Pith. sign in

REVIEW 3 major objections 6 minor 177 references

AgentOmnia claims that one taxonomy-driven loop—covering data synthesis, post-training, evaluation, and self-evolution—can lift a 30B-scale agent from a 22.9% to a 41.7% macro-average across four agent benchmarks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 03:32 UTC pith:4AISA6GW

load-bearing objection Solid engineering report with honest limitations: the headline OmniaBench gain is partly circular, but the three external benchmark gains make the central claim credible enough to referee. the 3 major comments →

arxiv 2607.23124 v1 pith:4AISA6GW submitted 2026-07-25 cs.AI cs.CL

AgentOmnia: Scaling Agentic Models for Full-Scenario Applications

Hao Jiang , Gangtao Xin , Yingdi Huang , Guojie Zhu , Jiangshan Zhang , Xinyuan Lin , Yunkun Xu , Chengyu Shen
show 30 more authors
Wenlong Fei Jiawei Li Yujie Fu Sichen Kang Tingyu Xie Yedi Hu Jingren Zhang Hongcheng Gao Jianshu Zeng Chong Chen Chang Guo Chao Feng Feng Wang Fulin Lin Jinchao Ma Lang Mei Li Huang Liyan Liu Qing He Shuting Tao Siyu Mo Xiangnan Chen Xiaohan Yu Xiaoyang Li Yanheng Hou Yanyu Wu Zhihan Yang Wentao Zhang Yang Gao Zhao Cao
This is my paper
classification cs.AI cs.CL
keywords full-scenario agentic scalingagentic post-trainingenvironment synthesistask synthesisweak-to-strong supervisionrollback curriculum RLPRD-guided self-evolutionOmniaBench
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

AgentOmnia is trying to establish that agent capability can be scaled across the full space of real-world applications—consumer, business, and employee tasks—rather than improved one benchmark at a time. Its central claim is that a shared three-axis taxonomy (domain × capability × atomic difficulty) can align every stage from data synthesis to post-training to diagnosis, and that this alignment turns evaluation failures into structured product requirements for the next training round. Starting from a 30B-parameter thinking-model checkpoint, the authors report raising the OmniaBench challenging-subset pass rate from 9.16% to 37.11% and the macro-average over OmniaBench, τ²-Bench, DeepPlanning, and VitaBench from 22.86% to 41.69%, with gains spread across most domains, all capability dimensions, and all difficulty factors. If correct, the loop offers a formula for building agents that operate reliably across heterogeneous applications and for steering future training with evidence about where they fail.

Core claim

The paper's discovery claim is that full-scenario agentic scaling is tractable as a closed loop: define an extensible Domain × Capability × Atomic Difficulty taxonomy; instantiate it with bidirectional environment–task synthesis (5,018 executable, stateful environments with 255,375 tools and 52,361 tasks); generate verified training trajectories through capability-aware privileged guidance; post-train with supervised fine-tuning followed by rollback-based curriculum reinforcement learning that recovers signal from all-fail rollouts; and convert evaluation failures into Product Requirement Documents (PRDs) that drive the next synthesis round. On this recipe AgentOmnia improves its foundation

What carries the argument

The load-bearing mechanism is the Domain × Capability × Atomic Difficulty taxonomy: a three-axis coordinate system that indexes every environment, task, trajectory, verifier, and rubric, so that what is synthesized, what is trained on, what is measured, and what is diagnosed all share one vocabulary. Around it the paper builds a weak-to-strong synthesis and post-training pipeline: language models propose environments, tasks, and trajectories, while programs, solvers, state-transition checks, and rubrics supply correctness signals. Privileged guidance—planning-oriented or outcome-oriented—helps a teacher model solve tasks beyond its direct-rollout ability, and leakage checks strip privileged

Load-bearing premise

The claim rests on OmniaBench scores being a measure of generalization rather than an artifact of benchmark–training alignment: the benchmark shares the same taxonomy and parts of the synthesis methodology, and the paper itself cautions that the large OmniaBench gain should be read in that light.

What would settle it

A test that would settle it: split OmniaBench into tasks whose level-2 domains and synthesis style overlap the training taxonomy and tasks that do not, then compare AgentOmnia's improvement over its foundation model on each group; if the gain concentrates on the overlapping group, the headline OmniaBench improvement is benchmark alignment, not full-scenario scaling. A matched control trained on external-benchmark-style data alone would offer a second check.

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

If this is right

  • If the central claim holds, agent post-training no longer needs benchmark-specific recipes: one taxonomy-aligned loop produces gains on all four evaluated benchmark families from a single model.
  • Because gains appear across 76 of 90 level-1 domains, all ten capability dimensions, and all eight difficulty factors, the recipe yields a broad capability shift rather than a single-format specialization.
  • Verified trajectories produced with privileged guidance and program/solver checks can raise the ceiling beyond what the teacher model can solve by direct rollout, so post-training data quality no longer depends on the teacher's rollout success rate.
  • The rollback curriculum turns otherwise all-fail reinforcement-learning groups into learnable signal, addressing the sparse-reward problem that typically halts agentic RL on hard tasks.
  • Evaluation failures can be converted into PRDs that specify the next round's environments, tasks, and success criteria, making model improvement a traceable product-development loop rather than ad hoc data collection.

Where Pith is reading between the lines

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

  • My inference: the +27.95-point OmniaBench gain is the least conservative estimate of the method's effect; the paper's own caveat about taxonomy alignment and synthesis-methodology reuse means the three external-benchmark gains are the cleaner estimate of pure transfer.
  • My inference: a decisive control would hold the training recipe fixed and randomize only the taxonomy coordinates used to target synthesis; if taxonomically targeted data beats equal-sized randomly selected data, the coordinate system itself is causally doing work.
  • My inference: for multi-round PRD evolution, the main risk is over-fitting to diagnostic signals; the paper's practice of excluding external benchmarks from target construction and checking transfer on them is a guard that should remain mandatory in every round.
  • My inference: the PRD interface generalizes to industrial settings where product teams draft requirements; if that path works, deployment logs and stakeholder specifications can become training inputs without manual data engineering.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper presents AgentOmnia, a closed-loop framework for full-scenario agentic AI: a Domain × Capability × Atomic Difficulty taxonomy aligns data synthesis, post-training, evaluation, and PRD-guided self-evolution. The authors construct 5,018 stateful environments, 52,361 tasks across DAG-, program-, and solver-based synthesis pipelines, and 53K SFT / 5K RL training instances. Starting from Qwen3-30B-A3B-Thinking-2507, the trained AgentOmnia-30B-A3B reports a pass-rate gain from 9.16% to 37.11% on the OmniaBench challenging subset and a four-benchmark macro-average gain from 22.86% to 41.69%, with double-digit gains also on three external benchmarks (τ²-Bench, DeepPlanning, VitaBench). The paper further reports fine-grained OmniaBench improvements across application splits, capability dimensions, and atomic difficulty factors, and a preliminary one-round PRD-guided self-evolution study.

