islam, Mohammed Eunus Ali, and Md Rizwan Parvez

    Md. Ashraful Islam, Mohammed Eunus Ali, and Md Rizwan Parvez. Mapcoder: Multi-agent code generation for competitive problem solving. Association for Computational Linguistics, 2024

  10. [19]

    Ashraful Islam, Mohammed Eunus Ali, and Md Rizwan Parvez

    Md. Ashraful Islam, Mohammed Eunus Ali, and Md Rizwan Parvez. Codesim: Multi-agent code generation and problem solving through simulation-driven planning and debugging. Association for Computational Linguistics, 2025

  11. [20]

    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. 23

  12. [21]

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

    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. arXiv preprint arXiv:2503.09516, 2025

  13. [22]

    Reveal: Self-evolving code agents via iterative generation-verification, 2025

    Yiyang Jin, Kunzhao Xu, Hang Li, Xueting Han, Yanmin Zhou, Cheng Li, and Jing Bai. Reveal: Self-evolving code agents via iterative generation-verification, 2025. URLhttps://arxiv.org/abs/2506.11442

  14. [23]

    Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension

    Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension. arXiv preprint arXiv:1705.03551, 2017

  15. [24]

    Sequence-level knowledge distillation

    Yoon Kim and Alexander M Rush. Sequence-level knowledge distillation. In Proceedings of the 2016 conference on empirical methods in natural language processing, pages 1317–1327, 2016

  16. [25]

    Natural questions: a benchmark for question answering research

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. Natural questions: a benchmark for question answering research. Transactions of the Association for Computa...

  17. [26]

    Camel: Communicative agents for "mind" exploration of large language model society

    Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for "mind" exploration of large language model society. InThirty-seventh Conference on Neural Information Processing Systems, 2023

  18. [27]

    Websailor: Navigating super-human reasoning for web agent, 2025

    Kuan Li, Zhongwang Zhang, Huifeng Yin, Liwen Zhang, Litu Ou, Jialong Wu, Wenbiao Yin, Baixuan Li, Zhengwei Tao, Xinyu Wang, Weizhou Shen, Junkai Zhang, Dingchu Zhang, Xixi Wu, Yong Jiang, Ming Yan, Pengjun Xie, Fei Huang, and Jingren Zhou. Websailor: Navigating super-human rea...

  19. [28]

    Search-o1: Agentic search-enhanced large reasoning models

    Xiaoxi Li, Guanting Dong, Jiajie Jin, Yuyao Zhang, Yujia Zhou, Yutao Zhu, Peitian Zhang, and Zhicheng Dou. Search-o1: Agentic search-enhanced large reasoning models. arXiv preprint arXiv:2501.05366, 2025

  20. [29]

    Webthinker: Empowering large reasoning models with deep research capability

    Xiaoxi Li, Jiajie Jin, Guanting Dong, Hongjin Qian, Yutao Zhu, Yongkang Wu, Ji-Rong Wen, and Zhicheng Dou. Webthinker: Empowering large reasoning models with deep research capability. arXiv preprint arXiv:2504.21776, 2025

  21. [30]

    Torl: Scaling tool-integrated rl

    Xuefeng Li, Haoyang Zou, and Pengfei Liu. Torl: Scaling tool-integrated rl. arXiv preprint arXiv:2503.23383, 2025

  22. [31]

    Competition-level code generation with alphacode

    Yujia Li, David Choi, Junyoung Chung, Nate Kushman, Julian Schrittwieser, Rémi Leblond, Tom Eccles, James Keeling, Felix Gimeno, Agustin Dal Lago, et al. Competition-level code generation with alphacode. Science, 378(6624):1092–1097, 2022

  23. [32]

    Let’s verify step by step

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. arXiv preprint arXiv:2305.20050, 2023

  24. [33]

    Inference-time scaling for generalist reward modeling

    Zijun Liu, Peiyi Wang, Runxin Xu, Shirong Ma, Chong Ruan, Peng Li, Yang Liu, and Yu Wu. Inference-time scaling for generalist reward modeling. arXiv preprint arXiv:2504.02495, 2025

  25. [34]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017

  26. [35]

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

    Xinji Mai, Haotian Xu, Weinong Wang, 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

  27. [36]

    When not to trust language models: Investigating effectiveness of parametric and non-parametric memories

    Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Daniel Khashabi, and Hannaneh Hajishirzi. When not to trust language models: Investigating effectiveness of parametric and non-parametric memories. arXiv preprint arXiv:2212.10511, 2022

  28. [37]

    Gaia: a benchmark for general ai assistants

    Grégoire Mialon, Clémentine Fourrier, Thomas Wolf, Yann LeCun, and Thomas Scialom. Gaia: a benchmark for general ai assistants. In The Twelfth International Conference on Learning Representations, 2023

  29. [38]

    American invitational mathematics examination (aime) 2024

    Mathematical Association of America (MAA). American invitational mathematics examination (aime) 2024. Competitive mathematics examination, 2024

  30. [39]

    American invitational mathematics examination (aime) 2025

    Mathematical Association of America (MAA). American invitational mathematics examination (aime) 2025. Competitive mathematics examination, 2025

  31. [40]

    Codeforces

    Guilherme Penedo, Anton Lozhkov, Hynek Kydlíˇcek, Loubna Ben Allal, Edward Beeching, Agustín Piqueres Lajarín, Quentin Gallouédec, Nathan Habib, Lewis Tunstall, and Leandro von Werra. Codeforces. Hugging Face, 2025

  32. [41]

    Humanity’s last exam

    Long Phan, Alice Gatti, Ziwen Han, Nathaniel Li, Josephina Hu, Hugh Zhang, Chen Bo Calvin Zhang, Mohamed Shaaban, John Ling, Sean Shi, et al. Humanity’s last exam. arXiv preprint arXiv:2501.14249, 2025

  33. [42]

    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. arXiv preprint arXiv:2210.03350, 2022. 24

  34. [43]

    Toolrl: Reward is all tool learning needs

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

  35. [44]

    Alita: Generalist agent enabling scalable agentic reasoning with minimal predefinition and maximal self-evolution

    Jiahao Qiu, Xuan Qi, Tongcheng Zhang, Xinzhe Juan, Jiacheng Guo, Yifu Lu, Yimin Wang, Zixin Yao, Qihan Ren, Xun Jiang, et al. Alita: Generalist agent enabling scalable agentic reasoning with minimal predefinition and maximal self-evolution. arXiv preprint arXiv:2505.20286, 2025

  36. [45]

    Qwen2.5 technical report, 2025

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, ...

