Pith. sign in

REVIEW 4 major objections 6 minor 4 cited by

O$^2$-Searcher: A Searching-based Agent Model for Open-Domain Open-Ended Question Answering

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A 3B-parameter reinforcement-learning search agent can outperform much larger prompted LLM agents on open-ended multi-finding questions while matching 7B-parameter search agents on closed-ended factoid QA.

desk verdict Solid closed-ended search-agent work and a worthwhile open-ended training recipe; the open-ended superiority claim currently rests on a self-referential benchmark and needs independent targets before it can be believed. read the letter →

arxiv 2505.16582 v2 pith:BXRD33ND submitted 2025-05-22 cs.CL cs.AI

Jianbiao Mei , Tao Hu , Daocheng Fu , Licheng Wen , Xuemeng Yang , Rong Wu , Pinlong Cai , Xinyu Cai
show 6 more authors
Xing Gao Yu Yang Chengjun Xie Botian Shi Yong Liu Yu Qiao
This is my paper
classification cs.CLcs.AI
keywords open-endedquestionansweringsearchagentreinforcementlearninggrouprelativepolicyoptimizationretrieval-augmentedgenerationfactoidgroundedlanguagemodels
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

The paper sets out to show that a small language model can become a capable open-domain search agent by learning when to search, what to query, and how to synthesize several findings into one answer, rather than by memorizing more facts. It introduces O2-Searcher, built on a 3B-parameter backbone and trained with reinforcement learning in a locally simulated search environment built from cached web pages and Wikipedia. Its central claim is that this agent substantially outperforms leading LLM agents, including much larger commercial ones, on a new open-ended benchmark called O2-QA, and that on closed-ended benchmarks it is state of the art among similarly sized models and comparable to 7B-parameter search agents. The method matters because open-ended questions have no single correct answer, so they have mostly been left out of both training and evaluation for search agents. If the claim holds, cheap small models can be trained in simulation and then used for live web search on questions that require gathering diverse, up-to-date information.

What carries the argument

The central object is the locally simulated search environment coupled with a unified GRPO training loop. The agent emits <think>, <search>, and <answer> actions, and the environment returns condensed learnings from a cache of curated web pages for open-ended questions and a Wikipedia index for closed-ended ones. Three rewards shape open-ended behavior: a format reward for well-formed, non-duplicated Markdown findings; a diversity reward that penalizes semantically redundant queries; and a factual reward that matches predicted findings to reference findings with the Hungarian algorithm and scores the matched pairs by embedding F1. The closed-ended reward is a rule-based exact-match check inside the answer tags. Together these signals teach a single 3B model to identify question type, choose how many queries and search rounds to use, and decide when to stop searching.

What would settle it

Rerun O2-QA with reference answers written by human experts from sources outside the cached corpus; if O2-Searcher's margin over the prompted baselines collapses or reverses, the claimed open-ended superiority is an artifact of reward alignment rather than general answering skill.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that open-ended, multi-finding question answering can be turned into a reinforcement-learning problem and solved by a 3B-parameter model. O2-Searcher interacts with a local search engine: it reasons inside <think> tags, issues up to five <query> tags per <search> round, receives condensed <learnings> from a cached corpus, and finally emits a Markdown list of key findings in <answer> tags. The paper constructs O2-QA, 300 manually curated open-ended questions across domains such as AI, philosophy, geopolitics, healthcare, and the metaverse, with roughly 30,000 cached web pages; 240 questions are used for training and 60 for testing. Against prompt-engineered commercial and open-source agents, SFT baselines, and previous RL search baselines, O2-Searcher reports the highest embedding-F1 and LLM-judged LFS scores on O2-QA, and it maintains comparable performance when switched to live web search. On closed-ended benchmarks, trained on only 1,200 NQ and HotpotQA questions, it reports the best average exact match among 3B-parameter systems and an average close to a 7B search agent trained on roughly 190,000 samples.

Load-bearing premise

The open-ended results rest on treating O2-QA's reference answers, which are produced by a prompted LLM searching the same cached pages the agent trains on, as an independent standard of correctness.

Editorial extensions

If this is right

  • If the central claim is right, search competence is learnable with far less data, since 1,200 closed-ended RL questions match a 7B search agent trained on roughly 190,000 samples.
  • Small models can compete with much larger commercial agents on open-ended synthesis, which lowers the compute and cost bar for deploying grounded question answering.
  • Unified training over open-ended and closed-ended data stabilizes RL and produces adaptive search behavior, with more queries and search rounds for multi-hop and open-ended questions.
  • Training in a local simulated search environment transfers to live web search with comparable scores, suggesting simulation is a sufficient and much cheaper training ground.
  • RL-trained search agents outperform SFT-trained ones on the same data, implying that rollout-time reward feedback, not just imitation, is what builds the search skill.

Reading between the lines

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

  • Editorial inference: because the O2-QA reference answers are generated by a prompt-engineered LLM agent searching the same cache the trained agent sees, and the factual reward compares against those same answers, the open-ended scores may measure alignment with that generator rather than absolute answer quality.
  • Editorial inference: the composite reward design is generic enough to be applied to other multi-finding generation tasks such as report writing, which the paper sketches but does not evaluate.
  • Editorial inference: the smooth transfer from local cache to open web suggests the agent learned a general search policy, but the fixed cache also leaves open the possibility that it overfit to query patterns specific to the 300 curated questions.
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

4 major / 6 minor

Summary. The paper proposes O2-Searcher, a 3B-parameter LLM agent trained with GRPO to interact with a locally simulated search engine for both closed-ended and open-ended question answering. The method uses a unified reward combining format, diversity, and embedding-F1 factual rewards for open-ended answers, and exact-match rewards for closed-ended answers. The authors construct O2-QA, a 300-question open-ended benchmark, and report that O2-Searcher surpasses much larger prompted agents on O2-QA while matching 7B Search-R1 on closed-ended benchmarks such as NQ, HotpotQA, TriviaQA, and others.