Significance. If the gains generalize, the framework is a significant engineering and methodological advance: it integrates task-space definition, executable environment synthesis, verifiable trajectory generation, SFT/RL post-training, and diagnosis-driven data evolution into one loop. The strongest evidence is the consistent double-digit improvement over the base model on three independently authored external benchmarks (τ²-Bench, DeepPlanning, VitaBench), which is not attributable to the companion benchmark alone. The paper is also transparent about the main confound: the headline OmniaBench result is explicitly qualified in §7.2 as reflecting taxonomy alignment and partial reuse of the synthesis methodology. The scale of the synthesized corpus and the effort to unify evaluation protocols are notable strengths. However, the central leadership claim — the highest four-benchmark macro-average — relies on including the companion benchmark; the external-only macro-average is below two recent baselines. The fine-grained domain/capability/difficulty claims rest entirely on OmniaBench, whose task instances are deduplicated but whose environment/tool-level overlap with the training corpus is not quan

major comments (3)
  1. [§7.1, §7.2; Tables 9–12] The headline +27.95 pp OmniaBench gain and the fine-grained generalization claims (76/90 domains, all ten capability dimensions, all eight atomic difficulty factors) rest on OmniaBench, which is companion work [22] by overlapping authors and shares the same taxonomy and synthesis methodology as the AgentOmnia training corpus. §7.1 states only that OmniaBench task instances are 'deduplicated against the AgentOmnia post-training corpus'; this is a task-level check and does not rule out overlap at the level of environments, tool schemas, state containers, initial-state generators, or verifier templates. The §7.2 concession that the 'larger OmniaBench gain should be interpreted in light of its taxonomy alignment and partial reuse of the synthesis methodology' is not quantified. Please provide environment-level overlap statistics between the OmniaBench challenging set and the training environ
  2. [§7.2, Table 9] The four-benchmark macro-average lead over Agents-A1 is 41.69 vs. 41.52, a difference of 0.17 percentage points. The paper reports no confidence intervals, bootstrap estimates, repeated-run variance, or significance testing. Given the OmniaBench challenging subset has 644 tasks and several external track sizes are small, this lead is plausibly within sampling noise. The 'highest' claim should be either accompanied by uncertainty quantification or restricted to a more modest statement such as 'competitive with' the top agentic post-trained baselines.
  3. [§5.2, §7.2; Figure 14] The paper presents RCRL and privileged guidance as named contributions, but the experiments do not isolate their downstream contribution. Figure 14 shows that privileged guidance increases verified trajectory yield during synthesis, and Table 9 reports the combined SFT+RL recipe, but there is no comparison such as SFT-only, SFT+RL without RCRL, SFT+RL without privileged guidance, or ablation of the rollback curriculum. Consequently, the reader cannot determine whether the external benchmark gains come from the data corpus, the SFT stage, the RCRL objective, or privileged trajectory synthesis. This is not fatal to the framework-level claim, but it is load-bearing for the specific algorithmic contributions claimed in §5.2.
minor comments (6)
  1. [Table 9] The column headers 'Avg. S', 'T Avg.', 'T Comp.', 'T CS', 'T PS' are not self-explanatory; the caption should explain the nesting of Shopping/Travel averages and the components of DeepPlanning.
  2. [Throughout] The name of the external benchmark is written inconsistently as 'τ 2-Bench' and 'τ²-Bench'. Please unify.
  3. [Figure 18] The KL divergences are computed over the displayed top-50 level-1 domains, not necessarily the full 90-domain taxonomy. Please state the exact domain set used for the KL calculation.
  4. [§7.2, Figure 17] The '76 of 90 domains improve' claim is based on the 644-task challenging subset; some domain cells contain very few tasks. Please report per-domain task counts and consider excluding domains with fewer than a minimum number of tasks from the aggregate percentage claim.
  5. [§7.3, Table 13] The one-round self-evolution results are presented without variance or significance measures. The paper appropriately calls them preliminary, but a statement about the number of diagnostic tasks, the PRD generation cost, and any filtering would help reproducibility.
  6. [General] The paper does not include an availability statement for the trained model or the training data. Even if only OmniaBench is released, please state explicitly what artifacts will be made public.

Circularity Check

2 steps flagged

OmniaBench-based claims are partly circular: the main gain is measured on an overlapping companion benchmark sharing the training taxonomy, and the self-evolution gain is evaluated on the same benchmark used to derive its training targets; only the three external benchmarks give independent support.

specific steps
  1. self citation load bearing [Section 7.2, Main Results (paragraph after Table 9)]
    "The larger OmniaBench gain should be interpreted in light of its taxonomy alignment and partial reuse of the synthesis methodology, although its task instances are deduplicated from the training corpus. Improvements on all three independently developed external benchmarks provide complementary evidence that the effect extends beyond the companion suite."

    OmniaBench is companion work [22] by overlapping authors, and the same Domain × Capability × Atomic Difficulty taxonomy is used to drive training-data synthesis and to define the benchmark: §2 says OmniaBench 'instantiates this design' and 'During synthesis, the same coordinates are used to track coverage and specify task difficulty.' Thus the +27.95 pp OmniaBench gain and the 'highest four-benchmark macro-average' partly measure a model evaluated on a benchmark constructed from the same taxonomy, methodology, and author team as the training corpus, rather than an independent generalization result. The paper's own sentence concedes this alignment and reuse.

  2. fitted input called prediction [Section 7.3 / Table 13 caption]
    "OmniaBench uses the challenging subset that provides the diagnostic signals for PRD target construction, while τ2-Bench, DeepPlanning, and VitaBench are excluded from target construction and used to assess transfer."