  37. [46]

    ‘smolagents‘: a smol library to build great agentic systems.https://github.com/huggingface/smolagents, 2025

    Aymeric Roucher, Albert Villanova del Moral, Thomas Wolf, Leandro von Werra, and Erik Kaunismäki. ‘smolagents‘: a smol library to build great agentic systems.https://github.com/huggingface/smolagents, 2025

  38. [47]

    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, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

  39. [48]

    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

  40. [49]

    Taskcraft: Automated generation of agentic tasks

    Dingfeng Shi, Jingyi Cao, Qianben Chen, Weichen Sun, Weizhen Li, Hongxuan Lu, Fangchen Dong, Tianrui Qin, King Zhu, Minghao Yang, et al. Taskcraft: Automated generation of agentic tasks. arXiv preprint arXiv:2506.10055, 2025

  41. [50]

    R1-searcher: Incentivizing the search capability in llms via reinforcement learning

    Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. R1-searcher: Incentivizing the search capability in llms via reinforcement learning. arXiv preprint arXiv:2503.05592, 2025

  42. [51]

    Zerosearch: Incentivize the search capability of llms without searching

    Hao Sun, Zile Qiao, Jiayan Guo, Xuanbo Fan, Yingyan Hou, Yong Jiang, Pengjun Xie, Yan Zhang, Fei Huang, and Jingren Zhou. Zerosearch: Incentivize the search capability of llms without searching. arXiv preprint arXiv:2505.04588, 2025

  43. [52]

    Simpledeepsearcher: Deep information seeking via web-powered reasoning trajectory synthesis

    Shuang Sun, Huatong Song, Yuhao Wang, Ruiyang Ren, Jinhao Jiang, Junjie Zhang, Fei Bai, Jia Deng, Wayne Xin Zhao, Zheng Liu, et al. Simpledeepsearcher: Deep information seeking via web-powered reasoning trajectory synthesis. arXiv preprint arXiv:2505.16834, 2025

  44. [53]

    Agent kb: Leveraging cross-domain experience for agentic problem solving

    Xiangru Tang, Tianrui Qin, Tianhao Peng, Ziyang Zhou, Daniel Shao, Tingting Du, Xinming Wei, Peng Xia, Fang Wu, He Zhu, Ge Zhang, Jiaheng Liu, Xingyao Wang, Sirui Hong, Chenglin Wu, Hao Cheng, Chi Wang, and Wangchunshu Zhou. Agent kb: Leveraging cross-domain experience for age...

  45. [54]

    Webshaper: Agentically data synthesizing via information-seeking formalization, 2025

    Zhengwei Tao, Jialong Wu, Wenbiao Yin, Junkai Zhang, Baixuan Li, Haiyang Shen, Kuan Li, Liwen Zhang, Xinyu Wang, Yong Jiang, Pengjun Xie, Fei Huang, and Jingren Zhou. Webshaper: Agentically data synthesizing via information-seeking formalization, 2025. URLhttps://arxiv.org/abs...

  46. [55]

    Qwq: Reflect deeply on the boundaries of the unknown, November 2024

    Qwen Team. Qwq: Reflect deeply on the boundaries of the unknown, November 2024. URL https://qwenlm.github.io/ blog/qwq-32b-preview/

  47. [56]

    Verl-tool: A version of verl to support tool use, 2025

    TIGER-AI-Lab. Verl-tool: A version of verl to support tool use, 2025. URLhttps://github.com/TIGER-AI-Lab/verl-tool . Accessed: 2025-08-04

  48. [57]

    Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. arXiv preprint arXiv:2212.10509, 2022

  49. [58]

    Musique: Multihop questions via single-hop question composition

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. Musique: Multihop questions via single-hop question composition. Transactions of the Association for Computational Linguistics, 10:539–554, 2022

  50. [59]

    Otc: Optimal tool calls via reinforcement learning

    Hongru Wang, Cheng Qian, Wanjun Zhong, Xiusi Chen, Jiahao Qiu, Shijue Huang, Bowen Jin, Mengdi Wang, Kam-Fai Wong, and Heng Ji. Otc: Optimal tool calls via reinforcement learning. arXiv e-prints, pages arXiv–2504, 2025

  51. [60]

    Stepsearch: Igniting llms search ability via step-wise proximal policy optimization

    Ziliang Wang, Xuhui Zheng, Kang An, Cijun Ouyang, Jialu Cai, Yuhang Wang, and Yichao Wu. Stepsearch: Igniting llms search ability via step-wise proximal policy optimization. arXiv preprint arXiv:2505.15107, 2025

  52. [61]

    Corag: A cost-constrained retrieval optimization system for retrieval-augmented generation

    Ziting Wang, Haitao Yuan, Wei Dong, Gao Cong, and Feifei Li. Corag: A cost-constrained retrieval optimization system for retrieval-augmented generation. arXiv preprint arXiv:2411.00744, 2024. 25

  53. [62]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022

  54. [63]

    Browsecomp: A simple yet challenging benchmark for browsing agents, 2025

    Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, and Amelia Glaese. Browsecomp: A simple yet challenging benchmark for browsing agents, 2025. URL https://arxiv.org/abs/2504.12516

  55. [64]

    Autotir: Autonomous tools integrated reasoning via reinforcement learning, 2025

    Yifan Wei, Xiaoyan Yu, Yixuan Weng, Tengfei Pan, Angsheng Li, and Li Du. Autotir: Autonomous tools integrated reasoning via reinforcement learning, 2025. URLhttps://arxiv.org/abs/2507.21836

  56. [65]

    Webdancer: Towards autonomous information seeking agency

    Jialong Wu, Baixuan Li, Runnan Fang, Wenbiao Yin, Liwen Zhang, Zhengwei Tao, Dingchu Zhang, Zekun Xi, Yong Jiang, Pengjun Xie, et al. Webdancer: Towards autonomous information seeking agency. arXiv preprint arXiv:2505.22648, 2025

  57. [66]

    Simpletir: End-to-end reinforcement learning for multi-turn tool-integrated reasoning.https://simpletir.notion.site/report, 2025

    Zhenghai Xue, Longtao Zheng, Qian Liu, Yingru Li, Zejun Ma, and Bo An. Simpletir: End-to-end reinforcement learning for multi-turn tool-integrated reasoning.https://simpletir.notion.site/report, 2025. Notion Blog

  58. [67]

    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

  59. [68]

    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

  60. [69]

