Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

OpenReward: Learning to Reward Long-form Agentic Tasks via Reinforcement Learning

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A reward model that searches external sources before judging can outperform static LLM judges on knowledge-intensive long-form responses.

desk verdict A genuinely useful recipe for training evidence-grounded judges, but the R_tool term is underspecified and the in-domain numbers are built on the same synthetic premise as the training data. read the letter →

arxiv 2510.24636 v3 pith:P2Z5KBPG submitted 2025-10-28 cs.CL

Ziyou Hu , Zhengliang Shi , Minghang Zhu , Haitao Li , Teng Sun , Pengjie Ren , Suzan Verberne , Zhaochun Ren This is my paper
classification cs.CL
keywords rewardmodelingtool-augmentedevaluationlong-formtextreinforcementlearningLLM-as-judgepreferencedatasynthesisRLHF
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

OpenReward (OpenRM) is a generative reward model trained to invoke external search tools before deciding which of two long-form answers is better. The paper argues that existing reward models fail on knowledge-intensive, open-ended responses because they rely on internal knowledge, and shows that a small model taught to gather evidence can surpass much larger static judges. Training pairs are synthesized without human labels: a strong LLM answers a document-grounded query with and without access to the reference document, and the grounded answer is treated as the better one. On three new in-domain benchmarks the model reports 91.33% average judgment accuracy versus 63.23% for the best tool-augmented baseline, and it generalizes to out-of-domain benchmarks while using only 27K training examples. The authors also demonstrate gains in inference-time response selection and in data selection for direct preference optimization.

What carries the argument

The central mechanism is the agentic judgment trajectory: the reward model interleaves reasoning tokens with tool calls (search over reference corpora) and uses the returned evidence to verify claims before choosing between answers. The composite reward R = R_EM + sign(R_EM)·λ·R_tool supervises both the terminal judgment and each tool selection, and the synthetic preference pairs from document-grounded versus ungrounded generation provide the learning signal. GRPO turns the sparse final accuracy into dense group-relative advantages, so the model learns when and how to search.

What would settle it

Construct a test set of preference pairs in which the ungrounded response is factually accurate and more concise, while the grounded response is verbose and padded with irrelevant retrieved details, and label the ungrounded response as better. If OpenRM consistently prefers the grounded-style response, it has learned to detect document-grounding rather than genuine quality. A second probe: evaluate OpenRM on pairs where both responses were generated with document access, so no style cue is available; chance-level accuracy would indicate the model relies on the contrast rather than content.

Watch

Extended reading notes

Core claim

OpenRM is a tool-augmented long-form reward model. Given a query and two candidate responses, it iteratively decides which tool to call, retrieves external evidence, reasons over it, and then outputs a preference judgment. The training objective combines an exact-match reward on the final judgment with a dense reward on intermediate tool selection, optimized with Group Relative Policy Optimization; tool-use credit is granted only when the final judgment is correct, which prevents reward hacking. The preference data is synthesized on a large scale through a controllable pipeline that samples reference documents, generates target-aware queries from them, and produces a positive response with d

Load-bearing premise

The load-bearing premise is that a response written with access to a reference document is always better than one written without it; if an ungrounded response is equally accurate or the grounded response adds irrelevant detail, the training signal and the new benchmarks inherit that distortion.

Editorial extensions

If this is right

  • If OpenRM's results hold, reward models for long-form and knowledge-intensive evaluation should be equipped with external tool access rather than relying only on parametric knowledge.
  • The synthetic data pipeline removes the need for human pairwise annotations, making it feasible to scale preference data for new domains where reference corpora exist.
  • The composite reward design shows that supervising intermediate tool use jointly with final accuracy avoids both lazy searching and over-searching reward hacking.
  • Downstream alignment (DPO) improves when preference pairs are chosen by OpenRM, suggesting tool-augmented reward signals can benefit language model training, not just evaluation.
  • A relatively small (7B-parameter) model trained this way can outperform much larger static LLM judges on the tested long-form tasks.