Significance. If the open-ended results were measured against independent human-verified targets, the paper would make a meaningful contribution: a small RL-trained search agent that matches a 7B agent trained on far more closed-ended data, plus a reusable local search environment and open-sourced code. The closed-ended experiments on standard benchmarks are a concrete strength, and the ablation in Table 3 gives useful evidence about reward design. However, the open-ended benchmark construction currently conflates the gold-answer generation, the training reward, and the evaluation metric, so the central open-ended claim is not yet supported.

major comments (4)
  1. [Section 2.2, Section 2.4, Section 3.2] Section 3.2 says O2-QA answers are 'derived directly from the content collected during the search process within the rollout of the prompt-engineered LLM agent' from Section 2.2; Section 2.4 uses the embedding-F1 of Eqs. (8)-(10) as the RL factual reward; and Section 3.2 then uses that same F1 as the primary evaluation metric. The gold answers, the training reward, and the headline metric are therefore all generated by one pipeline, and the policy is trained to imitate the same agent that produced the gold answers. The superiority in Table 1 is thus evidence of reward alignment rather than of general open-ended answering skill. Please re-validate with human-verified gold answers and an independent metric (e.g., human preference) before claiming general open-ended superiority.
  2. [Section 2.2 and Section 3.2] Section 2.2 states that 'all retrieved content is cached to construct the retriever index' for the 300 open-ended questions; Section 3.2 splits those 300 into 240 training and 60 test questions. Consequently, pages retrieved specifically for the 60 test questions are in the local MeiliSearch index during evaluation, so the agent can retrieve the exact documents from which the gold answers were distilled. This is a form of test-set leakage that inflates O2-QA scores. Please rebuild the index excluding all pages retrieved for test questions (or demonstrate that such pages do not affect retrieval rankings), and clarify whether any cold-start or SFT trajectories for the 60 test questions were used.
  3. [Tables 1 and 2] All results are single runs on a 60-question open-ended test set, with no standard deviations, confidence intervals, or significance tests. Given the small test size and the metric-alignment issues above, the reported gaps (e.g., F1 0.2236 vs. 0.1012 in Table 1) are hard to interpret without uncertainty estimates. Please report multiple RL seeds and bootstrap confidence intervals for the open-ended results, and ideally for the closed-ended results in Table 2 as well.
  4. [Section 3.1 and Appendix A.2, Table 3] Section 3.1 says the main model is trained for 200 RL steps, while Table 3 in Appendix A.2 trains all variants for 300 steps and reports a full-reward open-ended F1 of 0.2523, which is higher than the 0.2236 reported in Table 1 for the same configuration. Please state which checkpoint is used in Table 1 and align the ablation training budgets with the main experiments so the comparisons are apples-to-apples.
minor comments (6)
  1. [Section 2.4] In the first sentence of the reward-design overview, 'we combine closed-ended questions with closed-ended questions' should read 'open-ended and closed-ended questions.'
  2. [Table 2] In the Search-R1-base row, the numbers '0.5870.4350.273' are concatenated; insert separators for TriviaQA, PopQA, and 2Wiki.
  3. [Equation (3)] The phrase 'potentially constrained, e.g. u03b11 + u03b12 = 1, treating u03b12 as a separate penalty weight' is unclear; please state exactly which hyperparameters are constrained and what values are used.
  4. [Figure 3] The examples labelled (a) open-ended and (b) closed-ended appear reversed: the left panel answers '2003' (closed-ended) and the right panel lists sustainable-manufacturing findings (open-ended).
  5. [Section 3.2] LFS uses Doubao-1.5-pro-32k, which is also a baseline in Table 1; using the same model as judge and competitor invites evaluator-bias concerns. Consider a judge model not among the baselines or report agreement with human ratings.
  6. [Section 4] The claim of being 'the first work exploring LLM evaluation and training on open-ended questions' overstates novelty; prior open-ended QA benchmarks exist. Suggest softening or citing them.

Circularity Check

2 steps flagged · score 6.0 of 10

O2-QA open-ended 'prediction' is partly circular: the gold answers are LLM-agent outputs from the same cached corpus, and the reported F1 is the same embedding-F1 used as the RL reward.

  1. self definitional [Section 3.2, 'Benchmarks' (O2-QA construction)]
    "we construct an open-ended dataset, termed O2-QA, derived from and intrinsically linked to the knowledge corpus developed in Sec. 2.2. ... The answers for this dataset are derived directly from the content collected during the search process within the rollout of the prompt-engineered LLM agent, as described in Sec. 2.2."

    The reference answers for O2-QA are defined by the output of a prompt-engineered LLM agent searching the same locally cached web corpus that forms the agent's training/evaluation environment. There is no independent human-verified ground truth for open-ended answer quality, so the benchmark measures agreement with a particular LLM-agent-derived target distribution rather than an external standard. The paper itself says O2-QA is 'intrinsically linked' to the knowledge corpus, making the benchmark target and the training/search environment the same constructed object.

  2. fitted input called prediction [Section 2.4, 'Reward Design', Eqs. (8)-(11); Section 3.2, 'Metrics']
    "To evaluate the factual correctness of extracted finding items, we compute the F1 score by comparing the items derived from the predicted answer apred against those from the ground-truth answer agt. ... the aggregated F1 score based on the filtered pairs as the final factual reward ro,f1. ... For open-ended questions, evaluation relies on the F1 score (aligned with the RL training reward) and LLM-assessed Finding Similarity (LFS)."