    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

  61. [70]

    Auto-rag: Autonomous retrieval-augmented generation for large language models

    Tian Yu, Shaolei Zhang, and Yang Feng. Auto-rag: Autonomous retrieval-augmented generation for large language models. arXiv preprint arXiv:2411.19443, 2024

  62. [71]

    Simplerl-zoo: Investigating and taming zero reinforcement learning for open base models in the wild

    Weihao Zeng, Yuzhen Huang, Qian Liu, Wei Liu, Keqing He, Zejun Ma, and Junxian He. Simplerl-zoo: Investigating and taming zero reinforcement learning for open base models in the wild. arXiv preprint arXiv:2503.18892, 2025

  63. [72]

    Flowmind: automatic workflow generation with llms

    Zhen Zeng, William Watson, Nicole Cho, Saba Rahimi, Shayleen Reynolds, Tucker Balch, and Manuela Veloso. Flowmind: automatic workflow generation with llms. In Proceedings of the Fourth ACM International Conference on AI in Finance, pages 73–81, 2023

  64. [73]

    Evolvesearch: An iterative self-evolving search agent

    Dingchu Zhang, Yida Zhao, Jialong Wu, Baixuan Li, Wenbiao Yin, Liwen Zhang, Yong Jiang, Yufeng Li, Kewei Tu, Pengjun Xie, et al. Evolvesearch: An iterative self-evolving search agent. arXiv preprint arXiv:2505.22501, 2025

  65. [74]

    Aflow: Automating agentic workflow generation

    Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xionghui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, et al. Aflow: Automating agentic workflow generation. arXiv preprint arXiv:2410.10762, 2024

  66. [75]

    Process vs

    Wenlin Zhang, Xiangyang Li, Kuicai Dong, Yichao Wang, Pengyue Jia, Xiaopeng Li, Yingyi Zhang, Derong Xu, Zhaocheng Du, Huifeng Guo, et al. Process vs. outcome reward: Which is better for agentic rag reinforcement learning. arXiv preprint arXiv:2505.14069, 2025

  67. [76]

    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, 36:46595–46623, 2023

  68. [77]

    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

  69. [78]

    Openresearcher: Unleashing ai for accelerated scientific research

    Yuxiang Zheng, Shichao Sun, Lin Qiu, Dongyu Ru, Cheng Jiayang, Xuefeng Li, Jifan Lin, Binjie Wang, Yun Luo, Renjie Pan, et al. Openresearcher: Unleashing ai for accelerated scientific research. arXiv preprint arXiv:2408.06941, 2024

  70. [79]

    Deepresearcher: Scaling deep research via reinforcement learning in real-world environments

    Yuxiang Zheng, Dayuan Fu, Xiangkun Hu, Xiaojie Cai, Lyumanshan Ye, Pengrui Lu, and Pengfei Liu. Deepresearcher: Scaling deep research via reinforcement learning in real-world environments. arXiv preprint arXiv:2504.03160, 2025

  71. [80]

    Agents: An open-source framework for autonomous language agents

    Wangchunshu Zhou, Yuchen Eleanor Jiang, Long Li, Jialong Wu, Tiannan Wang, Shi Qiu, Jintian Zhang, Jing Chen, Ruipu Wu, Shuai Wang, et al. Agents: An open-source framework for autonomous language agents. arXiv preprint arXiv:2309.07870, 2023

  72. [81]

    Symbolic learning enables self-evolving agents

    Wangchunshu Zhou, Yixin Ou, Shengwei Ding, Long Li, Jialong Wu, Tiannan Wang, Jiamin Chen, Shuai Wang, Xiaohua Xu, Ningyu Zhang, et al. Symbolic learning enables self-evolving agents. arXiv preprint arXiv:2406.18532, 2024. 26

  73. [82]

    Oagents: An empirical study of building effective agents, 2025

    He Zhu, Tianrui Qin, King Zhu, Heyuan Huang, Yeyi Guan, Jinxiang Xia, Yi Yao, Hanhao Li, Ningning Wang, Pai Liu, Tianhao Peng, Xin Gui, Xiaowan Li, Yuhui Liu, Yuchen Eleanor Jiang, Jun Wang, Changwang Zhang, Xiangru Tang, Ge Zhang, Jian Yang, Minghao Liu, Xitong Gao, Wangchuns...

  74. [83]

    Scaling test- time compute for llm agents, 2025

    King Zhu, Hanhao Li, Siwei Wu, Tianshun Xing, Dehua Ma, Xiangru Tang, Minghao Liu, Jian Yang, Jiaheng Liu, Yuchen Eleanor Jiang, Changwang Zhang, Chenghua Lin, Jun Wang, Ge Zhang, and Wangchunshu Zhou. Scaling test- time compute for llm agents, 2025. URLhttps://arxiv.org/abs/2...

  75. [84]

    Search for NASA Astronomy Pictures of the Day from the first week of August 2015 to find the relevant image with city lights on the horizon

  76. [85]

    Identify the city shown in that picture

  77. [86]

    Determine the namesake of that city (the person the city is named after)

  78. [87]

    Search for landmark buildings in Chicago named after this person to confirm the connection

  79. [88]

    Find out which architectural firm designed the identified landmark building in Chicago

  80. [89]

    Extract and provide the first name in the architectural firm’s name as of June 2023. </plan> Step 2 <think> The user’s task involves finding a specific NASA Astronomy Picture of the Day (APOD) from the first week of August 2015 that shows the lights of a city on the horizon. F...

  81. [90]

    An archive of all existing APOD pages (current date through

    [Astronomy Picture of the Day Archive 2015 - NASA](https://apod.nasa.gov/apod/archivepix.html) This archive list links to previous daily APOD pages from the current date through January 1, 2015. An archive of all existing APOD pages (current date through

  82. [91]

    [Astronomy Picture of the Day Archive - NASA](https://apod.nasa.gov/apod/archivepixFull.html) 2008 August 30: The View from Everest 2008 August 29: Generations of Stars in W5 2008 August 28: Fermi’s First Light 2008 August 27: IC 5146: The Cocoon Nebula

  83. [92]

    [APOD: 2015 August 15 - Perihelion Approaches - NASA](https://apod.nasa.gov/apod/ap150815.html) Date published: Aug 15, 2015 The stark image of the 4 kilometer wide, double-lobed nucleus in bright sunlight and dark shadows was taken by the Rosetta spacecraft’s science

  84. [93]

    1, 2015 (Credit: NASA/Bill Ingalls)

    [Spot The Station - NASA](https://www.nasa.gov/spot-the-station/) The International Space Station pictured in this 30-second exposure image over Elkton, Virginia, on Aug. 1, 2015 (Credit: NASA/Bill Ingalls). How do I spot the

  85. [94]

    </observation> Step 3 <think> Step 1 of the task is to identify the NASA Astronomy Picture of the Day (APOD) from the first week of August 2015 showing city lights on the horizon

    [2015 August 1 - Stripping ESO 137-001 - APOD](https://apod.nasa.gov/apod/ap150801.html) Date published: Aug 1, 2015 Evident in Hubble’s near visible light data, bright star clusters have formed in the stripped material along the short, trailing blue streaks. </observation> St...