    In the one-round self-evolution study, PRDs are generated from OmniaBench failure diagnoses and guide construction of 121 environments and 804 tasks; the model is then evaluated on the same OmniaBench challenging subset and the improvement from 37.11% to 38.49% is reported as 'model improvement.' Because the training data is synthesized from the evaluation benchmark's own failures, the OmniaBench gain is partly forced by construction: the model is trained on material derived from the test set and then re-measured on that test set. The external-benchmark gains are the only non-forced part of this sub-claim.

full rationale

AgentOmnia's central evaluation narrative has two partly circular components. First, the headline OmniaBench gain (+27.95 pp) and the 'highest four-benchmark macro-average' rely on OmniaBench, companion work [22] by overlapping authors, and the paper itself states that the same taxonomy is used for synthesis and evaluation, with partial reuse of the synthesis methodology; the gain therefore partly reflects benchmark–training alignment rather than independent generalization. Second, the PRD-guided self-evolution study derives its training environments and tasks from OmniaBench failure diagnoses and then reports improvement on the same OmniaBench challenging subset; that gain reduces by construction because the evaluation set supplied the training signal. The three external benchmarks (τ2-Bench, DeepPlanning, VitaBench) are independently developed and are explicitly excluded from PRD target construction, so the core post-training improvement claim retains genuine non-circular support; however, the external-only macro-average is lower than Agents-A1 and Nex-N2-Mini, so the leadership claim depends on the circular companion benchmark. No uniqueness-theorem or ansatz-smuggling pattern was found. Overall, the paper discloses the main alignment issue in §7.2, but the disclosure confirms rather than removes the circularity, and the self-evolution step is a concrete case of a reported 'prediction' that is partly an artifact of test-set-derived training. Score 6: one or more reported gains reduce by construction, while the framework's central transfer claim still has independent external evidence.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced. The central claim rests on two classes of assumptions: synthetic environments/verifiers faithfully represent real agent tasks, and evaluation, especially OmniaBench, is a fair measure of generalization. Several hand-set hyperparameters (RCRL prefixes, clip thresholds, loss weights) influence the reported gains but are not derived, and their values are only partially disclosed.

free parameters (5)
  • RCRL initial gold-prefix increment ratio M^(0) = 20% of golden trajectory length
    Chosen coarsely in Algorithm 1 to localize trajectory bottlenecks; directly controls the curriculum but no principled derivation is given.
  • RL clip thresholds ε_low, ε_high = 0.2, 0.28
    Set following prior asymmetric-clipping work; not derived for this agentic regime.
  • Hybrid objective weights λ_CE, λ_RL = not reported
    Weights in Eq. (11) balance golden-prefix cross-entropy against on-policy GRPO loss; values and selection process are not disclosed.
  • Reward threshold R_th and max rollback step δ_max = not reported
    Parameters in Eq. (10) adaptively expand or shrink the golden prefix; their values are not given, yet the curriculum depends on them.
  • SFT corpus size and RL task count = 53K SFT samples; 5K RL tasks
    Chosen via convergence-aware budgeting and pass@K filtering; these resource choices shape all reported results.
axioms (5)
  • domain assumption Synthesized executable environments are faithful proxies for real-world ToC/ToB/ToE tasks.
    Section 4.2 grounds environments in code and LLM-based utility/realism scoring, but no real-user validation is provided; the transfer claim rests on this.
  • domain assumption Rule-, rubric-, and verifier-based reward signals correctly identify task success.
    Section 5.2.2 uses these as RL rewards; reward-hacking audits are themselves LLM-based, so correctness of the evaluators is assumed rather than proven against external ground truth.
  • domain assumption OmniaBench task instances are free of meaningful training contamination because they were deduplicated against the training corpus.
    Section 7.1 asserts deduplication and manual curation, but no contamination analysis or independent audit is provided; the paper concedes partial methodology reuse in §7.2.
  • domain assumption Privileged guidance does not leak into retained training trajectories.
    Section 4.5.3 reports high but imperfect evidence-grounding pass rates, and the leakage check is an automated LLM/rule filter, not human verification.
  • domain assumption Group-relative RL with binary rule/rubric rewards improves the policy as expected under the clipped objective in Eq. (8).
    The GRPO-style update, with KL removed and asymmetric clipping, is treated as a reliable optimizer; no convergence analysis or comparison to alternative RL objectives is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 48784 in / 14544 out tokens · 147024 ms · 2026-08-01T03:32:48.283339+00:00 · methodology

0 comments
read the original abstract

Large language model agents have advanced rapidly, yet progress remains fragmented across domains, capabilities, task difficulty, and interaction settings. We frame this as full-scenario agentic scaling and present AgentOmnia, a framework coordinating task-space definition, data synthesis, post-training, evaluation, and improvement across To-Consumer (ToC), To-Business (ToB), and To-Employee (ToE) applications. An extensible Domain x Capability x Atomic Difficulty taxonomy aligns these stages and enables fine-grained diagnosis with OmniaBench. AgentOmnia combines bidirectional environment-task synthesis with tool-dependency, program-structured, and solver-based pipelines, constructing 5,018 stateful environments with 255,375 tools and 52,361 tasks. Programs, solvers, and verifiers provide correctness signals, while supervised fine-tuning, online agentic reinforcement learning, and a rollback curriculum support post-training. Evaluation failures translate into Product Requirement Documents (PRDs) for targeted self-evolution. Starting from Qwen3-30B-A3B-Thinking-2507, AgentOmnia raises the pass rate on the OmniaBench challenging subset from 9.16% to 37.11% and the macro-average across OmniaBench, $\tau^2$-Bench, DeepPlanning, and VitaBench from 22.86% to 41.69%. Under a unified protocol,it leads the evaluated agentic post-trained baselines on OmniaBench and retains the highest four-benchmark macro-average. It also surpasses Qwen3-235B-A22B-Thinking-2507 on all four benchmarks and exceeds Qwen3.5-35B-A3B on the macro-average. Gains span three application splits, ten capability dimensions, eight atomic-difficulty factors, and 76 of 90 level-1 domains, indicating broad rather than category-specific improvement. A one-round study provides initial evidence for PRD-guided self-evolution, motivating validation at larger scales and in industrial settings.

discussion (0)

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

Reference graph

Works this paper leans on

177 extracted references · 55 linked inside Pith

  1. [1]

    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. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  2. [2]

    Toolformer: Language models can teach themselves to use tools

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools. InAdvances in Neural Information Processing Systems, volume 36, 2023

    2023

  3. [3]

    ToolLLM: Facilitating large language models to master 16000+ real-world APIs

    Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Lauren Hong, Runchu Tian, Ruobing Xie, Jie Zhou, Mark Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun. ToolLLM: Facilitating large language models to master 16000+ real-world APIs. InThe Twelfth International Conference on Learning...