    The RL objective's factual reward is an embedding-F1 against the LLM-agent-derived ground-truth items, and the reported open-ended evaluation metric is the same F1 score, explicitly 'aligned with the RL training reward.' Therefore the headline open-ended F1 gain is a measure of how well the model optimized the very same target used to score it. The evaluation does not supply an independent check of the open-ended superiority claim; it partially reduces to reward-target fitting by construction.

full rationale

The circularity is confined to the open-ended O2-QA evaluation, which is the paper's central contribution. Section 3.2 states that O2-QA answers are derived from the content collected during the rollout of the prompt-engineered LLM agent over the same knowledge corpus used for training, and Section 2.4 defines the RL factual reward as an embedding-F1 against those same answers. Section 3.2 then reports the open-ended F1 as 'aligned with the RL training reward.' Thus the open-ended benchmark's target and the model's optimization objective are the same object, so the claim that O2-Searcher 'significantly surpasses leading LLM agents' on O2-QA is at least partly a measure of reward alignment rather than independent open-ended answer quality. The closed-ended results on NQ, HotpotQA, TriviaQA, PopQA, 2WikiMultiHopQA, Musique, and Bamboogle use standard external ground truth and are not affected by this circularity; they provide independent evidence for the method's search-and-reasoning effectiveness. No load-bearing self-citation chain was found: Search-R1 and other cited methods are external works without author overlap. A score of 6 reflects that one central 'prediction' (open-ended superiority) is partially constructed by the benchmark and metric design, while the closed-ended claims retain independent content.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central open-ended claim rests on hand-tuned reward hyperparameters and on assumptions that the benchmark ground truths and the embedding-F1 metric capture answer quality. There are no newly postulated physical or theoretical entities. The largest burden is the domain assumption that LLM-generated ground truths from the same corpus are valid external targets.

free parameters (5)
  • Reward weights {alpha_0, alpha_1, alpha_2} = {0.5, 0.5, 3}
    Hand-set weights for format reward in Eq. 3; no sensitivity analysis beyond component ablation in Table 3.
  • Similarity weights {w_0, w_1, w_2} = {0.5, 0.3, 0.2}
    Hand-set weights for the pairwise similarity score s in Eq. 4.
  • Final reward weights {gamma_0, gamma_1, gamma_2} = {0.4, 0.4, 0.2}
    Hand-set weights for the weighted sum of format, diversity, and factual rewards in Eq. 11.
  • Similarity thresholds s_thr and s_theta = 0.6 and 0.75
    Hand-set thresholds for detecting similar findings and for matching items in the F1 reward.
  • Training data mix and size = 240 open-ended + 1200 closed-ended questions, 4:1 sampling ratio
    Chosen by the authors; affects the balance between open-ended and closed-ended capabilities.
assumptions (5)
  • domain assumption Open-ended answer quality is measurable by F1 between sets of key findings via embedding similarity.
    This is the core of both the RL reward (Eq. 10) and the evaluation metric; no human validation is provided for this proxy.
  • domain assumption Ground-truth answers produced by a prompt-engineered LLM agent plus commercial LLM distillation are accurate and complete.
    Section 2.2 and Section 3.2 describe this process; errors in these targets propagate to training and evaluation.
  • domain assumption The locally cached corpus, built from Serper API and condensed by commercial LLMs, adequately simulates the open web.
    Section 2.2; the paper does compare local vs. web search, but training is done only on the local corpus.
  • domain assumption The 300 manually curated O2-QA questions are representative of open-ended questions in the real world.
    Section 2.2 and Section 4 acknowledge sampling bias from manual curation.
  • standard math GRPO is a valid optimization algorithm for this RL setup.
    Eq. 1 uses the standard GRPO objective from DeepSeekMath; the paper does not prove convergence but relies on established results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of O$^2$-Searcher: A Searching-based Agent Model for Open-Domain Open-Ended Question Answering." pith.science (2026). https://pith.science/paper/BXRD33ND

@misc{pith2026250516582,
  author       = {Pith},
  title        = {Pith review of: O$^2$-Searcher: A Searching-based Agent Model for Open-Domain Open-Ended Question Answering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BXRD33ND}},
  note         = {Machine review of arXiv:2505.16582}
}
abstract

Large Language Models (LLMs), despite their advancements, are fundamentally limited by their static parametric knowledge, hindering performance on tasks requiring open-domain up-to-date information. While enabling LLMs to interact with external knowledge environments is a promising solution, current efforts primarily address closed-end problems. Open-ended questions, which characterized by lacking a standard answer or providing non-unique and diverse answers, remain underexplored. To bridge this gap, we present O$^2$-Searcher, a novel search agent leveraging reinforcement learning to effectively tackle both open-ended and closed-ended questions in the open domain. O$^2$-Searcher leverages an efficient, locally simulated search environment for dynamic knowledge acquisition, effectively decoupling the external world knowledge from model's sophisticated reasoning processes. It employs a unified training mechanism with meticulously designed reward functions, enabling the agent to identify problem types and adapt different answer generation strategies. Furthermore, to evaluate performance on complex open-ended tasks, we construct O$^2$-QA, a high-quality benchmark featuring 300 manually curated, multi-domain open-ended questions with associated web page caches. Extensive experiments show that O$^2$-Searcher, using only a 3B model, significantly surpasses leading LLM agents on O$^2$-QA. It also achieves SOTA results on various closed-ended QA benchmarks against similarly-sized models, while performing on par with much larger ones.