  86. [95]

    Marquette had a population of 20,629 at the

    [Marquette, Michigan - Wikipedia](https://en.wikipedia.org/wiki/Marquette,_Michigan) 29 Marquette is named after Jacques Marquette, a French Jesuit missionary who had explored the Great Lakes region. Marquette had a population of 20,629 at the

  87. [96]

    [About the City - City of Marquette](https://www.marquettemi.gov/commission/about-the-city/) On August 21, 1850, the name was changed to honor Jacques Marquette, the French Jesuit missionary who had explored the region

  88. [97]

    [How did Marquette get its name? - The Mining Journal](https://www.miningjournal.net/news/superior_history/2019/12/how-did-marquette-get-its-name/) Date published: Dec 5, 2019 The name may reflect a local legend that Jesuit missionary Jacques Marquette (1637-1675) preached a s...

  89. [98]

    “Back in the 1600’s he set up several missions, including

    [NMU discusses Father Jacques Marquette, the City of Marquette’s ...](https://www.uppermichiganssource.com/2025/05/19/nmu-discusses-father-jacques- marquette-city-marquettes-namesake/) Date published: May 19, 2025 The presentation focused on Father Jacques Marquette, a French ...

  90. [99]

    [Jacques Marquette - Wikipedia](https://en.wikipedia.org/wiki/Jacques_Marquette) Jacques Marquette SJ sometimes known as Père Marquette or James Marquette, was a French Jesuit missionary who founded Michigan’s first European settlement, ... </observation> Step 4 <think> Reflec...

  91. [100]

    Completed in 1894, the Marquette Building brings Chicago’s early history to life in an artistic and elegant setting

    [Marquette Building | Chicago Architecture Center](https://www.architecture.org/city-tours/marquette-building) Marquette Building . Completed in 1894, the Marquette Building brings Chicago’s early history to life in an artistic and elegant setting

  92. [101]

    [Marquette Building (Chicago) - Wikipedia](https://en.wikipedia.org/wiki/Marquette_Building_(Chicago)) The building was named after Father Jacques Marquette, the first European settler in Chicago, who explored the Chicago region in 1674 and wintered in the

  93. [102]

    [Marquette Building - Chicago Landmarks - Landmark Details](https://webapps1.chicago.gov/landmarksweb/web/landmarkdetails.htm?lanId=1363) The building is named for Jacques Marquette, a French Jesuit missionary and explorer who, in 1674-75, wintered in the area that is now Chicago

  94. [103]

    It was created as a collection of first-class

    [Marquette Building | Chicago Architecture Center](https://www.architecture.org/online-resources/buildings-of-chicago/marquette-building) The Marquette Building stands as an ideal example of a speculative 19th-century commercial office building. It was created as a collection ...

  95. [104]

    In 1975

    [The Historic Marquette Building - MacArthur Foundation](https://www.macfound.org/press/article/historic-marquette-building) Date published: Apr 25, 2016 Designed and built by Holabird & Roche in 1895, it was named for the French Jesuit missionary and explorer Jacques Marquett...

  96. [105]

    Purdy, architects of the firm Holabird & Roche

    [Marquette Building (Chicago) - Wikipedia](https://en.wikipedia.org/wiki/Marquette_Building_(Chicago)) It was designed by William Holabird and Martin Roche, with Coydon T. Purdy, architects of the firm Holabird & Roche

  97. [106]

    Holabird & Root, their successor firm, is in the

    [The Architects: William Holabird, Martin Roche, and Edward Renwick](https://marquette.macfound.org/slide/the-architects.html) William Holabird and Martin Roche formed a firm in 1880, with Edward Renwick as their first employee. Holabird & Root, their successor firm, is in the

  98. [107]

    The building opened in 1895 as the nation’s first modern skyscraper to celebrate,

    [Marquette Building by Wiss, Janney, Elstner Associates, Inc.](https://architizer.com/projects/marquette-building/) The seventeen-story Marquette Building was designed by Holabird & Roche. The building opened in 1895 as the nation’s first modern skyscraper to celebrate,

  99. [108]

    The seventeen story building was designed to resemble a stately

    [marquette building - Harboe Architects](https://www.harboearch.com/marquette-building.html) The Marquette Building was originally designed by Holabird and Roche and completed in 1895. The seventeen story building was designed to resemble a stately

  100. [109]

    [[PDF] MARQUETTE BUILDING](https://buildingchicago.wordpress.com/wp-content/uploads/2013/05/marquette-design-and-construction.pdf) The architectural firm Holabird & Roche was founded in. 1880. • William Holabird & Martin Roche designed the. Marquette Building. Edward Renwick w...

  101. [110]

    The model correctly identified the relevant APOD as the one from August 3-7, 2015, which featured the city lights of Marquette, Michigan

    **Verify the Suggested Answer against the Task Steps:** a. The model correctly identified the relevant APOD as the one from August 3-7, 2015, which featured the city lights of Marquette, Michigan. b. The city is Marquette. c. The namesake of Marquette is Jacques Marquette. d. ...

  102. [111]

    It accurately follows all the required steps and provides the precise piece of information requested

    **Evaluate the Match and Score:** The suggested answer, ’Holabird’, is the correct output of the entire logical chain described in the task. It accurately follows all the required steps and provides the precise piece of information requested. Therefore, it is completely correc...