    2024

  4. [4]

    Patil, Tianjun Zhang, Xin Wang, and Joseph E

    Shishir G. Patil, Tianjun Zhang, Xin Wang, and Joseph E. Gonzalez. Gorilla: Large language model connected with massive apis. In Amir Globersons, Lester Mackey, Danielle Belgrave, Angela Fan, Ulrich Paquet, Jakub M. Tomczak, and Cheng Zhang, editors,Advances in Neural Information Processing Systems 37: Annual Conference on Neural Information Processing Sy...

    2024

  5. [5]

    Tooltalk: Evaluating tool-usage in a conversational setting.CoRR, abs/2311.10775, 2023

    Nicholas Farn and Richard Shin. Tooltalk: Evaluating tool-usage in a conversational setting.CoRR, abs/2311.10775, 2023. doi: 10.48550/ARXIV.2311.10775. URLhttps://doi.org/10.48550/arXiv.2311.10775

    Pith/arXiv arXiv doi:10.48550/arxiv.2311.10775 2023

  6. [6]

    API-Bank: A comprehensive benchmark for tool-augmented LLMs.arXiv preprint arXiv:2304.08244, 2023

    Minghao Li, Yingxiu Zhao, Bowen Yu, Feifan Song, Hangyu Li, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. API-Bank: A comprehensive benchmark for tool-augmented LLMs.arXiv preprint arXiv:2304.08244, 2023. URL https://arxiv.org/abs/2304.08244

    Pith/arXiv arXiv 2023

  7. [7]

    AgentBench: Evaluating LLMs as agents

    Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, Shudan Zhang, Xiang Deng, Aohan Zeng, Zhengxiao Du, Chenhui Zhang, Sheng Shen, Tianjun Zhang, Yu Su, Huan Sun, Minlie Huang, Yuxiao Dong, and Jie Tang. AgentBench: Evaluating LLMs as agents. InThe Twelfth International Conference on Learning R...

    2024

  8. [8]

    WebArena: A realistic web environment for building autonomous agents

    Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig. WebArena: A realistic web environment for building autonomous agents. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  9. [9]

    AndroidWorld: A dynamic benchmarking environment for autonomous agents.arXiv preprint arXiv:2405.14573, 2024

    Christopher Rawles, Sarah Clinckemaillie, Yifan Chang, Jonathan Waltz, Gabrielle Lau, Marybeth Fair, Alice Li, William Bishop, Wei Li, Folawiyo Campbell-Ajala, Daniel Toyama, Robert Berry, Divya Tyamagundlu, Timothy Lillicrap, and Oriana Riva. AndroidWorld: A dynamic benchmarking environment for autonomous agents.arXiv preprint arXiv:2405.14573, 2024. URL...

    Pith/arXiv arXiv 2024

  10. [10]

    OSWorld: Benchmarking multimodal agents for open-ended tasks in real computer environments.arXiv preprint arXiv:2404.07972, 2024

    Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, Yitao Liu, Yiheng Xu, Shuyan Zhou, Silvio Savarese, Caiming Xiong, Victor Zhong, and Tao Yu. OSWorld: Benchmarking multimodal agents for open-ended tasks in real computer environments.arXiv preprint arXiv:2404.07972, 20...

    Pith/arXiv arXiv 2024

  11. [11]

    τ 2-Bench: Evaluating Conversational Agents in a Dual-Control Environment.arXiv preprint arXiv:2506.07982, 2025

    Victor Barres, Honghua Dong, Soham Ray, Xujie Si, and Karthik Narasimhan. τ 2-Bench: Evaluating Conversational Agents in a Dual-Control Environment.arXiv preprint arXiv:2506.07982, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    DeepPlanning: Benchmarking long-horizon agentic planning with verifiable constraints.arXiv preprint arXiv:2601.18137, 2026

    Yinger Zhang, Shutong Jiang, Renhao Li, Jianhong Tu, Yang Su, Lianghao Deng, Xudong Guo, Chenxu Lv, and Junyang Lin. DeepPlanning: Benchmarking long-horizon agentic planning with verifiable constraints.arXiv preprint arXiv:2601.18137, 2026

    arXiv 2026

  13. [13]

    VitaBench: Benchmarking LLM agents with versatile interactive tasks in real-world applications.arXiv preprint arXiv:2509.26490, 2025

    Wei He, Yueqing Sun, Hongyan Hao, Xueyuan Hao, Zhikang Xia, Qi Gu, Chengcheng Han, Dengchang Zhao, Hui Su, Kefeng Zhang, Man Gao, Xi Su, Xiaodong Cai, Xunliang Cai, Yang Yu, and Yunke Zhao. VitaBench: Benchmarking LLM agents with versatile interactive tasks in real-world applications.arXiv preprint arXiv:2509.26490, 2025

    arXiv 2025

  14. [14]

    The tool decathlon: Benchmarking language agents for diverse, realistic, and long-horizon task execution.arXiv preprint arXiv:2510.25726, 2025

    Junlong Li, Wenshuo Zhao, Jian Zhao, Weihao Zeng, Haoze Wu, Xiaochen Wang, Rui Ge, Yuxuan Cao, Yuzhen Huang, Wei Liu, Junteng Liu, Zhaochen Su, Yiyang Guo, Fan Zhou, Lueyang Zhang, Juan Michelini, Xingyao Wang, Xiang Yue, Shuyan Zhou, Graham Neubig, and Junxian He. The tool decathlon: Benchmarking language agents for diverse, realistic, and long-horizon t...

    arXiv 2025

  15. [15]

    Gaia2: Benchmarking LLM agents on dynamic and asynchronous environments.arXiv preprint arXiv:2602.11964, 2026

    Romain Froger, Pierre Andrews, Matteo Bettini, Amar Budhiraja, Ricardo Silveira Cabral, Virginie Do, Emilien Garreau, Jean-Baptiste Gaya, Hugo Laurençon, Maxime Lecanu, Kunal Malkan, Dheeraj Mekala, Pierre Ménard, Gerard Moreno- Torres Bertran, Ulyana Piterbarg, Mikhail Plekhanov, Mathieu Rita, Andrey Rusakov, Vladislav V orotilov, Mengjue Wang, Ian Yu, A...

    arXiv 2026

  16. [16]

