pith. sign in

arxiv: 2606.02372 · v1 · pith:5M7KFBF7new · submitted 2026-06-01 · 💻 cs.AI · cs.CL

COMAP: Co-Evolving World Models and Agent Policies for LLM Agents

Youwei Liu , Jian Wang , Hanlin Wang , Wenjie Li This is my paper

Pith reviewed 2026-06-28 14:25 UTC · model grok-4.3

classification 💻 cs.AI cs.CL
keywords LLM agentsworld modelsco-evolutionself-distillationtask planningweb navigationtool useembodied AI
0
0 comments X

The pith

Co-evolving world models and agent policies improves LLM performance without external rewards.

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

The paper presents COMAP as a way for language agents to improve their world models and decision policies together in a closed loop. At each step the world model forecasts outcomes for possible actions, the agent judges how trustworthy those forecasts are and adjusts its choice, and the resulting interaction data is fed back to refine the world model through self-distillation. This setup targets the mismatch between a static world model and an improving agent while removing the need for outside reward signals or verifiers. If the loop works as described, agents become better at long-horizon planning on embodied, web, and tool-use tasks by steadily increasing the accuracy of their internal predictions.

Core claim

COMAP co-evolves textual world models and agent policies by letting the agent perform future-aware reflection on the world model's predicted state feedback and then using the resulting on-policy trajectories to update the world model via self-distillation, producing measurable gains in prediction accuracy and task success across multiple benchmarks.

What carries the argument

The closed-loop interaction in which the agent's reliability estimates of world-model feedback generate on-policy trajectories that are distilled back into the world model.

If this is right

  • World-model prediction accuracy increases steadily across co-evolution iterations.
  • Agents produce more effective long-horizon decisions on embodied planning, web navigation, and tool-use tasks.
  • The framework delivers consistent outperformance over fixed-world-model baselines, including a 16.75 percent relative gain with the Qwen3-4B model.
  • No external reward signals or verifiers are required for the improvement cycle to operate.

Where Pith is reading between the lines

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

  • The same loop could be tested in environments where ground-truth rewards are unavailable or expensive to obtain.
  • Extending the self-distillation step to larger models might reduce reliance on human-curated data for agent training.
  • The reliability-estimation component could be studied separately to see how its accuracy affects overall loop stability.

Load-bearing premise

The agent's own judgments about whether the world model's feedback is reliable are accurate enough to produce useful training data without any external verification.

What would settle it

Running multiple rounds of the co-evolution loop on one of the reported benchmarks and observing that world-model prediction accuracy does not rise while agent success rate stays flat or falls.

Figures

Figures reproduced from arXiv: 2606.02372 by Hanlin Wang, Jian Wang, Wenjie Li, Youwei Liu.

Figure 1
Figure 1. Figure 1: Conceptual illustration of the co-evolution of [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed COMAP framework. The textual world model and the agent policy co-evolve through three key mechanisms: on-policy self-distillation of the world model, future-aware policy reflection, and dynamic interaction between the model and the policy. back to the agent policy, yielding a foresight￾augmented trajectory h + t = (ht , adraft t , sˆt+1) for the final policy decision. 3 Method We p… view at source ↗
Figure 3
Figure 3. Figure 3: Component ablations of COMAP on ALFWorld. We report leave-one-component-out results on Qwen3-4B and Qwen3-8B. Policy-side results represent success rates, while world-model-side results represent Delta-F1 scores. smaller backbones, where future-conditioned re￾flection can compensate for weaker planning ability. Compared with prompting-only, test-time evolv￾ing, and fixed world-modeling methods, COMAP furth… view at source ↗
Figure 6
Figure 6. Figure 6: Change-token NLL under off-policy SFT and [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: Adoption ratio of the world state gate (Qwen3- [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Benchmark-specific rollout training-data distributions. Each panel reports the composition of agent-policy [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Model size scaling of COMAP. We evaluate COMAP with Qwen3-4B, Qwen3-8B, and Qwen3-30B-A3B backbones on ALFWorld. The left panel reports the average agent-policy performance, while the right panel reports the average world-model Delta-F1. The marker size indicates the relative backbone scale. C.2 Additional Clarification of COMAP Mechanisms This section provides additional details on the deci￾sion flow, Fut… view at source ↗
Figure 9
Figure 9. Figure 9: Prompt template for draft action generation in the agent policy. The policy proposes a candidate action [PITH_FULL_IMAGE:figures/full_fig_p017_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Prompt template for reflection with future state in the agent policy. The policy uses the future-state signal [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Prompt template for the one-step textual world model. Given the current textual state and an action, the [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
read the original abstract