  103. [112]

    [Eighth Grade (film) - Wikipedia](https://en.wikipedia.org/wiki/Eighth_Grade_(film)) Burnham began writing the screenplay in 2014, initially in an attempt to reflect on his own anxiety and onstage panic attacks he experienced during his career

  104. [113]

    It is based on the

    [Do what matters, no matter what! Factorizing positive activities ...](https://pmc.ncbi.nlm.nih.gov/articles/PMC9490392/) Symptoms of anxiety were measured with the Generalized Anxiety Disorder-7 (GAD-7; Spitzer et al., 2006; German version: Löwe et al., 2008). It is based on the

  105. [114]

    A beautiful, culinary panic attack - Cut To The Take](https://cuttothetake.com/review-the-bear/) Date published: Aug 10, 2024 Review: The Bear

    [Review: The Bear. A beautiful, culinary panic attack - Cut To The Take](https://cuttothetake.com/review-the-bear/) Date published: Aug 10, 2024 Review: The Bear. A beautiful, culinary panic attack ; Release year: 2022 ; Creator: Christopher Storer. ; Directors: Christopher St...

  106. [115]

    rape case in October 2013 attempted to take her own life after she was attacked anonymously

    [[PDF] Unmasking the Teen Cyberbully - IdeaExchange@UAkron](https://ideaexchange.uakron.edu/cgi/viewcontent.cgi?article=2410&context=akronlawreview) distinct effects on social anxiety, symptoms of depression, [and] ... rape case in October 2013 attempted to take her own life a...

  107. [116]

    Aaronson’s clinical interests include panic disorder, generalized anxiety disorder, social anxiety, posttraumatic stress disorder, and

    [What is Depression and How Do I Know If I Have It?](https://adaa.org/learn-from-us/from-the-experts/blog-posts/consumer/what-depression-and-how-do-i- know-if-i-have-it) Date published: May 8, 2017 Dr. Aaronson’s clinical interests include panic disorder, generalized anxiety d...

  108. [117]

    [Eighth Grade (film) - Wikipedia](https://en.wikipedia.org/wiki/Eighth_Grade_(film)) It stars Elsie Fisher as Kayla, a teenager attending middle school who struggles with anxiety but strives to gain social acceptance from her peers

  109. [118]

    Aquamarine

    [10 Movies That Portray What Anxiety Is *Really* Like - Verywell Mind](https://www.verywellmind.com/films-about-anxiety-8723579) Date published: Oct 16, 2024 “Aquamarine” · “Annie Hall” · “Black Swan” · “Inside Out 2” · “Eighth Grade” · “Whiplash” · “Good Will Hunting” · ”Mean Girls“

  110. [119]

    Anxiety is a Nicolas Cage special! Garden State (2004): low level anxiety/

    [Movies where the protagonist suffers from anxiety - Reddit](https://www.reddit.com/r/MovieSuggestions/comments/15p2ozx/movies_where_the_protago- nist_suffers_from_anxiety/) Date published: Aug 12, 2023 Matchstick Men (2003): Protagonist suffers from severe OCD. Anxiety is a N...

  111. [120]

    And to play his onscreen alter ego, he chose a 13-year-old

    [Making of ’Eighth Grade’: How Bo Burnham Brought His Anxiety to ...](https://www.hollywoodreporter.com/movies/movie-features/making-eighth-grade- how-bo-burnham-brought-his-anxiety-screen-1162239/) Date published: Nov 20, 2018 The 28-year-old comedian decided to make an autob...

  112. [121]

    </observation> Step 5 <think> Looking at the problem-solving trace, I can see several critical issues that require reflection

    [Movies about Anxiety - Mental Health Related Films](https://www.therapyroute.com/article/movies-about-anxiety-mental-health-related-films-by-e-sinisi) Date published: Aug 8, 2024 We invite you to explore mental health through movies that raise awareness, open dialogue, presen...

  113. [123]

    United States · Official sites · Official site · Official site (Japan) · Language

    [Eighth Grade (2018) - IMDb](https://www.imdb.com/title/tt7014006/) Details · Release date · August 3, 2018 (United States) · Country of origin. United States · Official sites · Official site · Official site (Japan) · Language

  114. [124]

    Release Date (Theaters): Aug 3, 2018, Wide

    [Eighth Grade | Rotten Tomatoes](https://www.rottentomatoes.com/m/eighth_grade) Original Language: English. Release Date (Theaters): Aug 3, 2018, Wide. Release Date (Streaming): Sep 25, 2018. Box Office (Gross USA): $13.5M. Runtime: 1h 33m

  115. [125]

    RELEASE DATE

    [Eighth Grade | Official Trailer HD | A24 - YouTube](https://www.youtube.com/watch?v=y8lFgF_IjPw) Date published: Mar 14, 2018 EIGHTH GRADE – Now available for rent or purchase (links below). RELEASE DATE ... Comments. 8.5K. a teen movie w/ a character who actually has

  116. [126]

    [Eighth Grade (2018) - Release info - IMDb](https://www.imdb.com/title/tt7014006/releaseinfo/) Release Date: United States January 19, 2018(Sundance Film Festival) United States March 9, 2018(South by Southwest Film Festival)

  117. [127]

    Thirteen-year-old Kayla endures the tidal wave of

    [Everything You Need to Know About Eighth Grade Movie (2018)](https://www.movieinsider.com/m16179/eighth-grade) Eighth Grade in US theaters July 13, 2018 starring Elsie Fisher, Josh Hamilton, Emily Robinson. Thirteen-year-old Kayla endures the tidal wave of ... </observation> ...

  118. [129]

    [what’s the worst anxiety/panic attack you’ve ever had? - Reddit](https://www.reddit.com/r/Anxiety/comments/1bw2kgn/whats_the_worst_anxietypanic_- attack_youve_ever_had/) Date published: Apr 5, 2024 I had several panic attacks that have brought me similar symptoms, but this wa...

  119. [130]

    [What do I do if I am at school and my anxiety starts coming ... - Quora](https://www.quora.com/What-do-I-do-if-I-am-at-school-and-my-anxiety-starts- coming-up-when-I-have-a-panic-attack-I-start-feeling-really-nauseous) Date published: Mar 17, 2021 The last time I had a panic ...

  120. [131]

    [A Social Anxiety: Afraid of People.(2011) This is the documentary I ...](https://www.reddit.com/r/Documentaries/comments/6wdzob/a_social_anxiety_- afraid_of_people2011_this_is_the/) Date published: Aug 27, 2017 My sixth grade English teacher told me a similar story. ... https...

  121. [132]

    panic attacks

    [How to Cure Anxiety - One Workaholic’s Story, Six Techniques That ...](https://tim.blog/2014/02/19/anxiety-attacks-2/) Date published: Feb 19, 2014 Write down everything you worry about on a daily or weekly basis. 3-5 words per item is fine. For instance, “panic attacks.” 2. ...