    AgentTuning: Enabling generalized agent abilities for LLMs

    Aohan Zeng, Mingdao Liu, Rui Lu, Bowen Wang, Xiao Liu, Yuxiao Dong, and Jie Tang. AgentTuning: Enabling generalized agent abilities for LLMs. InFindings of the Association for Computational Linguistics: ACL 2024, pages 3053–3077, 2024. 38

    2024

  17. [18]

    AgentSkiller: Scaling generalist agent intelligence through semantically integrated cross-domain data synthesis.arXiv preprint arXiv:2602.09372, 2026

    Zexu Sun, Bokai Ji, Hengyi Cai, Shuaiqiang Wang, Lei Wang, Guangxia Li, and Xu Chen. AgentSkiller: Scaling generalist agent intelligence through semantically integrated cross-domain data synthesis.arXiv preprint arXiv:2602.09372, 2026. URL https://arxiv.org/abs/2602.09372

    arXiv 2026

  18. [19]

    Agent-World: Scaling real-world environment synthesis for evolving general agent intelligence.arXiv preprint arXiv:2604.18292, 2026

    Guanting Dong, Junting Lu, Junjie Huang, Wanjun Zhong, Longxiang Liu, Shijue Huang, Zhenyu Li, Yang Zhao, Xiaoshuai Song, Xiaoxi Li, Jiajie Jin, Yutao Zhu, Hanbin Wang, Fangyu Lei, Qinyu Luo, Mingyang Chen, Zehui Chen, Jiazhan Feng, Ji-Rong Wen, and Zhicheng Dou. Agent-World: Scaling real-world environment synthesis for evolving general agent intelligence...

    Pith/arXiv arXiv 2026

  19. [20]

    Agent world model: Infinity synthetic environments for agentic reinforcement learning.arXiv preprint arXiv:2602.10090, 2026

    Zhaoyang Wang, Canwen Xu, Boyi Liu, Yite Wang, Siwei Han, Zhewei Yao, Huaxiu Yao, and Yuxiong He. Agent world model: Infinity synthetic environments for agentic reinforcement learning.arXiv preprint arXiv:2602.10090, 2026

    Pith/arXiv arXiv 2026

  20. [21]

    Scaling the horizon, not the parameters: Reaching trillion-parameter performance with a 35b agent.arXiv preprint arXiv:2606.30616, 2026

    Lei Bai, Zongsheng Cao, Yang Chen, Zhiyao Cui, Shangheng Du, Yue Fan, Shiyang Feng, Zijie Guo, Haonan He, Liang He, et al. Scaling the horizon, not the parameters: Reaching trillion-parameter performance with a 35b agent.arXiv preprint arXiv:2606.30616, 2026. URLhttps://arxiv.org/abs/2606.30616

    Pith/arXiv arXiv 2026

  21. [22]

    Omniabench: Benchmarking general ai agents across diverse scenarios, 2026

    Chengyu Shen, Yujie Fu, Gangtao Xin, Yanheng Hou, Wenlong Fei, Guojie Zhu, Jiawei Li, Hongcheng Gao, Runming He, Zhen Hao Wong, Meiyi Qiang, Hao Liang, Zhao Cao, Hao Jiang, Chong Chen, and Wentao Zhang. Omniabench: Benchmarking general ai agents across diverse scenarios, 2026. URLhttps://arxiv.org/abs/2607.14989

    Pith/arXiv arXiv 2026

  22. [23]

    AppWorld: A controllable world of apps and people for benchmarking interactive coding agents.arXiv preprint arXiv:2407.18901, 2024

    Harsh Trivedi, Tushar Khot, Mareike Hartmann, Ruskin Manku, Vinty Dong, Edward Li, Shashank Gupta, Ashish Sabharwal, and Niranjan Balasubramanian. AppWorld: A controllable world of apps and people for benchmarking interactive coding agents.arXiv preprint arXiv:2407.18901, 2024. URLhttps://arxiv.org/abs/2407.18901

    Pith/arXiv arXiv 2024

  23. [24]

    Laradji, Manuel Del Verme, Tom Marty, Léo Boisvert, Megh Thakkar, Quentin Cappart, David Vazquez, Nicolas Chapados, and Alexandre Lacoste

    Alexandre Drouin, Maxime Gasse, Massimo Caccia, Issam H. Laradji, Manuel Del Verme, Tom Marty, Léo Boisvert, Megh Thakkar, Quentin Cappart, David Vazquez, Nicolas Chapados, and Alexandre Lacoste. WorkArena: How capable are web agents at solving common knowledge work tasks?arXiv preprint arXiv:2403.07718, 2024. URL https://arxiv.org/abs/ 2403.07718

    Pith/arXiv arXiv 2024

  24. [25]

    CRMArena: Understanding the capacity of LLM agents to perform professional CRM tasks in realistic environments.arXiv preprint arXiv:2411.02305, 2024

    Kung-Hsiang Huang, Akshara Prabhakar, Sidharth Dhawan, Yixin Mao, Huan Wang, Silvio Savarese, Caiming Xiong, Philippe Laban, and Chien-Sheng Wu. CRMArena: Understanding the capacity of LLM agents to perform professional CRM tasks in realistic environments.arXiv preprint arXiv:2411.02305, 2024. URLhttps://arxiv.org/abs/2411.02305

    Pith/arXiv arXiv 2024

  25. [26]

    Gdpval: Evaluating ai model performance on real-world economically valuable tasks.arXiv preprint arXiv:2510.04374, 2025

    Tejal Patwardhan, Rachel Dias, Elizabeth Proehl, Grace Kim, Michele Wang, Olivia Watkins, Simón Posada Fishman, Marwan Aljubeh, Phoebe Thacker, Laurance Fauconnet, et al. Gdpval: Evaluating ai model performance on real-world economically valuable tasks.arXiv preprint arXiv:2510.04374, 2025

    Pith/arXiv arXiv 2025

  26. [27]

    OfficeBench: Benchmarking language agents across multiple applications for office automation.arXiv preprint arXiv:2407.19056, 2024

    Zilong Wang, Yuedong Cui, Li Zhong, Zimin Zhang, Da Yin, Bill Yuchen Lin, and Jingbo Shang. OfficeBench: Benchmarking language agents across multiple applications for office automation.arXiv preprint arXiv:2407.19056, 2024. URL https: //arxiv.org/abs/2407.19056

    Pith/arXiv arXiv 2024

  27. [28]