Equipping language agents with world models enables them to anticipate environment dynamics and evaluate candidate actions before execution. However, existing textual world models are typically fixed after training, preventing them from adapting to the on-policy state-action distributions induced by an evolving agent. Meanwhile, agent-improvement methods often rely on external rewards or verifiers, limiting their applicability in realistic interactive environments. In this paper, we propose COMAP, a novel framework that co-evolves textual world models and agent policies through closed-loop interaction. At each decision step, the world model predicts future state feedback for candidate actions, and the agent performs future-aware reflection by estimating the reliability of this feedback and refining its action accordingly. The resulting on-policy trajectories are then used to update the world model via self-distillation, allowing it to better match the agent's evolving interaction distribution. Across embodied task planning, Web navigation, and tool-use benchmarks, COMAP consistently outperforms competitive baselines, e.g., +16.75% relative improvement with Qwen3-4B. Further analyses show that the co-evolutionary loop improves the world model's prediction accuracy over time and leads to more effective long-horizon decision-making. Our code is available at: https://github.com/loyiv/CoMAP.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces COMAP, a framework for co-evolving textual world models (WMs) and LLM agent policies in closed-loop interaction without external rewards or verifiers. At each step the WM predicts future state feedback for candidate actions; the agent performs future-aware reflection by estimating the reliability of that feedback and refining its action; the resulting on-policy trajectories are then used to update the WM via self-distillation so that the WM better matches the agent's evolving distribution. Experiments on embodied task planning, web navigation, and tool-use benchmarks report consistent gains (e.g., +16.75% relative with Qwen3-4B) and show improving WM prediction accuracy and long-horizon decision-making over iterations.

Significance. If the co-evolutionary loop can be shown to produce calibrated reliability estimates and genuine on-policy improvement without external signals, the result would be significant for autonomous agent training in realistic interactive environments where fixed WMs and external verifiers are impractical.

major comments (2)
  1. [Abstract / method] Abstract and method description: the central claim that the closed loop improves both WM accuracy and policy over iterations rests on the agent's internal reliability estimates being sufficiently calibrated to yield useful self-distillation targets. No analysis, calibration plots, or external verification of these estimates is provided, leaving open the possibility that early errors are amplified rather than corrected.
  2. [Abstract] Abstract: the reported gains (e.g., +16.75% relative) are attributed to the co-evolution loop, yet the manuscript supplies no ablations that isolate the contribution of the reliability-estimation + self-distillation step from other factors such as additional trajectory collection or prompt changes.
minor comments (2)
  1. [Abstract] The abstract states that code is available at a GitHub link, but no details on reproducibility (hyperparameters, exact prompts, or evaluation protocols) are given in the text.
  2. [Method] Notation for the reliability estimation step is introduced without a formal definition or pseudocode, making it difficult to assess how the estimate is computed from the WM output.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and commit to revisions that strengthen the manuscript's claims regarding calibration and component contributions.

read point-by-point responses

  1. Referee: [Abstract / method] Abstract and method description: the central claim that the closed loop improves both WM accuracy and policy over iterations rests on the agent's internal reliability estimates being sufficiently calibrated to yield useful self-distillation targets. No analysis, calibration plots, or external verification of these estimates is provided, leaving open the possibility that early errors are amplified rather than corrected.

    Authors: We agree that the manuscript would benefit from explicit analysis of the reliability estimates' calibration. While the current version reports that the co-evolutionary loop improves world model prediction accuracy over iterations (providing indirect support that self-distillation targets are useful), we do not include calibration plots or external verification. In the revision we will add these analyses, including plots of predicted vs. observed reliability and comparisons against external signals where available, to rule out error amplification. revision: yes

  2. Referee: [Abstract] Abstract: the reported gains (e.g., +16.75% relative) are attributed to the co-evolution loop, yet the manuscript supplies no ablations that isolate the contribution of the reliability-estimation + self-distillation step from other factors such as additional trajectory collection or prompt changes.

    Authors: We concur that dedicated ablations are needed to isolate the reliability-estimation and self-distillation components. The current experiments demonstrate overall gains and improving WM accuracy, but do not control for trajectory volume or prompt variations. We will add these ablations in the revised manuscript, comparing the full framework against controlled variants that disable the reliability step or self-distillation while matching other factors. revision: yes

Circularity Check

0 steps flagged

No circularity: standard iterative self-training loop with external benchmark validation

full rationale

The COMAP framework describes an iterative process of world-model prediction, agent reliability estimation for action refinement, trajectory collection, and self-distillation updates. This is a procedural training loop, not a derivation or prediction that reduces to fitted inputs by construction. No equations, uniqueness theorems, ansatzes, or self-citations are invoked in the abstract or description that would create self-definitional or load-bearing circularity. Performance gains are reported as empirical results on external benchmarks (embodied planning, web navigation, tool-use), which are independently falsifiable. The central claim therefore remains self-contained against external evaluation rather than tautological.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, new entities, or formal axioms are stated. The framework implicitly relies on the assumption that LLMs can perform reliable self-assessment of prediction quality.