  122. [133]

    It suggests that the user may have followed a broken link or typed the URL incorrectly

    **Web Search Query-Related Information:** - The web page indicates that the requested content was not found. It suggests that the user may have followed a broken link or typed the URL incorrectly

  123. [134]

    **Clues and Ideas-Related Information:** - The page does not contain any content related to anonymous 8th grade blogs, movie-induced anxiety attacks, or mentions of a mother and a crush from the specified time period (2005-2013)

  124. [135]

    - [LJ Home](URL): A link to the LiveJournal home page

    **Relevant Links with Descriptions:** - [Recent Entries of this journal](URL): A link to view recent entries in the journal. - [LJ Home](URL): A link to the LiveJournal home page. - [FAQ](URL): A link to the LiveJournal frequently asked questions page. - [Site Map](URL): A lin...

  125. [136]

    While his main intention was to write a story about

    [Eighth Grade (film) - Wikipedia](https://en.wikipedia.org/wiki/Eighth_Grade_(film)) As a touring comedian, Bo Burnham suffered from panic attacks related to his performances since 2013. While his main intention was to write a story about

  126. [137]

    anxiety and experiencing panic attacks on stage

    [Bo Burnham - Wikipedia](https://en.wikipedia.org/wiki/Bo_Burnham) ... anxiety and experiencing panic attacks on stage. He went on to make his filmmaking debut as the writer and director of the drama film Eighth Grade (2018)

  127. [139]

    Months passed and

    [The Darkest Term: Teacher Stress and Depression](https://teachingbattleground.wordpress.com/2013/12/17/the-darkest-term-teacher-stress-and-depression/) Date published: Dec 17, 2013 I went to the doctors because I was ill a lot and, once I’d explained symptoms, he medicated me...

  128. [140]

    anxiety attacks and episodes of insanity

    [Why I went from being a top student to an expelled dropout](https://acestoohigh.com/2013/07/23/why-i-went-from-being-a-top-student-to-an-expelled- dropout/) Date published: Jul 23, 2013 ... anxiety attacks and episodes of insanity. Now I am a suicidal delinquent, only want to...

  129. [141]

    If I read a facebook/twitter/

    [True Story: I Have A Severe Phobia - Yes and Yes](https://www.yesandyes.org/2012/06/true-story-i-have-severe-phobia.html) My anxiety and panic attacks would leave me scared that I was going to throw up, not that I was having a heart attack or dying. If I read a facebook/twitter/

  130. [142]

    Panic Attack

    [Grading The Intensity Of The Types Of Panic Attacks - Mind Ease](https://mindease.io/wellness-blog/types-of-panic-attacks-intensity-scale) Date published: Aug 22, 2023 “Panic Attack.” [2]: National Institute of Mental Health. “Panic Disorder: When Fear Overwhelms.” [3]: Mayo ...

  131. [143]

    panic attack that put me out of commission for the rest of the day

    [Emily Gould - Exposed - Blog-Post Confidential - Gawker](https://www.nytimes.com/2008/05/25/magazine/25internet-t.html) Date published: May 25, 2008 ... panic attack that put me out of commission for the rest of the day. I started having panic attacks — breathless bouts of te...

  132. [144]

    social/emotional denial of service attack

    [Blog Archive » A low-tech solution - Shtetl-Optimized](https://scottaaronson.blog/?p=6576) Date published: Jul 19, 2022 I started blogging in October 2005. It took until July 2022 for me to endure a full-scale “social/emotional denial of service attack” (not

  133. [145]

    [[PDF] ONE TEACHER’S AUTOBIOGRAPHICAL NARRATIVE INQUIRY ...](https://open.library.ubc.ca/media/stream/pdf/24/1.0353195/4) was guided, after a panic attack in 2012, in how to practice self-compassion and acceptance that I began to feel the grip of my anxiety disorder lessen. I ...

  134. [146]

    [what’s the worst anxiety/panic attack you’ve ever had? - Reddit](https://www.reddit.com/r/Anxiety/comments/1bw2kgn/whats_the_worst_anxietypanic_- attack_youve_ever_had/) Date published: Apr 5, 2024 I had several panic attacks that have brought me similar symptoms, but this wa...

  135. [147]

    anxiety attacks and episodes of insanity

    [Why I went from being a top student to an expelled dropout](https://acestoohigh.com/2013/07/23/why-i-went-from-being-a-top-student-to-an-expelled- dropout/) Date published: Jul 23, 2013 As soon as my grades started to decline, which was around 8th grade ... anxiety attacks an...

  136. [148]

    [Anxiety Symptoms, Mindfulness and the Enlargement of the Self](https://www.afterpsychotherapy.com/anxiety-symptoms-and-mindfulness/) Date published: Jan 19, 2012 In an early post about anxiety symptoms, I discussed my personal experience with a panic attack several years ago

  137. [149]

    She was no longer engaged with friends or family and reported not

    [Author: Sydney Sparrow - Family Service of Roanoke Valley](https://fsrv.org/author/guestwriters/page/6/) She was also suffering panic attacks, in which she felt anxious and afraid of dying. She was no longer engaged with friends or family and reported not

  138. [150]

    [What do I do if I am at school and my anxiety starts coming ... - Quora](https://www.quora.com/What-do-I-do-if-I-am-at-school-and-my-anxiety-starts- coming-up-when-I-have-a-panic-attack-I-start-feeling-really-nauseous) Date published: Mar 17, 2021 What should I do if I have a...

  139. [151]

    There is no direct mention of an 8th-grade student experiencing anxiety attacks or any anonymous blog posts from 2005-2013

    **Web Search Query-Related Information:** - The provided web page content is from DVD Talk, a site dedicated to reviews and discussions of DVDs, Blu-rays, and related media. There is no direct mention of an 8th-grade student experiencing anxiety attacks or any anonymous blog p...

  140. [152]

    The reviews are primarily focused on films and their technical qualities

    **Clues and Ideas-Related Information:** - The page contains a list of recent reviews and articles, but none of them appear to match the specific criteria of an 8th-grade student’s blog post about anxiety attacks after watching a movie. The reviews are primarily focused on fil...

  141. [153]

    **Relevant Links with Descriptions:** - [DVD Talk Home Page](https://www.dvdtalk.com/) - [Review of Pulp Fiction (4K Ultra HD)](https://www.dvdtalk.com/reviews/11657/pulp-fiction-4k-ultra-hd-steelbook/) - [Review of The Power of the Dog (Criterion Collection)](https://www.dvdt...