Figures

Figures reproduced from arXiv: 2505.16582 by the authors.

Figure 1
Figure 1. Illustration of different characteristics of closed [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The construction of the knowledge cor￾pus for open-ended questions. All retrieved content is cached to construct the retriever index. For closed-ended queries, we follow [17] and locally index the 2018 Wikipedia dump [20] to facilitate efficient struc￾tured knowledge retrieval. By integrating these two types of content, we establish a special￾ized search environment E with a locally hosted knowledge corpus, enabling… view at source ↗
Figure 3
Figure 3. We use multi-round conversations for modeling action trajectories to enhance interactivity. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: GRPO training with the interaction with the search environment. The policy model is [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: The evolution of response length, reward value, and valid search results across different [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Search behavior across different datasets in the inference stage. † denotes gen￾eral datasets, ∗ denotes multi-hop datasets. Study of valid search across different datasets. We further analyze the search behavior across dif￾ferent question types during the testing stag…
Figure 7
Figure 7. Figure 7: A closed-ended case produced by O2 -Searcher Case Study [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: A closed-ended case produced by Search-R1 [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: An open-ended case produced by our O2 -Searcher. gather information and generate the final key findings for the open-ended question, demonstrating the generalization and adaptation ability to different types of questions. A.4 Limitations and Broader Impacts O 2 -Search…
Figure 10
Figure 10. Figure 10: Illustration of the collaborative report writing process involving O [PITH_FULL_IMAGE:figures/full_fig_p022_10.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

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

  1. Fetch-then-Explore: Decoupling Selection from Extraction over a Persistent Workspace for Search Agents

    cs.AI 2026-08 conditional novelty 6.0 of 10

    Fetch-then-Explore, which stores fetched pages in a per-question workspace and extracts evidence on demand with grep/read, beats visit-and-read and browsing baselines on BrowseComp across three LLM backbones.

  2. MemHarness: Memory Is Reconstructed, Not Replayed

    cs.AI 2026-07 conditional novelty 6.0 of 10

    End-to-end GRPO training teaches a unified LLM policy to reconstruct retrieved experiences into state-aligned guidance, outperforming verbatim memory replay on ALFWorld and WebShop.

  3. The Agent's First Day: Benchmarking Learning, Exploration, and Scheduling in the Workplace Scenarios

    cs.AI 2026-01 conditional novelty 6.0 of 10

    Trainee-Bench evaluates LLM agents on scheduling, active exploration, and continual learning in randomized workplace scenarios; the best current model completes only 35% of tasks.

  4. ParallelSearch: Train your LLMs to Decompose Query and Search Sub-queries in Parallel with Reinforcement Learning

    cs.CL 2025-08 unverdicted novelty 6.0 of 10

    ParallelSearch applies reinforcement learning with parallel-decomposition rewards to teach LLM search agents to issue independent sub-queries concurrently, lifting QA accuracy by 2.9% on average and 12.7% on paralleli...

Reference graph

Works this paper leans on

111 extracted references · 36 canonical work pages · cited by 4 Pith papers

  1. [1]

    Open deep search: Democratizing search with open-source reasoning agents.arXiv preprint arXiv:2503.20201, 2025

    Salaheddin Alzubi, Creston Brooks, Purva Chiniya, Edoardo Contente, Chiara von Gerlach, Lucas Irwin, Yihan Jiang, Arda Kaz, Windsor Nguyen, Sewoong Oh, et al. Open deep search: Democratizing search with open-source reasoning agents.arXiv preprint arXiv:2503.20201, 2025

    arXiv 2025

  2. [2]

    Self-rag: Learning to retrieve, generate, and critique through self-reflection

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self-rag: Learning to retrieve, generate, and critique through self-reflection. InThe Twelfth International Conference on Learning Representations, 2023

    2023

  3. [3]

    Scaling instruction-finetuned language models

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models. Journal of Machine Learning Research, 25(70):1–53, 2024

    2024

  4. [4]

    Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168, 2021

    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

    arXiv

  5. [5]

    Process reinforcement through implicit rewards.arXiv preprint arXiv:2502.01456, 2025

    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

    arXiv 2025

  6. [6]

    Plan-and-act: Improving planning of agents for long-horizon tasks

    Lutfi Eren Erdogan, Nicholas Lee, Sehoon Kim, Suhong Moon, Hiroki Furuta, Gopala Anumanchipalli, Kurt Keutzer, and Amir Gholami. Plan-and-act: Improving planning of agents for long-horizon tasks. arXiv preprint arXiv:2503.09572, 2025

    arXiv 2025

  7. [7]

    Trustgeogen: Scalable and formal-verified data engine for trustworthy multi-modal geometric problem solving.arXiv preprint arXiv:2504.15780, 2025

    Daocheng Fu, Zijun Chen, Renqiu Xia, Qi Liu, Yuan Feng, Hongbin Zhou, Renrui Zhang, Shiyang Feng, Peng Gao, Junchi Yan, et al. Trustgeogen: Scalable and formal-verified data engine for trustworthy multi-modal geometric problem solving.arXiv preprint arXiv:2504.15780, 2025

    arXiv 2025

  8. [8]

    Gemini deep research, 12 2024

    Google. Gemini deep research, 12 2024

    2024

Show all 111 references
  1. [9]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    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

    2025 arXiv

  2. [10]

    Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps.arXiv preprint arXiv:2011.01060, 2020

    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

    2011 arXiv

  3. [11]

    Metagpt: Meta programming for multi-agent collaborative framework.arXiv preprint arXiv:2308.00352, 3(4):6, 2023

    Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. Metagpt: Meta programming for multi-agent collaborative framework.arXiv preprint arXiv:2308.00352, 3(4):6, 2023

    2023 arXiv

  4. [12]

    Tree-planner: Efficient close-loop task planning with large language models.arXiv preprint arXiv:2310.08582, 2023

    Mengkang Hu, Yao Mu, Xinmiao Yu, Mingyu Ding, Shiguang Wu, Wenqi Shao, Qiguang Chen, Bin Wang, Yu Qiao, and Ping Luo. Tree-planner: Efficient close-loop task planning with large language models.arXiv preprint arXiv:2310.08582, 2023

    arXiv

  5. [13]

    A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.ACM Transactions on Information Systems, 43(2):1– 55, 2025

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions.ACM Transactions on Information...