axioms (1)
  • domain assumption LLM agents can produce sufficiently accurate estimates of the reliability of world-model feedback to generate useful training trajectories.
    The method depends on the agent performing 'future-aware reflection' by estimating reliability without external signals.

pith-pipeline@v0.9.1-grok · 5754 in / 1230 out tokens · 22885 ms · 2026-06-28T14:25:49.477835+00:00 · methodology

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

65 extracted references · 10 canonical work pages

  1. [1]

    Advances in neural information processing systems , volume=

    Language models meet world models: Embodied experiences enhance language models , author=. Advances in neural information processing systems , volume=

  2. [2]

    and Le, Quoc V

    Wei, Jason and Wang, Xuezhi and Schuurmans, Dale and Bosma, Maarten and Ichter, Brian and Xia, Fei and Chi, Ed H. and Le, Quoc V. and Zhou, Denny , title =. Advances in Neural Information Processing Systems , editor =. 2022 , url =

    2022

  3. [3]

    Mohit Shridhar and Jesse Thomason and Daniel Gordon and Yonatan Bisk and Winson Han and Roozbeh Mottaghi and Luke Zettlemoyer and Dieter Fox , bibsource =. 2020. doi:10.1109/CVPR42600.2020.01075 , pages =

    work page doi:10.1109/cvpr42600.2020.01075 2020

  4. [4]

    S cience W orld: Is your Agent Smarter than a 5th Grader?

    Wang, Ruoyao and Jansen, Peter and C \^o t \'e , Marc-Alexandre and Ammanabrolu, Prithviraj. S cience W orld: Is your Agent Smarter than a 5th Grader?. EMNLP. 2022

    2022

  5. [5]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 Technical Report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  6. [6]

    arXiv preprint arXiv:2412.15115 , year =

    Qwen2.5 Technical Report , author =. arXiv preprint arXiv:2412.15115 , year =

    Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2407.10671 , year =

    Qwen2 Technical Report , author =. arXiv preprint arXiv:2407.10671 , year =

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2407.21783 , year=

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    Pith/arXiv arXiv

  9. [9]

    2023 , html =

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle =. 2023 , html =

    2023

  10. [10]

    ST e C a: Step-level Trajectory Calibration for LLM Agent Learning

    Wang, Hanlin and Wang, Jian and Leong, Chak Tou and Li, Wenjie. ST e C a: Step-level Trajectory Calibration for LLM Agent Learning. Findings of the Association for Computational Linguistics: ACL 2025. 2025. doi:10.18653/v1/2025.findings-acl.604

    work page doi:10.18653/v1/2025.findings-acl.604 2025

  11. [11]

    Tianzhe Chu and Yuexiang Zhai and Jihan Yang and Shengbang Tong and Saining Xie and Dale Schuurmans and Quoc V Le and Sergey Levine and Yi Ma , booktitle=

  12. [12]

    Trial and Error: Exploration-Based Trajectory Optimization of LLM Agents

    Song, Yifan and Yin, Da and Yue, Xiang and Huang, Jie and Li, Sujian and Lin, Bill Yuchen. Trial and Error: Exploration-Based Trajectory Optimization of LLM Agents. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/v1/2024.acl-long.409

    work page doi:10.18653/v1/2024.acl-long.409 2024

  13. [13]

    Watch Every Step! LLM Agent Learning via Iterative Step-level Process Refinement

    Xiong, Weimin and Song, Yifan and Zhao, Xiutian and Wu, Wenhao and Wang, Xun and Wang, Ke and Li, Cheng and Peng, Wei and Li, Sujian. Watch Every Step! LLM Agent Learning via Iterative Step-level Process Refinement. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024. doi:10.18653/v1/2024.emnlp-main.93

    work page doi:10.18653/v1/2024.emnlp-main.93 2024

  14. [14]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Virtualhome: Simulating household activities via programs , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  15. [15]

    Advances in Neural Information Processing Systems , volume=

    Camel: Communicative agents for" mind" exploration of large language model society , author=. Advances in Neural Information Processing Systems , volume=

  16. [16]

    Transactions on Machine Learning Research , year=

    Voyager: An Open-Ended Embodied Agent with Large Language Models , author=. Transactions on Machine Learning Research , year=

  17. [17]

    Advances in Neural Information Processing Systems , volume=

    Reflexion: Language agents with verbal reinforcement learning , author=. Advances in Neural Information Processing Systems , volume=

  18. [18]

    Advances in Neural Information Processing Systems , volume=

    On the planning abilities of large language models-a critical investigation , author=. Advances in Neural Information Processing Systems , volume=