Reading between the lines

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

  • The synthetic preference label assumes document-grounded responses are always higher quality; if that assumption fails, the model may learn to detect document-grounded style rather than genuine quality. An editor's test: run OpenRM on pairs where the ungrounded response is factually superior to see if accuracy collapses.
  • The approach's generality depends on retrieval tool reliability; in domains without a clean reference corpus (e.g., subjective or creative writing), the evidence-gathering step may add noise rather than signal.
  • Because the reward signal is exact match on a synthetic label, the same RL recipe could be adapted to other verifiable outcome functions (e.g., matching a known fact in the response), potentially extending OpenRM beyond pairwise comparison.
  • The reported in-domain gains might be partially inflated by the construction of the benchmarks from the same documents used to synthesize training data; true out-of-domain generalization should be probed on tasks with different evidence sources.
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 OpenRM, a tool-augmented generative reward model trained with GRPO on 27K synthesized pairwise examples. Given a query and two candidate answers, OpenRM interleaves reasoning with calls to Wikipedia/arXiv search, then outputs a preference. The training reward in Eq. (3) combines a terminal EM reward with an intermediate tool-use reward. The authors synthesize training pairs by prompting an LLM to answer with and without a reference document; the same mechanism is used to construct three new in-domain test sets. Experiments report large in-domain gains (Table 2: 91.33% vs. 63.23% for the best tool-augmented baseline), competitive out-of-domain results on PandaLM/RewardBench (Table 3), and downstream gains in data selection and DPO alignment.

Significance. The idea of making a reward model actively retrieve evidence and training the whole policy with RL is timely and potentially useful, and the paper has several explicit strengths: a concrete training pipeline, a stated code release, human evaluation of judgment quality (§5.3), and a data-efficiency comparison. If the central claims were valid, OpenRM would be an interesting step toward long-form evaluators that are not limited by parametric knowledge. However, the empirical foundation is currently not reliable: the in-domain benchmarks are generated by the same synthetic preference rule used for training, so the headline accuracy numbers do not measure what they claim. The tool-use reward in Eq. (3) is never formally defined, making the method irreproducible as written. These issues are load-bearing for the paper's main claims, so the current manuscript cannot be accepted.

major comments (4)
  1. [§3.3, §4, §5.1 (Table 2)] The in-domain benchmarks are not independent of the training signal. Training pairs are built by making the positive response from the query+reference document and the negative response from the query alone (§3.3), and the three new test sets are 'consistent with the training data source' (§4). The binary label in both cases is therefore the presence/absence of document grounding during generation. A model can score 91.33% by detecting the document-grounded style (e.g., concrete entity names and detailed facts) rather than by judging answer quality. The reported 91.33% vs. 63.23% in Table 2 is inflated by construction and cannot support the paper's central claim. The human evaluation in §5.3 rates the self-containment and factuality of the judgments but does not validate the ground-truth preference labels, so it does not break this circularity.
  2. [§3.2, Eq. (3)] The intermediate reward R_tool is never formally defined. The text says it assigns +1 to 'correct tool selections' and 0 to incorrect ones, but the paper does not specify what constitutes a correct tool selection, how ground-truth tool-selection labels are derived for the synthetic data, or how R_tool is computed from a trajectory. Since Eq. (3) and the GRPO advantage in Eq. (4) depend on this term, the training objective cannot be implemented or reproduced from the manuscript. The notation is also inconsistent: Eq. (3) uses sign(R_EM), while §5.2 uses '1_REM'; and the free weighting parameter λ is not given a value in §4.
  3. [§3.3, §5.3] The core synthetic-label assumption is untested. The paper states that a response generated with the reference document is 'grounded, informative, and factually accurate,' while the no-document response 'typically leads to incomplete, hallucinated, or less reliable content.' The word 'typically' admits exceptions, yet the data pipeline assigns binary labels. No human agreement, quality metric, or diagnostic is reported to confirm that the with-document response is reliably the better answer. If the no-document response is often also accurate, or if the with-document response merely adds length or detail without improving correctness, the model learns a surface cue rather than genuine quality. A concrete test, such as human preference agreement on a sample of the 27K pairs, is needed before the training signal can be accepted.
  4. [§5.1, Tables 2 and 3] The reported differences lack uncertainty estimates or significance testing. Table 2 reports single accuracy numbers per condition, and Table 3 shows OpenRM beating RRM-7B by only 0.40 points on average (78.54 vs. 78.14), with RewardBench actually +0.88 and PandaLM -1.68; no confidence intervals are given. The paper only runs a t-test in §5.2 between two ablation variants, not for the headline comparisons. Given the small margins and the benchmark-construction issues above, the claim of 'substantially outperforms' is not supported empirically.
minor comments (6)
  1. [§1] The contribution list jumps from (iii) to (v); item (iv) is missing.
  2. [§5.2] The text refers to 'Eq. 3.1' and '1_REM' where Eq. (3) and the indicator function are meant; please unify the notation.
  3. [Table 5] The benchmark name is misspelled as 'TrutuFulQA'; should be TruthfulQA.
  4. [Throughout] The name is inconsistent between 'OPENRM' and 'OpenReward' (title vs. §3.1); choose one.
  5. [Table 4] The header row repeats 'Accuracy (↑)' for every column, which obscures the three-point-scale nature of the human ratings.
  6. [Abstract] The code URL appears as 'on/gtbOpenRM'; please provide a working link or remove it.