    2025

  6. [14]

    Gpt-4o system card.arXiv preprint arXiv:2410.21276, 2024

    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. 10

    2024 arXiv

  7. [15]

    Forward-backward reasoning in large language models for mathematical verification.arXiv preprint arXiv:2308.07758, 2023

    Weisen Jiang, Han Shi, Longhui Yu, Zhengying Liu, Yu Zhang, Zhenguo Li, and James T Kwok. Forward-backward reasoning in large language models for mathematical verification.arXiv preprint arXiv:2308.07758, 2023

    2023 arXiv

  8. [16]

    Swe-bench: Can language models resolve real-world github issues?arXiv preprint arXiv:2310.06770, 2023

    Carlos E Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. Swe-bench: Can language models resolve real-world github issues?arXiv preprint arXiv:2310.06770, 2023

    2023 arXiv

  9. [17]

    Search-r1: Training llms to reason and leverage search engines with reinforcement learning.arXiv preprint arXiv:2503.09516, 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.arXiv preprint arXiv:2503.09516, 2025

    2025 arXiv

  10. [18]

    Triviaqa: A large scale distantly supervised challenge dataset for reading comprehension.arXiv preprint arXiv:1705.03551, 2017

    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

    2017 arXiv

  11. [19]

    Reinforcement learning: A survey

    Leslie Pack Kaelbling, Michael L Littman, and Andrew W Moore. Reinforcement learning: A survey. Journal of artificial intelligence research, 4:237–285, 1996

    1996

  12. [20]

    Dense passage retrieval for open-domain question answering

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick SH Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. InEMNLP (1), pages 6769–6781, 2020

    2020

  13. [21]

    Reliability across parametric and external knowledge: Understanding knowledge handling in llms.arXiv preprint arXiv:2502.13648, 2025

    Youna Kim, Minjoon Choi, Sungmin Cho, Hyuhng Joon Kim, Sang-goo Lee, and Taeuk Kim. Reliability across parametric and external knowledge: Understanding knowledge handling in llms.arXiv preprint arXiv:2502.13648, 2025

    2025 arXiv

  14. [22]

    The hungarian method for the assignment problem.Naval research logistics quarterly, 2(1-2):83–97, 1955

    Harold W Kuhn. The hungarian method for the assignment problem.Naval research logistics quarterly, 2(1-2):83–97, 1955

    1955

  15. [23]

    Natural questions: a benchmark for question answering research.Transactions of the Association for Computational Linguistics, 7:453–466, 2019

    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 Computat...

    2019

  16. [24]

    Retrieval-augmented generation for knowledge- intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474, 2020

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge- intensive nlp tasks.Advances in neural information processing systems...

    2020

  17. [25]

    Camel: Commu- nicative agents for" mind" exploration of large language model society.Advances in Neural Information Processing Systems, 36:51991–52008, 2023

    Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Commu- nicative agents for" mind" exploration of large language model society.Advances in Neural Information Processing Systems, 36:51991–52008, 2023

    2023

  18. [26]

    Codei/o: Condensing reasoning patterns via code input-output prediction.arXiv preprint arXiv:2502.07316, 2025

    Junlong Li, Daya Guo, Dejian Yang, Runxin Xu, Yu Wu, and Junxian He. Codei/o: Condensing reasoning patterns via code input-output prediction.arXiv preprint arXiv:2502.07316, 2025

    2025 arXiv

  19. [27]

    Search-o1: Agentic search-enhanced large reasoning models.arXiv preprint arXiv:2501.05366, 2025

    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

    2025 arXiv

  20. [28]

    Openmanus: An open-source framework for building general ai agents.https://github.com/mannaandpoem/OpenManus, 2025

    Xinbin Liang, Jinyu Xiang, Zhaoyang Yu, Jiayi Zhang, and Sirui Hong. Openmanus: An open-source framework for building general ai agents.https://github.com/mannaandpoem/OpenManus, 2025

    2025

  21. [29]

    Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

    2024 arXiv

  22. [30]

    Inference- time scaling for generalist reward modeling.arXiv preprint arXiv:2504.02495, 2025

    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

    2025

  23. [31]

    Large language model agent: A survey on methodology, applications and challenges.arXiv preprint arXiv:2503.21460, 2025

    Junyu Luo, Weizhi Zhang, Ye Yuan, Yusheng Zhao, Junwei Yang, Yiyang Gu, Bohan Wu, Binqi Chen, Ziyue Qiao, Qingqing Long, et al. Large language model agent: A survey on methodology, applications and challenges.arXiv preprint arXiv:2503.21460, 2025

    2025 arXiv

  24. [32]

    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

    2022 arXiv

  25. [33]

    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. InThe Twelfth International Conference on Learning Representations, 2023. 11

    2023

  26. [34]