    SpreadsheetBench: Towards challenging real world spreadsheet manipulation.arXiv preprint arXiv:2406.14991, 2024

    Zeyao Ma, Bohan Zhang, Jing Zhang, Jifan Yu, Xiaokang Zhang, Xiaohan Zhang, Sijia Luo, Xi Wang, and Jie Tang. SpreadsheetBench: Towards challenging real world spreadsheet manipulation.arXiv preprint arXiv:2406.14991, 2024. URL https://arxiv.org/abs/2406.14991

    Pith/arXiv arXiv 2024

  28. [29]

    Agentsynth: Scalable task generation for generalist computer-use agents.arXiv (Cornell University), abs/2506.14205, 2025

    Jingxu Xie, Dylan Xu, Xuandong Zhao, and Dawn Song. Agentsynth: Scalable task generation for generalist computer-use agents.arXiv (Cornell University), abs/2506.14205, 2025. doi: 10.48550/arxiv.2506.14205

    doi:10.48550/arxiv.2506.14205 2025

  29. [30]

    Envscaler: Scaling tool- interactive environments for llm agent via programmatic synthesis, 2026

    Xiaoshuai Song, Haofei Chang, Guanting Dong, Yutao Zhu, Ji-Rong Wen, and Zhicheng Dou. Envscaler: Scaling tool- interactive environments for llm agent via programmatic synthesis, 2026. URLhttps://arxiv.org/abs/2601.05808

    Pith/arXiv arXiv 2026

  30. [31]

    Genenv: Difficulty-aligned co-evolution between LLM agents and environment simulators.CoRR, abs/2512.19682, 2025

    Jiacheng Guo, Ling Yang, Peter Chen, Qixin Xiao, Yinjie Wang, Xinzhe Juan, Jiahao Qiu, Ke Shen, and Mengdi Wang. Genenv: Difficulty-aligned co-evolution between LLM agents and environment simulators.CoRR, abs/2512.19682, 2025. doi: 10.48550/ARXIV.2512.19682. URLhttps://doi.org/10.48550/arXiv.2512.19682

    doi:10.48550/arxiv.2512.19682 2025

  31. [32]

    Graph2eval: Automatic multimodal task generation for agents via knowledge graphs.CoRR, abs/2510.00507,

    Yurun Chen, Xavier Hu, Yuhan Liu, Ziqi Wang, Zeyi Liao, Lin Chen, Feng Wei, Yuxi Qian, Bo Zheng, Keting Yin, and Shengyu Zhang. Graph2eval: Automatic multimodal task generation for agents via knowledge graphs.CoRR, abs/2510.00507,

    arXiv

  32. [33]

    Close the loop: Synthesizing infinite tool-use data via multi-agent role-playing.arXiv preprint arXiv:2512.23611, 2025

    Yuwen Li, Wei Zhang, Zelong Huang, Mason Yang, Jiajun Wu, Shawn Guo, Huahao Hu, Lingyi Sun, Jian Yang, Mingjie Tang, and Byran Dai. Close the loop: Synthesizing infinite tool-use data via multi-agent role-playing.arXiv preprint arXiv:2512.23611, 2025. URLhttps://arxiv.org/abs/2512.23611

    arXiv 2025

  33. [34]

    URLhttps://doi.org/10.48550/arXiv.2510.00507

    doi: 10.48550/ARXIV.2510.00507. URLhttps://doi.org/10.48550/arXiv.2510.00507

    doi:10.48550/arxiv.2510.00507

  34. [35]

    Reasoning with language model is planning with world model

    Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhiting Hu. Reasoning with language model is planning with world model. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 8154–8173, 2023

    2023

  35. [36]

    The false promise of imitating proprietary language models

    Arnav Gudibande, Eric Wallace, Charlie Snell, Xinyang Geng, Hao Liu, Pieter Abbeel, Sergey Levine, and Dawn Song. The false promise of imitating proprietary language models. InThe Twelfth International Conference on Learning Representations, 2024. 39

    2024

  36. [37]

    Qwen3-30B-A3B-Thinking-2507

    Qwen Team. Qwen3-30B-A3B-Thinking-2507. Hugging Face model card, 2025. URL https://huggingface.co/Qwen/ Qwen3-30B-A3B-Thinking-2507

    2025

  37. [38]

    Qwen3 technical report.CoRR, abs/2505.09388, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.CoRR, abs/2505.09388, 2025. doi: 10.48550/ARXIV.2505.09388. URL https://doi.org/10.48550/arXiv.2505.09388

    Pith/arXiv arXiv doi:10.48550/arxiv.2505.09388 2025

  38. [39]

    Weak-to-strong generalization: Eliciting strong capabilities with weak supervision

    Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold Aschenbrenner, Yining Chen, Adrien Ecoffet, Manas Joglekar, Jan Leike, Ilya Sutskever, and Jeffrey Wu. Weak-to-strong generalization: Eliciting strong capabilities with weak supervision. InProceedings of the 41st International Conference on Machine Learning, 2024

    2024

  39. [40]

    Nex-N2-mini

    Nex AGI. Nex-N2-mini. Hugging Face model card, 2026. URLhttps://huggingface.co/nex-agi/Nex-N2-mini

    2026

  40. [41]

    AgentGym: Evolving large language model-based agents across diverse environments.arXiv preprint arXiv:2406.04151, 2024

    Zhiheng Xi, Yiwen Ding, Wenxiang Chen, Boyang Hong, Honglin Guo, Junzhe Wang, Dingwen Yang, Chenyang Liao, Xin Guo, Wei He, et al. AgentGym: Evolving large language model-based agents across diverse environments.arXiv preprint arXiv:2406.04151, 2024

    Pith/arXiv arXiv 2024

  41. [42]

    DigiRL: Training in-the-wild device-control agents with autonomous reinforcement learning

    Hao Bai, Yifei Zhou, Mert Cemri, Jiayi Pan, Alane Suhr, Sergey Levine, and Aviral Kumar. DigiRL: Training in-the-wild device-control agents with autonomous reinforcement learning. InAdvances in Neural Information Processing Systems, volume 37, 2024

    2024

  42. [43]

    Self-instruct: Aligning language models with self-generated instructions

    Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics, pages 13484–13508, 2023

    2023

  43. [45]

    URLhttps://arxiv.org/abs/2508.18669

    Pith/arXiv arXiv

  44. [46]

    Self-refine: Iterative refinement with self-feedback.arXiv preprint arXiv:2303.17651, 2023