Circularity Check

1 steps flagged · score 5.0 of 10

In-domain benchmarks share the exact synthetic preference rule used for training, so the headline in-domain gain is partly self-referential; independent OOD and human evaluations keep the paper from being fully circular.

  1. self definitional [Sections 3.2-3.3, Section 4 (Benchmark and Evaluation Protocol), Table 2]
    "Second, R EM measures the correctness of the final prediction by checking whether the model's output a i exactly matches the ground-truth label, i.e., EM(ai). ... The positive response is generated using both the query and its corresponding reference document, guaranteeing that the answer is grounded, informative, and factually accurate. In contrast, the negative response is generated using the query alone, without access to the reference document ... These datasets are consistent with the training data source and are used to assess the model's in-domain performance."

    The label 'better response' is defined identically in training and test: the response generated with the reference document is positive, and the one generated without it is negative. R_EM supervises by exact match to this ground-truth label (Eq. 3), and the three new in-domain benchmarks are generated by the same synthesis pipeline, explicitly 'consistent with the training data source.' Hence Table 2's 91.33% in-domain accuracy primarily measures whether the model has learned to reproduce the synthetic document-access rule (e.g., by recognizing document-grounded style), not independently validated long-form quality discrimination. The paper itself hedges that the negative response is only 'typically' worse, so the rule is an untested premise rather than a proven quality signal.

full rationale

The derivation chain is self-contained in its equations: GRPO with Eqs. 3-5 is applied to the synthesized pairs, and the out-of-domain results in Table 3 and human evaluation in Table 4 provide independent evidence. No load-bearing self-citation or imported uniqueness theorem is present. The circularity is partial and concentrated on the in-domain claim: the positive/negative definition used to train the reward model is reused to construct the three new evaluation datasets, and the final reward R_EM is exact match to that same definition. This makes the headline in-domain comparison (91.33 vs 63.23) a within-pipeline consistency check rather than an independent demonstration of quality judgment. The independent benchmarks (PandaLM/RewardBench, UltraFeedback, TruthfulQA/MMLU-Pro/TriviaQA) prevent a higher score, but they do not rescue the claim that the three newly collected datasets show 'substantially outperforms existing reward modeling approaches' in long-form quality. Separately (not scored as circularity), R_tool in Eq. 3 is claimed to give +1/0 for correct/incorrect tool selection but no ground-truth tool-selection labels are defined, so that component is under-specified rather than reducible to an input.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the synthetic labeling assumption (doc-grounded answer is better), the existence of a tool-use oracle that the paper does not define, and the reliability of retrieval. These are not standard axioms; they are load-bearing assumptions specific to this method.

free parameters (1)
  • λ (tool reward weight) = 0.5 (default; 1.0 in one ablation)
    Equation (3) balances the tool-selection reward with the exact-match reward through λ. The paper never states the default in Sec. 4; the ablation in Sec. 5.2 compares λ=0.5 and λ=1.0, implying 0.5 is the default. The learned behavior depends on this hand-chosen value.
assumptions (4)
  • domain assumption The positive response generated with the reference document is of higher quality than the negative response generated without it.
    This defines all synthetic preference labels in Sec. 3.3 and the ground truth for the three in-domain benchmarks in Sec. 5.1. No human validation of this premise is provided beyond a 30-case evaluation of the trained model's judgments.
  • ad hoc to paper Correct tool selections can be automatically identified for a tool-use reward.
    Eq. (3) requires R_tool, and Sec. 5.2 says correct tool selections get +1, but no verifier, rule, or computed criterion is given anywhere in the paper, making this an ad-hoc unstated assumption.
  • domain assumption The external tools (Wikipedia 2018 dump, arXiv via LitSearch) return evidence relevant enough to support correct judgments.
    The method's entire value proposition is that retrieved evidence improves judgments. Imperfect retrieval is acknowledged as a limitation (Sec. 7), but the approach still depends on tool reliability.
  • standard math GRPO provides correct credit assignment for the composite reward.
    Uses GRPO objective from Shao et al. (2024) (Eqs. 4-5) without derivation; a standard, trusted background result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OpenReward: Learning to Reward Long-form Agentic Tasks via Reinforcement Learning." pith.science (2026). https://pith.science/paper/P2Z5KBPG