    Deep research system card

    OpenAI. Deep research system card. Technical report, OpenAI, 2 2025

    2025

  27. [35]

    Memgpt: Towards llms as operating systems

    Charles Packer, Vivian Fang, Shishir_G Patil, Kevin Lin, Sarah Wooders, and Joseph_E Gonzalez. Memgpt: Towards llms as operating systems. 2023

    2023

  28. [36]

    Advancing reasoning in large language models: Promising methods and approaches.arXiv preprint arXiv:2502.03671, 2025

    Avinash Patil. Advancing reasoning in large language models: Promising methods and approaches.arXiv preprint arXiv:2502.03671, 2025

    2025 arXiv

  29. [37]

    Check your facts and try again: Improving large language models with external knowledge and automated feedback.arXiv preprint arXiv:2302.12813, 2023

    Baolin Peng, Michel Galley, Pengcheng He, Hao Cheng, Yujia Xie, Yu Hu, Qiuyuan Huang, Lars Liden, Zhou Yu, Weizhu Chen, et al. Check your facts and try again: Improving large language models with external knowledge and automated feedback.arXiv preprint arXiv:2302.12813, 2023

    2023 arXiv

  30. [38]

    Measuring and narrowing the compositionality gap in language models.arXiv preprint arXiv:2210.03350, 2022

    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

    2022 arXiv

  31. [39]

    Making language models better tool learners with execution feedback.arXiv preprint arXiv:2305.13068, 2023

    Shuofei Qiao, Honghao Gui, Chengfei Lv, Qianghuai Jia, Huajun Chen, and Ningyu Zhang. Making language models better tool learners with execution feedback.arXiv preprint arXiv:2305.13068, 2023

    2023 arXiv

  32. [40]

    Toolllm: Facilitating large language models to master 16000+ real-world apis.arXiv preprint arXiv:2307.16789, 2023

    Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al. Toolllm: Facilitating large language models to master 16000+ real-world apis.arXiv preprint arXiv:2307.16789, 2023

    2023 arXiv

  33. [41]

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

    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

    2024 arXiv

  34. [42]

    R1-searcher: Incentivizing the search capability in llms via reinforcement learning.arXiv preprint arXiv:2503.05592, 2025

    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

    2025 arXiv

  35. [43]

    Reinforcement learning.Journal of Cognitive Neuroscience, 11(1):126–134, 1999

    Richard S Sutton, Andrew G Barto, et al. Reinforcement learning.Journal of Cognitive Neuroscience, 11(1):126–134, 1999

    1999

  36. [44]

    Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions.arXiv preprint arXiv:2212.10509, 2022

    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

    2022 arXiv

  37. [45]

    Musique: Multihop questions via single-hop question composition.Transactions of the Association for Computational Linguistics, 10:539–554, 2022

    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

    2022

  38. [46]

    Karthik Valmeekam, Matthew Marquez, Alberto Olmo, Sarath Sreedharan, and Subbarao Kambhampati. Planbench: An extensible benchmark for evaluating large language models on planning and reasoning about change.Advances in Neural Information Processing Systems, 36:38975–38987, 2023

    2023

  39. [47]

    Plan- and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models.arXiv preprint arXiv:2305.04091, 2023

    Lei Wang, Wanyu Xu, Yihuai Lan, Zhiqiang Hu, Yunshi Lan, Roy Ka-Wei Lee, and Ee-Peng Lim. Plan- and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models.arXiv preprint arXiv:2305.04091, 2023

    2023 arXiv

  40. [48]

    Text embeddings by weakly-supervised contrastive pre-training.arXiv preprint arXiv:2212.03533, 2022

    Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Ma- jumder, and Furu Wei. Text embeddings by weakly-supervised contrastive pre-training.arXiv preprint arXiv:2212.03533, 2022

    2022 arXiv

  41. [49]

    Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers.Advances in neural information processing systems, 33:5776–5788, 2020

    Wenhui Wang, Furu Wei, Li Dong, Hangbo Bao, Nan Yang, and Ming Zhou. Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers.Advances in neural information processing systems, 33:5776–5788, 2020

    2020

  42. [50]

    Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171, 2022

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models.arXiv preprint arXiv:2203.11171, 2022

    2022 arXiv

  43. [51]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    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

    2022

  44. [52]

    Avatar: Optimizing llm agents for tool usage via contrastive reasoning.Advances in Neural Information Processing Systems, 37:25981–26010, 2024

    Shirley Wu, Shiyu Zhao, Qian Huang, Kexin Huang, Michihiro Yasunaga, Kaidi Cao, Vassilis Ioannidis, Karthik Subbian, Jure Leskovec, and James Y Zou. Avatar: Optimizing llm agents for tool usage via contrastive reasoning.Advances in Neural Information Processing Systems, 37:259...

    2024

  45. [53]

    Omnithink: Expanding knowledge boundaries in machine writing through thinking.arXiv preprint arXiv:2501.09751, 2025

    Zekun Xi, Wenbiao Yin, Jizhan Fang, Jialong Wu, Runnan Fang, Ningyu Zhang, Jiang Yong, Pengjun Xie, Fei Huang, and Huajun Chen. Omnithink: Expanding knowledge boundaries in machine writing through thinking.arXiv preprint arXiv:2501.09751, 2025

    2025

  46. [54]

    Deepseek-prover: Advancing theorem proving in llms through large-scale synthetic data

    Huajian Xin, Daya Guo, Zhihong Shao, Zhizhou Ren, Qihao Zhu, Bo Liu, Chong Ruan, Wenda Li, and Xiaodan Liang. Deepseek-prover: Advancing theorem proving in llms through large-scale synthetic data. arXiv preprint arXiv:2405.14333, 2024