  142. [155]

    anonymous 8th grade blog movie anxiety attack mother crush dvdtalk review

    [rulins/fast_eval · Datasets at Hugging Face](https://huggingface.co/datasets/rulins/fast_eval) Their first blog post mentions the writer is in 8th grade. This post also mentions that they had anxiety attacks because they were watching a specific movie. </observation> Step 14 ...

  143. [156]

    Parse the input: Read N and the N strings S_1, S_2, ..., S_N

  144. [157]

    Find M = the maximum length among all S_i

  145. [158]

    Prepare a matrix of size M x N filled with ’*’ characters

  146. [159]

    − Specifically, for each character S_i[j], place it at (M − j − 1, N − i) in the matrix

    For each string S_i (1 <= i <= N), place its characters in reverse order in the appropriate column: − The first character of S_i should go to the last row of its column, the second to the one above, etc. − Specifically, for each character S_i[j], place it at (M − j − 1, N − i)...

  147. [160]

    no trailing *

    After filling, for each row (T_i), remove any trailing ’*’ characters to satisfy the "no trailing *" condition

  148. [161]

    N = {N}") print(f

    Print each row as a string. </plan> <think> Let’s implement step 1 and 2: read the input, and compute N, the strings, and M (the max length). We should also prepare a test to ensure our parsing and max−length finding works. </think> <code> ‘‘‘py # Test input parsing and max le...

  149. [162]

    The first output line is built from the *last* characters of each string (properly padded with ’*’ where needed)

  150. [163]

    Test 1 Output:

    The vertical writing is such that the *leftmost* column in the output corresponds to the *bottom* row of the input strings. Therefore, our row order in the result should be reversed before printing. Let’s fix this by reversing the result list before output. </reflection> <thin...

  151. [164]

    - Lett=time (in minutes) spent in the coffee shop

    Assign variables: - Lets=walking speed in km/h. - Lett=time (in minutes) spent in the coffee shop. - Walking distance = 9 km

  152. [165]

    - At speeds+2: walking time= 9 s+2 hours; total time= 9 s+2 + t 60 hours=2+ 24 60 =2.4hours

    Express total time equations for both scenarios: - At speeds: walking time= 9 s hours; total time= 9 s + t 60 hours=4hours. - At speeds+2: walking time= 9 s+2 hours; total time= 9 s+2 + t 60 hours=2+ 24 60 =2.4hours

  153. [166]

    Set up two equations using those relationships

  154. [167]

    Solve the system of equations to findsandt

  155. [168]

    - Total time including coffee break= 9 s+ 1 2 + t 60 hours

    Compute the new total time when Aya walks ats+ 1 2 kilometers per hour: - New walking time= 9 s+ 1 2 hours. - Total time including coffee break= 9 s+ 1 2 + t 60 hours

  156. [169]

    s = {s_val}

    Convert this total time to minutes for the final answer. </plan> <think> Let’s start by expressing the two main equations based on the problem statement: First scenario (speeds): 9 s + t 60 =4 Second scenario (speeds+2): 9 s+2 + t 60 =2.4 Subtracting the second from the first ...

  157. [170]

    Find the density of honey and mayonnaise at 25 celsius from a reliable chemistry source

  158. [171]

    Calculate the mass of 1 gallon of honey and 1 gallon of mayonnaise

  159. [172]

    Calculate the mass of 1 cup of honey (in grams or ounces)

  160. [173]

    Iteratively subtract the mass of a cup of honey from the original mass of the gallon, counting the number of cups removed, until the remaining mass of honey is less than that of the mayonnaise

  161. [174]

    Cups to remove:

    Report the number of cups required. I will first search for the densities of honey and mayonnaise at 25 celsius from a reliable chemistry source. </think> <web_search> Density of honey and mayonnaise at 25 celsius </web_search> <observation> [url:https://physics.info/density/]...

  162. [175]

    think: • Format: <think> [step-by-step reasoning] </think> • Function Description: –Provide your step by step reasoning process

  163. [176]

    • Function Requirements –Singleplanfunction only, output as the first function

    plan: • Format: <plan> [high-level steps] </plan> • Function Description: –First make sure you understand the mathematical problem; –Identify the mathematical concepts, theorems, or techniques needed; –Break down the problem into logical steps (e.g., simplification, substituti...

  164. [177]

    –No file operations

    code: • Format: <code> ```py code snippet with'print()' ``` </code> • Function Description: 47 –Use for numerical calculations, symbolic computation, or verification of mathematical results –Can be used to test conjectures, check edge cases, or visualize patterns –Must useprin...

  165. [178]

    observation: • Format: <observation> [Code Execution results, including stdout and stderr.] </observation> • Function Description: –Returns thecodeexecution results by an external python executor

  166. [179]

    –Check if the calculations are mathematically sound and support your approach

    reflection: • Format: <reflection> [Your mathematical reflections] </reflection> • Function Description: –Verify whether the computational results confirm your mathematical reasoning. –Check if the calculations are mathematically sound and support your approach. –Identify if t...

  167. [180]

    answer: • Format: <answer> \boxed{The final answer goes here.} </answer> Requirements:

  168. [183]

    4.reflectionreviews thecodeand theobservation, whilethinkconsiders the nextcodeaccording to plans

    Special Token Restriction: <plan>, <code>, <observation>, <reflection> and <answer> are special tokens and must not appear in free text, especially not within thethinkfunction. 4.reflectionreviews thecodeand theobservation, whilethinkconsiders the nextcodeaccording to plans. D...

  169. [184]

    • Function Requirements: –Callthinkbefore anycodeoranswerfunction

    think • Format: <think> [step-by-step reasoning] </think> • Function Description: –Provide your step-by-step reasoning process.thinkin different locations may focus on different targets. • Function Requirements: –Callthinkbefore anycodeoranswerfunction. –Follow the plan, decid...

  170. [185]

    • Function Requirements: –Singleplanfunction only, output as the first function

    plan • Format: <plan> [high-level steps] </plan> • Function Description: –First make sure you understand the task; –Break down the programming task into atomic, sequential sub-tasks; –For each sub-task, decide on the most efficient way at a high level; –Provide integration ste...

  171. [186]

    Use Format 2 for code that usessys.stdinorinput()to get test input

    code 48 • Use Format 1 for code with test input written in the code. Use Format 2 for code that usessys.stdinorinput()to get test input. (a) Format 1: Only Python markdown <code> ```py code snippet without sys.stdin ``` </code> (b) Format 2: A Python markdown and a sh markdown...