@misc{pith2026251024636,
  author       = {Pith},
  title        = {Pith review of: OpenReward: Learning to Reward Long-form Agentic Tasks via Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P2Z5KBPG}},
  note         = {Machine review of arXiv:2510.24636}
}
read the original abstract

Reward models (RMs) have become essential for aligning large language models (LLMs), serving as scalable proxies for human evaluation in both training and inference. However, existing RMs struggle on knowledge-intensive and long-form tasks, where evaluating correctness requires grounding beyond the model's internal knowledge. This limitation hinders them from reliably discriminating subtle quality differences, especially when external evidence is necessary. To address this, we introduce OpenRM, a tool-augmented long-form reward model that systematically judges open-ended responses by invoking external tools to gather relevant evidence. We train OpenRM with Group Relative Policy Optimization (GRPO) on over 27K synthesized pairwise examples generated through a controllable data synthesis framework. The training objective jointly supervises intermediate tool usage and final outcome accuracy, incentivizing our reward model to learn effective evidence-based judgment strategies. Extensive experiments on three newly-collected datasets and two widely-used benchmarks demonstrate that OpenRM substantially outperforms existing reward modeling approaches. As a further step, we integrate OpenRM into both inference-time response selection and training-time data selection. This yields consistent gains in downstream LLM alignment tasks, highlighting the potential of tool-augmented reward models for scaling reliable long-form evaluation.

Figures

Figures reproduced from arXiv: 2510.24636 by the authors.

Figure 1
Figure 1. Illustration of the OPENRM framework where the reward model, when receiving the candidate responses, progressively invokes external tools to gather useful evidence, and then make the final judgment. comparisons or natural language evaluations (Li et al., 2024a; Zheng et al., 2023; Gu et al., 2024). Compared to traditional scalar reward models, generative reward models offer significant advantages in expressiveness a… view at source ↗
Figure 2
Figure 2. Illustration of the overlap between RewardBench and the training datasets of different [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Training process of the models trained with different variants of the vanilla training su [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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.

Reference graph

Works this paper leans on

19 extracted references · 16 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Lit- search: A retrieval benchmark for scientific literature search.arXiv preprint arXiv:2407.18940,

    Anirudh Ajith, Mengzhou Xia, Alexis Chevalier, Tanya Goyal, Danqi Chen, and Tianyu Gao. Lit- search: A retrieval benchmark for scientific literature search.arXiv preprint arXiv:2407.18940,

    arXiv

  2. [5]

    Ultrafeedback: Boosting language models with scaled ai feedback.arXiv preprint arXiv:2310.01377,

    Ganqu Cui, Lifan Yuan, Ning Ding, Guanming Yao, Bingxiang He, Wei Zhu, Yuan Ni, Guotong Xie, Ruobing Xie, Yankai Lin, et al. Ultrafeedback: Boosting language models with scaled ai feedback.arXiv preprint arXiv:2310.01377,

    arXiv

  3. [6]

    A survey on llm-as-a-judge.arXiv preprint arXiv:2411.15594,

    Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Ying- han Shen, Shengjie Ma, Honghao Liu, et al. A survey on llm-as-a-judge.arXiv preprint arXiv:2411.15594,

    arXiv

  4. [7]

    Reward reasoning model.arXiv preprint arXiv:2505.14674,

    Jiaxin Guo, Zewen Chi, Li Dong, Qingxiu Dong, Xun Wu, Shaohan Huang, and Furu Wei. Reward reasoning model.arXiv preprint arXiv:2505.14674,

    arXiv

  5. [8]

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

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

    arXiv

  6. [9]