    2024 arXiv

  47. [55]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report.arXiv preprint arXiv:2412.15115, 2024

    2024 arXiv

  48. [56]

    Gpt4tools: Teaching large language model to use tools via self-instruction.Advances in Neural Information Processing Systems, 36:71995–72007, 2023

    Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. Gpt4tools: Teaching large language model to use tools via self-instruction.Advances in Neural Information Processing Systems, 36:71995–72007, 2023

    2023

  49. [57]

    Hotpotqa: A dataset for diverse, explainable multi-hop question answering.arXiv preprint arXiv:1809.09600, 2018

    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

    2018 arXiv

  50. [58]

    Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems, 36:11809–11822, 2023

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.Advances in neural information processing systems, 36:11809–11822, 2023

    2023

  51. [59]

    Re- act: Synergizing reasoning and acting in language models

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

    2023

  52. [60]

    Cognitive mirage: A review of hallucinations in large language models.arXiv preprint arXiv:2309.06794, 2023

    Hongbin Ye, Tong Liu, Aijia Zhang, Wei Hua, and Weiqiang Jia. Cognitive mirage: A review of hallucinations in large language models.arXiv preprint arXiv:2309.06794, 2023

    2023 arXiv

  53. [61]

    Physics of language models: Part 2.1, grade- school math and the hidden reasoning process

    Tian Ye, Zicheng Xu, Yuanzhi Li, and Zeyuan Allen-Zhu. Physics of language models: Part 2.1, grade- school math and the hidden reasoning process. InThe Thirteenth International Conference on Learning Representations, 2024

    2024

  54. [62]

    Dapo: An open-source llm reinforcement learning system at scale.arXiv preprint arXiv:2503.14476, 2025

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

    2025 arXiv

  55. [63]

    Rankrag: Unifying context ranking with retrieval-augmented generation in llms.Advances in Neural Information Processing Systems, 37:121156–121184, 2024

    Yue Yu, Wei Ping, Zihan Liu, Boxin Wang, Jiaxuan You, Chao Zhang, Mohammad Shoeybi, and Bryan Catanzaro. Rankrag: Unifying context ranking with retrieval-augmented generation in llms.Advances in Neural Information Processing Systems, 37:121156–121184, 2024

    2024

  56. [64]

    Easytool: Enhancing llm-based agents with concise tool instruction.arXiv preprint arXiv:2401.06201, 2024

    Siyu Yuan, Kaitao Song, Jiangjie Chen, Xu Tan, Yongliang Shen, Ren Kan, Dongsheng Li, and De- qing Yang. Easytool: Enhancing llm-based agents with concise tool instruction.arXiv preprint arXiv:2401.06201, 2024

    2024 arXiv

  57. [65]

    Inference scaling for long-context retrieval augmented generation

    Zhenrui Yue, Honglei Zhuang, Aijun Bai, Kai Hui, Rolf Jagerman, Hansi Zeng, Zhen Qin, Dong Wang, Xuanhui Wang, and Michael Bendersky. Inference scaling for long-context retrieval augmented generation. arXiv preprint arXiv:2410.04343, 2024

    arXiv

  58. [66]

    Rest-mcts*: Llm self-training via process reward guided tree search, 2024a.URL https://arxiv

    Dan Zhang, Sining Zhoubian, Ziniu Hu, Yisong Yue, Yuxiao Dong, and Jie Tang. Rest-mcts*: Llm self-training via process reward guided tree search, 2024a.URL https://arxiv. org/abs/2406.03816

    arXiv

  59. [67]

    Siren’s song in the ai ocean: a survey on hallucination in large language models.arXiv preprint arXiv:2309.01219, 2023

    Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, et al. Siren’s song in the ai ocean: a survey on hallucination in large language models.arXiv preprint arXiv:2309.01219, 2023

    2023 arXiv

  60. [68]

    Towards lifelong learning of large language models: A survey.ACM Computing Surveys, 57(8):1–35, 2025

    Junhao Zheng, Shengjie Qiu, Chengming Shi, and Qianli Ma. Towards lifelong learning of large language models: A survey.ACM Computing Surveys, 57(8):1–35, 2025

    2025

  61. [69]

    Deepresearcher: Scaling deep research via reinforcement learning in real-world environments.arXiv preprint arXiv:2504.03160, 2025

    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

    2025 arXiv

  62. [70]

    Least-to-most prompting enables complex reasoning in large language models.arXiv preprint arXiv:2205.10625, 2022

    Denny Zhou, Nathanael Schärli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. Least-to-most prompting enables complex reasoning in large language models.arXiv preprint arXiv:2205.10625, 2022

    2022 arXiv

  63. [71]

    IND" and

    Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, et al. Webarena: A realistic web environment for building autonomous agents. arXiv preprint arXiv:2307.13854, 2023. 13 A Appendices A.1 Related Works...

    2023 arXiv

  64. [72]

    Conduct **multi-round SEARCH actions**, refining your queries over multiple iterations to gather comprehensive, high-quality information

  65. [73]

    Limit the number of SEARCH rounds to a maximum of **5**

  66. [74]

    Craft highly **relevant and non-redundant findings** that focus on delivering maximum value and insights to the user

  67. [75]

    Key finding 1: Green energy policies have led to a 15% increase in employment across renewable energy sectors in 2023

    Only proceed to the **final key findings presentation** once the SEARCH actions yield sufficient information to address the user’s request comprehensively. ### Examples ### SEARCH Action Execution Example: <think>Begin with an initial search to collect broad background informa...