  172. [187]

    observation • Format: <observation> [Code Execution results, including stdout and stderr.] </observation> • Returns thecodeexecution results by an external Python executor

  173. [188]

    expected results

    reflection • Format: <reflection> [Your reflections] </reflection> • Function Description: –Verifyobservationresult vs. expected results. –Explain why the code snippet execution result is wrong or correct. –Find potential bugs or edge cases. Decide whether more test cases shou...

  174. [189]

    • No example usage or test cases

    answer • Format: <answer> ```py [A complete code snippet] ``` </answer> • Include only the essential solution code necessary for the given task. • No example usage or test cases. • Ensure the code is readable and well-commented. Requirements:

  175. [190]

    Always followplan, (think,code,observation,reflection)*N,think,answersequences

  176. [191]

    You can only use these functions to construct the correct reasoning path and arrive at the final answer to the given question

  177. [192]

    rationale

    Special Token Restriction: <plan>, <code>, <observation>, <reflection> and <answer> are special tokens and must not appear in free text, especially not within thethinkfunction. 4.reflection reviews the code and the observation, while think considers the next code according to ...

  178. [193]

    Start with<think>and end with</think>

    think: Before using any plan, tool, reflection, or answer functions, you must use the think function to provide reasoning, arguments, and procedural steps for the function you intend to use next. Start with<think>and end with</think>

  179. [194]

    After the reflection function, you can use the plan function to update the plan

    plan: Given a given question, you must break it down into very detailed, fine-grained sub-questions to be executed using the tool function. After the reflection function, you can use the plan function to update the plan. Start with<plan>and end with</plan>

  180. [195]

    The tool label should be replaced with the exact tool name from the tool list below

    tool: You can use any tool from the tool list below to find information relevant to answering the question. The tool label should be replaced with the exact tool name from the tool list below

  181. [196]

    observation: The observation returned after using the tool

  182. [197]

    reflection: You evaluate the trajectory of the historical algorithm, effectively guiding the direction of your work towards the optimal path

  183. [198]

    suggested_answer: Based on the historical trajectory, you can come up with a suggested answer without checking the answer again

  184. [199]

    You will reflect on your historical trajectory and give your reasoning and thinking based on the credibility of the suggested answer

    double_check: After giving the suggested answer, you will do this step. You will reflect on your historical trajectory and give your reasoning and thinking based on the credibility of the suggested answer. If you are not confident in the suggested answer, you should rethink an...

  185. [200]

    The code will be executed in a sandbox environment and the results will be printed

    code: When dealing with precise calculations or data processing, you must use the code function to verify and validate your answers. The code will be executed in a sandbox environment and the results will be printed. Start with <code> followed by ```python and end with ``` fol...

  186. [201]

    Pi approximation: result

    answer: After checking the answer again and being 100 percent sure of the result, you will give the answer. Here is a list of some tools you can use: 1.<web_search> Search query that the web search tool needs to get information from the web</web_search>, for example: <web_sear...

  187. [202]

    If the information is not relevant to the query, you should search again with another search query until you get enough information and are very confident in getting the final answer

  188. [203]

    If you want to get other related information from the url, you can use crawl_page to crawl another url

  189. [204]

    If the information contains some deeper hints, you can use web_search or crawl_page again in a deeper loop based on the hints

    If you want to do a deeper search, you can first use the web_search tool to return a list of urls, and then use crawl_page to crawl a specific url to get detailed information. If the information contains some deeper hints, you can use web_search or crawl_page again in a deeper...

  190. [205]

    When dealing with precise calculations, numerical analysis, or any task requiring computational verification, you MUST use the code tool to verify your results before providing an answer

  191. [206]

    Use print() functoin to get the expected output you want! Trajectory Description

    When you call the Python executor, you must enclose your code in delimiters, that is, ```python your code ```, and then place <code></code> on the outside. Use print() functoin to get the expected output you want! Trajectory Description

  192. [207]

    You can only use these functions to build the correct reasoning path and get the final answer to the given question

  193. [208]

    Based on the result of the planning function, you can use the tool function multiple times to collect sufficient external knowledge before formulating your response

  194. [209]

    Function Correlation Description

    Special tag restrictions: <think>, <plan>, <web_search>, <crawl_page>, <code>, <observation>, <reflection>, <double_check>, <suggested_answer>and<answer>are special tags and must not appear in free text, especially in the think function. Function Correlation Description

  195. [210]

    Before each use of the plan, web_search, crawl_page, code, reflection, double_check or suggests_answer function, you must use the think function

  196. [211]

    If any scoring criteria in Reflection is poor, you need to re-plan

    You can use the Reflection function at any time. If any scoring criteria in Reflection is poor, you need to re-plan

  197. [212]

    If your <double_check> Score < 3, you should re-plan and arrange your thinking and reasoning process until you come up with your <suggested_answer> again

    Before getting <answer>, you should return <suggested_answer> first, and then return the suggested answer with a score >= 3 as the answer. If your <double_check> Score < 3, you should re-plan and arrange your thinking and reasoning process until you come up with your <suggeste...

  198. [213]

    Answer Tips

    When the question involves precise calculations, statistical analysis, or any mathematical operations, you MUST use the code function to verify your calculations before providing the final answer. Answer Tips

  199. [214]

    The answer should be as concise as possible and avoid detailed descriptions

    Do not give an answer easily unless you are absolutely sure. The answer should be as concise as possible and avoid detailed descriptions. For example, <answer>Beijing</answer>

  200. [215]

    If you are not sure, you must think, re-plan and try to find the definite answer based on the existing information before giving the final answer

    You must give a definite answer. If you are not sure, you must think, re-plan and try to find the definite answer based on the existing information before giving the final answer. The final answer cannot be insufficient or uncertain. The question must have a definite answer. T...

  201. [2013]

    Second, to find the first name and surname of a person who reviewed that movie on DVD Talk in February between 2000 and 2008

    where they mention being in 8th grade, having anxiety attacks from watching a specific movie (released in the US between 2002 and 2010), and other personal details. Second, to find the first name and surname of a person who reviewed that movie on DVD Talk in February between 2...

  202. [2018]

    I should investigate this further by searching for ’Eighth Grade movie release date’ to clarify if this is the same movie or a different one, as this could be a crucial lead

    This suggests there might be confusion with another movie titled ’Eighth Grade’ or that the term is being associated with coming-of-age films. I should investigate this further by searching for ’Eighth Grade movie release date’ to clarify if this is the same movie or a differe...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.