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

    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,

    arXiv

  7. [10]

    Rewardbench: Evaluating reward models for language modeling.arXiv preprint arXiv:2403.13787,

    Nathan Lambert, Valentina Pyatkin, Jacob Morrison, LJ Miranda, Bill Yuchen Lin, Khyathi Chandu, Nouha Dziri, Sachin Kumar, Tom Zick, Yejin Choi, et al. Rewardbench: Evaluating reward models for language modeling.arXiv preprint arXiv:2403.13787,

    arXiv

  8. [11]

    Llms-as-judges: a comprehensive survey on llm-based evaluation methods.arXiv preprint arXiv:2412.05579, 2024a

    Haitao Li, Qian Dong, Junjie Chen, Huixue Su, Yujia Zhou, Qingyao Ai, Ziyi Ye, and Yiqun Liu. Llms-as-judges: a comprehensive survey on llm-based evaluation methods.arXiv preprint arXiv:2412.05579, 2024a. Haitao Li, Qian Dong, Junjie Chen, Huixue Su, Yujia Zhou, Qingyao Ai, Ziyi Ye, and Yiqun Liu. Llms-as-judges: A comprehensive survey on llm-based evalua...

    arXiv

Show all 19 references
  1. [12]

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

    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, 2024a. 11 Preprint. Under review Chris Yuhao Liu, Liang Zeng, Jiacai Liu, Rui Yan, ...

    arXiv

  2. [13]

    Colbertv2: Effective and efficient retrieval via lightweight late interaction.arXiv preprint arXiv:2112.01488,

    Keshav Santhanam, Omar Khattab, Jon Saad-Falcon, Christopher Potts, and Matei Zaharia. Colbertv2: Effective and efficient retrieval via lightweight late interaction.arXiv preprint arXiv:2112.01488,

    arXiv

  3. [15]

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al

    URL https://arxiv.org/abs/2505.10320. 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 syste...

  4. [16]

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

    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 environ- ments.arXiv preprint arXiv:2504.03160,

    arXiv

  5. [17]

    Deepreview: Improving llm-based paper review with human-like deep thinking process.arXiv preprint arXiv:2503.08569,

    Minjun Zhu, Yixuan Weng, Linyi Yang, and Yue Zhang. Deepreview: Improving llm-based paper review with human-like deep thinking process.arXiv preprint arXiv:2503.08569,

    arXiv

  6. [18]

    Ttrl: Test-time reinforcement learning.arXiv preprint arXiv:2504.16084,

    Yuxin Zuo, Kaiyan Zhang, Li Sheng, Shang Qu, Ganqu Cui, Xuekai Zhu, Haozhan Li, Yuchen Zhang, Xinwei Long, Ermo Hua, et al. Ttrl: Test-time reinforcement learning.arXiv preprint arXiv:2504.16084,

    arXiv

  7. [19]

    <search>

    adopts a multi-stage preference optimization pipeline with iterative reward model updates.RM-R1(Chen et al., 2025b) is trained with reinforcement learning signals derived from rule-based reward functions. For all train-based reward models, we use the 7B versions re- leased and...

    2023

  8. [2021]

    Deepseekmath: Pushing the limits of mathemati- cal reasoning in open language models.arXiv preprint arXiv:2402.03300,

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

    arXiv

  9. [2023]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capa- bilities.arXiv preprint arXiv:2507.06261,

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capa- b...

    arXiv

  10. [2024]

    Judgelrm: Large reasoning models as a judge.arXiv preprint arXiv:2504.00050, 2025a

    Nuo Chen, Zhiyuan Hu, Qingyun Zou, Jiaying Wu, Qian Wang, Bryan Hooi, and Bingsheng He. Judgelrm: Large reasoning models as a judge.arXiv preprint arXiv:2504.00050, 2025a. Xiusi Chen, Gaotang Li, Ziqi Wang, Bowen Jin, Cheng Qian, Yu Wang, Hongru Wang, Yu Zhang, Denghui Zhang, ...

  11. [2025]

    Travelagent: An ai assistant for personalized travel planning.arXiv preprint arXiv:2409.08069,

    Aili Chen, Xuyang Ge, Ziquan Fu, Yanghua Xiao, and Jiangjie Chen. Travelagent: An ai assistant for personalized travel planning.arXiv preprint arXiv:2409.08069,

    arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.