    2023

  68. [76]

    Begin the task with an **analysis phase** to determine overarching themes and potential directions for inquiry

  69. [77]

    Execute **multi-round SEARCH actions** to gather comprehensive data

  70. [78]

    Review, analyze, and synthesize the obtained information into **key findings**

  71. [79]

    Present the **key findings** in a flat JSON array structure as per the Final Key Findings Presentation format. 19 When the LLM agent generates incorrect output (e.g., formatting errors), we use the following user prompt to guide correction: Error prompt for LLM agents The acti...

  72. [80]

    Preserve critical information, logical flow, and essential data points

  73. [81]

    Prioritize content relevance to the research query: <query>QUERY</query>

  74. [82]

    **Adjust length dynamically**: - If original content < 2K tokens, maintain original token count ±10% - If original content > 2K tokens, compress to∼2ktokens

  75. [83]

    Return a maximum of three distinct learnings

    Format output in clean Markdown without decorative elements **Prohibited**: - Adding content beyond source material - Truncating mid-sentence to meet token limits To extract the query-relevant learnings, we use the following prompt template: 20 Prompt template for learning ext...

  76. [84]

    You will receive two sets of findings: - Input findings: each separated by a newline - Target findings: each separated by a newline

  77. [85]

    For each input finding, you need to: - Analyze if it is semantically similar to any of the target findings - If a similar entry is found, pair them together - Each input finding can only be paired with one target finding - Each target finding can only be paired with one input finding

  78. [86]

    input finding 1

    Output Requirements: You need to output a list in JSON format, where each element is a pair: [ ["input finding 1", "matched target finding 1"], ["input finding 2", "matched target finding 2"], ... ] Similarity Judgment Criteria:

  79. [87]

    Core meanings should be identical or very close

  80. [88]

    Even if expressions differ, pair them if core concepts match

  81. [89]

    Partial overlap is not enough; main points must match

  82. [90]

    Do not include any explanatory text outside the JSON array

    If a finding contains multiple points, at least the main points must match Please ensure your output follows the strict JSON format for subsequent processing. Do not include any explanatory text outside the JSON array. If no matches are found, output an empty array []. Input F...

  83. [91]

    Output JSON array only, no additional explanatory text

  84. [92]

    Use the complete original text for each finding

  85. [93]

    O 2-Searcher is equipped with search capabilities to effectively gather and sift through pertinent online information, distilling key findings

    If no matches are found, output an empty array [] 21 A.6 Application of Writing Reports The task of report writing represents a quintessential open-domain, open-ended problem, where crafting a superior report demands comprehensive referencing and robust empirical support for i...

  86. [94]

    First, carefully analyze the user’s query to identify: - The core question or problem the user is trying to solve - Any specific aspects or dimensions they’re particularly interested in - The likely purpose of their research (practical application, theoretical understanding, etc.)

  87. [95]

    Create a logical outline structure that: - Starts with foundational concepts the user needs to understand - Progresses through increasingly specific or complex aspects of the topic - Concludes with practical applications or future implications relevant to the user’s query

  88. [96]

    Here are distilled key findings relevent to user’s query:<contents>CONTENTS</contents>

    Throughout the outline: - Maintain consistent terminology aligned with the user’s query - Ensure each section directly contributes to answering the user’s core question - Avoid tangential information even if interesting but not directly relevant The final outline should read a...

  89. [97]

    Introduction

    Represent the outline as a dictionary where: - Keys are chapters titles (e.g., "Introduction", "Analysis", "Conclusion"). - Values are lists of subsection titles or key points for each section

  90. [98]

    Create EXACTLY 5-7 first-level chapters (no more, no less)

  91. [99]

    For each chapter, include 1-4 second-level sections that explore specific aspects (no more, no less)

  92. [100]

    Outline must Include only up to the SECOND level of the title, please DONOT write the third level of the title or bullet points

  93. [101]

    ALL outline items should be writen in Chinese

  94. [102]

    Introduction

    Do NOT include any explanations, notes. Don’t use markdown code block label, just represent the dictionary in JSON format. Follow this operational protocol: Example: { "Introduction": { %ATTN: each chapter should be a dict not a list, "Background": "a concise description of th...

  95. [103]

    Analysis: Break down the information into meaningful insights

  96. [104]

    Synthesis: Connect ideas from different sources to form a cohesive narrative

  97. [105]

    Explanations: Provide clear and concise explanations of the content

  98. [106]

    Add in-text references (e.g., [1], [2]) based on the content index to reference the URLs where the information was sourced

  99. [107]

    Generate the contents using the Markdown format

    At the end of the section, list all the URLs referenced in the content. Generate the contents using the Markdown format. 23 Prompt template for final report generation Given the following initial query <query>QUERY</query> and contents <contents>CONTENTS</contents> of each sec...

  100. [108]

    Try to retain the original content to maintain the integrity and accuracy of the information

    Merge all sections into a single, cohesive report. Try to retain the original content to maintain the integrity and accuracy of the information. Ensure the report is well-structured and flows logically from one section to the next

  101. [109]

    Add in-text references (e.g., [1], [2]) to reference the URLs where the information was sourced

  102. [110]

    Provide a numbered list of the URLs corresponding to the in-text citations

    Include a References Section at the end of the report. Provide a numbered list of the URLs corresponding to the in-text citations

  103. [111]

    Why can’t Southeast Asia form a deep economic union like the EU?

    Generate the final report in Chinese, using Markdown format. Follow these formatting guidelines: - Use clear headings for main sections and subsections. - Ensure the content is thorough and comprehensive, organized with paragraphs for each subsection instead of listing items. ...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.