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark. Self-refine: Iterative refinement with self-feedback.arXiv preprint arXiv:2303.17651, 2023

    Pith/arXiv arXiv 2023

  45. [47]

    Self- rewarding language models

    Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. Self- rewarding language models. InProceedings of the 41st International Conference on Machine Learning, 2024

    2024

  46. [48]

    Reflexion: Language agents with verbal reinforcement learning.arXiv preprint arXiv:2303.11366, 2023

    Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning.arXiv preprint arXiv:2303.11366, 2023

    Pith/arXiv arXiv 2023

  47. [49]

    SEAL: Synergistic co-evolution of agents and learning environments.arXiv preprint arXiv:2605.24426, 2026

    Yihao Hu, Zhihao Wen, Xiujin Liu, Pan Wang, Xin Zhang, and Wei Wu. SEAL: Synergistic co-evolution of agents and learning environments.arXiv preprint arXiv:2605.24426, 2026. URLhttps://arxiv.org/abs/2605.24426

    Pith/arXiv arXiv 2026

  48. [50]

    SELF: Self-evolution with language feedback.arXiv preprint arXiv:2310.00533, 2023

    Jianqiao Lu, Wanjun Zhong, Wenyong Huang, Yufei Wang, Qi Zhu, Fei Mi, Baojun Wang, Weichao Wang, Xingshan Zeng, Lifeng Shang, Xin Jiang, and Qun Liu. SELF: Self-evolution with language feedback.arXiv preprint arXiv:2310.00533, 2023

    Pith/arXiv arXiv 2023

  49. [51]

    AgentEvolver: Towards efficient self-evolving agent system.arXiv preprint arXiv:2511.10395, 2025

    Yunpeng Zhai, Shuchang Tao, Cheng Chen, Anni Zou, Ziqian Chen, Qingxu Fu, Shinji Mai, Li Yu, Jiaji Deng, Zouying Cao, Zhaoyang Liu, Bolin Ding, and Jingren Zhou. AgentEvolver: Towards efficient self-evolving agent system.arXiv preprint arXiv:2511.10395, 2025. URLhttps://arxiv.org/abs/2511.10395

    arXiv 2025

  50. [52]

    Towards general agentic intelligence via environment scaling

    Runnan Fang, Shihao Cai, Baixuan Li, Jialong Wu, Guangyu Li, Wenbiao Yin, Xinyu Wang, Xiaobin Wang, Liangcai Su, Zhen Zhang, et al. Towards general agentic intelligence via environment scaling. InFindings of the Association for Computational Linguistics: ACL 2026, pages 17610–17621, 2026

    2026

  51. [53]

    Yu, Ran Xu, Xiang Li, and Lichao Sun

    Zhiling Yan, Dingjie Song, Hanrong Zhang, Wei Liang, Yuxuan Zhang, Yutong Dai, Lifang He, Philip S. Yu, Ran Xu, Xiang Li, and Lichao Sun. OpenSkill: Open-world self-evolution for LLM agents.arXiv preprint arXiv:2606.06741, 2026. URL https://arxiv.org/abs/2606.06741

    Pith/arXiv arXiv 2026

  52. [54]

    Autoforge: Automated environment synthesis for agentic reinforcement learning.arXiv preprint arXiv:2512.22857, 2025

    Shihao Cai, Runnan Fang, Jialong Wu, Baixuan Li, Xinyu Wang, Yong Jiang, Liangcai Su, Liwen Zhang, Wenbiao Yin, Zhen Zhang, et al. Autoforge: Automated environment synthesis for agentic reinforcement learning.arXiv preprint arXiv:2512.22857, 2025

    arXiv 2025

  53. [55]

    Hindsight hint distillation: Scaffolded reasoning for swe agents from cot-free answers, 2026

    Shengjie Wang, Guanghe Li, Zonghan Yang, and Yang Gao. Hindsight hint distillation: Scaffolded reasoning for swe agents from cot-free answers, 2026. URLhttps://arxiv.org/abs/2605.11556

    Pith/arXiv arXiv 2026

  54. [56]

    Self-distilled reasoner: On-policy self-distillation for large language models, 2026

    Siyan Zhao, Zhihui Xie, Mengchen Liu, Jing Huang, Guan Pang, Feiyu Chen, and Aditya Grover. Self-distilled reasoner: On-policy self-distillation for large language models, 2026. URLhttps://arxiv.org/abs/2601.18734

    Pith/arXiv arXiv 2026

  55. [57]

    Edge-opd: Internalizing privileged context with evidence guided on-policy distillation, 2026

    Aristotelis Lazaridis, Dylan Bates, Aman Sharma, Brian King, Vincent Lu, and Jack FitzGerald. Edge-opd: Internalizing privileged context with evidence guided on-policy distillation, 2026. URLhttps://arxiv.org/abs/2605.23493. 40

    Pith/arXiv arXiv 2026

  56. [58]

    Let it flow: Agentic crafting on rock and roll, building the rome model within an open agentic learning ecosystem.arXiv preprint arXiv:2512.24873, 2025

    Weixun Wang, XiaoXiao Xu, Wanhe An, Fangwen Dai, Wei Gao, Yancheng He, Ju Huang, Qiang Ji, Hanqi Jin, Xiaoyang Li, et al. Let it flow: Agentic crafting on rock and roll, building the rome model within an open agentic learning ecosystem.arXiv preprint arXiv:2512.24873, 2025

    arXiv 2025

  57. [59]

    Le, Tengyu Ma, and Adams Wei Yu

    Sang Michael Xie, Hieu Pham, Xuanyi Dong, Nan Du, Hanxiao Liu, Yifeng Lu, Percy Liang, Quoc V . Le, Tengyu Ma, and Adams Wei Yu. Doremi: Optimizing data mixtures speeds up language model pretraining, 2023. URL https: //arxiv.org/abs/2305.10429

    Pith/arXiv arXiv 2023

  58. [60]

    Data mixing optimization for supervised fine-tuning of large language models, 2025

    Yuan Li, Zhengzhong Liu, and Eric Xing. Data mixing optimization for supervised fine-tuning of large language models, 2025. URLhttps://arxiv.org/abs/2508.11953

    Pith/arXiv arXiv 2025

  59. [61]

    The verification horizon: No silver bullet for coding agent rewards.arXiv preprint arXiv:2606.26300, 2026

    Binghai Wang, Chenlong Zhang, Dayiheng Liu, Jiajun Zhang, Jiawei Chen, Mouxiang Chen, Rongyao Fang, Siyuan Zhang, Xuwu Wang, Yuheng Jing, et al. The verification horizon: No silver bullet for coding agent rewards.arXiv preprint arXiv:2606.26300, 2026

    Pith/arXiv arXiv 2026

  60. [62]

    A survey on llm-as-a-judge.The Innovation, 7(6), 2026

    Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, et al. A survey on llm-as-a-judge.The Innovation, 7(6), 2026

    2026

  61. [63]

    From reasoning to agentic: Credit assignment in reinforcement learning for large language models.arXiv preprint arXiv:2604.09459, 2026

    Chenchen Zhang. From reasoning to agentic: Credit assignment in reinforcement learning for large language models.arXiv preprint arXiv:2604.09459, 2026

    Pith/arXiv arXiv 2026

  62. [64]

    When speed kills stability: Demystifying RL collapse from the training-inference mismatch, 2025

    Jiacai Liu, Yingru Li, Yuqian Fu, Jiawei Wang, Qian Liu, and Yu Shen. When speed kills stability: Demystifying RL collapse from the training-inference mismatch, 2025. URLhttps://richardli.xyz/rl-collapse

    2025

  63. [65]

    Reproducing, analyzing, and detecting reward hacking in rubric-based reinforcement learning.arXiv preprint arXiv:2606.04923, 2026

    Xuekang Wang, Zhuoyuan Hao, Shuo Hou, Hao Peng, Juanzi Li, and Xiaozhi Wang. Reproducing, analyzing, and detecting reward hacking in rubric-based reinforcement learning.arXiv preprint arXiv:2606.04923, 2026

    Pith/arXiv arXiv 2026

  64. [66]

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

    Pith/arXiv arXiv 2024

  65. [67]

    Dapo: An open-source llm reinforcement learning system at scale.Advances in Neural Information Processing Systems, 38:113222–113244, 2026

    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.Advances in Neural Information Processing Systems, 38:113222–113244, 2026

    2026

  66. [68]

    Your efficient rl framework secretly brings you off-policy rl training, 2025

    Feng Yao, Liyuan Liu, Dinghuai Zhang, Chengyu Dong, Jingbo Shang, and Jianfeng Gao. Your efficient rl framework secretly brings you off-policy rl training, 2025. URLhttps://fengyao.notion.site/off-policy-rl

    2025

  67. [69]

    Stabilizing moe reinforcement learning by aligning training and inference routers.arXiv preprint arXiv:2510.11370, 2025

    Wenhan Ma, Hailin Zhang, Liang Zhao, Yifan Song, Yudong Wang, Zhifang Sui, and Fuli Luo. Stabilizing moe reinforcement learning by aligning training and inference routers.arXiv preprint arXiv:2510.11370, 2025

    arXiv 2025

  68. [70]

    Skyrl gym generator tutorial

    SkyRL Team. Skyrl gym generator tutorial. https://docs.skyrl.ai/docs/tutorials/skyrl_gym_generator, 2025

    2025

  69. [71]

    Guifang Duan, Yunkun Xu, Zhenyu Liu, and Jianrong Tan. Hand-in-hand guidance: an explore-exploit based reinforcement learning method for performance driven assembly-adjustment.IEEE Transactions on Industrial Informatics, 19(10):10045– 10055, 2023

    2023

  70. [72]

    Save, load and learn: Boosting agentic llms via rollback-based curriculum learning

    Xiaoyang Li, Weixun Wang, and Yancheng He. Save, load and learn: Boosting agentic llms via rollback-based curriculum learning. https://warm-pajama-44a.notion.site/Save-Load-and-Learn-Boosting-Agentic-LLMs-via-Rol lback-based-Curriculum-Learning-687a76d7970e831a91c501bafd9c7b2b, 2026

    2026

  71. [73]

    Gemini 3.5 Flash model card

    Google DeepMind. Gemini 3.5 Flash model card. Model card, May 2026. URL https://deepmind.google/models/mod el-cards/gemini-3-5-flash/

    2026

  72. [74]

    GPT-5.5 system card

    OpenAI. GPT-5.5 system card. System card, April 2026. URLhttps://openai.com/index/gpt-5-5-system-card/

    2026

  73. [75]

    Claude Opus 4.7 model report

    Anthropic. Claude Opus 4.7 model report. Anthropic Transparency Hub, April 2026. URL https://www.anthropic.com/ transparency/model-report

    2026

  74. [76]

    Qwen3-235B-A22B-Thinking-2507

    Qwen Team. Qwen3-235B-A22B-Thinking-2507. Hugging Face model card, 2025. URL https://huggingface.co/Qwe n/Qwen3-235B-A22B-Thinking-2507

    2025

  75. [77]

    Qwen3.7: The agent frontier

    Qwen Team. Qwen3.7: The agent frontier. Qwen blog, May 2026. URLhttps://qwen.ai/blog?id=qwen3.7

    2026

  76. [78]

    DeepSeek-V4: Towards highly efficient million-token context intelligence, 2026

    DeepSeek-AI. DeepSeek-V4: Towards highly efficient million-token context intelligence, 2026. URL https://huggingfac e.co/deepseek-ai/DeepSeek-V4-Pro

    2026

  77. [79]

    Nex-N1: Agentic models trained via a unified ecosystem for large-scale environment construction.arXiv preprint arXiv:2512.04987, 2025

    Yuxuan Cai, Lu Chen, Qiaoling Chen, Yuyang Ding, Liwen Fan, Wenjie Fu, Yufei Gao, Honglin Guo, et al. Nex-N1: Agentic models trained via a unified ecosystem for large-scale environment construction.arXiv preprint arXiv:2512.04987, 2025. URL https://arxiv.org/abs/2512.04987

    arXiv 2025

  78. [80]

    Qwen3.5: Towards native multimodal agents, February 2026

    Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026. URLhttps://qwen.ai/blog?id=qwen3.5. 41

    2026

  79. [81]

    Qwen3.6-35B-A3B: Agentic coding power, now open to all, April 2026

    Qwen Team. Qwen3.6-35B-A3B: Agentic coding power, now open to all, April 2026. URL https://qwen.ai/blog?id= qwen3.6-35b-a3b

    2026

  80. [82]

    Large language models are zero-shot reasoners

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. Large language models are zero-shot reasoners. InAdvances in Neural Information Processing Systems, volume 35, pages 22199–22213, 2022

    2022

Showing first 80 references.