  19. [19]

    arXiv preprint arXiv:2301.04104 , year=

    Mastering diverse domains through world models , author=. arXiv preprint arXiv:2301.04104 , year=

    Pith/arXiv arXiv

  20. [20]

    World Modeling Makes a Better Planner: Dual Preference Optimization for Embodied Task Planning

    Wang, Siyin and Fei, Zhaoye and Cheng, Qinyuan and Zhang, Shiduo and Cai, Panpan and Fu, Jinlan and Qiu, Xipeng. World Modeling Makes a Better Planner: Dual Preference Optimization for Embodied Task Planning. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.acl-long.1044

    work page doi:10.18653/v1/2025.acl-long.1044 2025

  21. [21]

    Reasoning with language model is planning with world model

    Hao, Shibo and Gu, Yi and Ma, Haodi and Hong, Joshua and Wang, Zhen and Wang, Daisy and Hu, Zhiting. Reasoning with Language Model is Planning with World Model. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2023.emnlp-main.507

    work page doi:10.18653/v1/2023.emnlp-main.507 2023

  22. [22]

    2, 2022-06-27 , author=

    A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27 , author=. Open Review , volume=

    2022

  23. [23]

    arXiv preprint arXiv:2506.22355 , year=

    Embodied AI Agents: Modeling the World , author=. arXiv preprint arXiv:2506.22355 , year=

    arXiv

  24. [24]

    W eb E volver: Enhancing Web Agent Self-Improvement with Co-evolving World Model

    Fang, Tianqing and Zhang, Hongming and Zhang, Zhisong and Ma, Kaixin and Yu, Wenhao and Mi, Haitao and Yu, Dong. W eb E volver: Enhancing Web Agent Self-Improvement with Co-evolving World Model. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.emnlp-main.454

    work page doi:10.18653/v1/2025.emnlp-main.454 2025

  25. [25]

    arXiv preprint arXiv:2510.08558 , year=

    Agent learning via early experience , author=. arXiv preprint arXiv:2510.08558 , year=

    Pith/arXiv arXiv

  26. [26]

    Advances in Neural Information Processing Systems , volume=

    Agent planning with world knowledge model , author=. Advances in Neural Information Processing Systems , volume=

  27. [27]

    E ^2 CL : Exploration-based Error Correction Learning for Embodied Agents

    Wang, Hanlin and Leong, Chak Tou and Wang, Jian and Li, Wenjie. E ^2 CL : Exploration-based Error Correction Learning for Embodied Agents. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.findings-emnlp.448

    work page doi:10.18653/v1/2024.findings-emnlp.448 2024

  28. [28]

    Advances in Neural Information Processing Systems , volume=

    Leveraging pre-trained large language models to construct and utilize world models for model-based task planning , author=. Advances in Neural Information Processing Systems , volume=

  29. [29]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    VAGEN: Reinforcing World Model Reasoning for Multi-Turn VLM Agents , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  30. [30]

    arXiv preprint arXiv:2510.15047 , year=

    Internalizing World Models via Self-Play Finetuning for Agentic RL , author=. arXiv preprint arXiv:2510.15047 , year=

    arXiv

  31. [31]

    arXiv preprint arXiv:2510.18135 , year=

    World-in-world: World models in a closed-loop world , author=. arXiv preprint arXiv:2510.18135 , year=

    arXiv

  32. [32]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    DMWM: Dual-Mind World Model with Long-Term Imagination , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  33. [33]

    Journal of Mathematical Analysis and Applications , volume=

    Optimal control of Markov processes with incomplete state information I , author=. Journal of Mathematical Analysis and Applications , volume=. 1965 , publisher=

    1965

  34. [34]

    arXiv preprint arXiv:2501.11425 , year=

    Agent-R: Training Language Model Agents to Reflect via Iterative Self-Training , author=. arXiv preprint arXiv:2501.11425 , year=

    arXiv

  35. [35]

    The Thirteenth International Conference on Learning Representations (ICLR) , year=

    AgentRefine: Enhancing Agent Generalization through Refinement Tuning , author=. The Thirteenth International Conference on Learning Representations (ICLR) , year=

  36. [36]

    arXiv preprint arXiv:2512.02556 , year =

    DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models , author =. arXiv preprint arXiv:2512.02556 , year =

    Pith/arXiv arXiv

  37. [37]

    arXiv preprint arXiv:2512.18832 , year=

    From Word to World: Can Large Language Models be Implicit Text-based World Models? , author=. arXiv preprint arXiv:2512.18832 , year=

    arXiv

  38. [38]

    Advances in Neural Information Processing Systems , volume =

    Shunyu Yao and Howard Chen and John Yang and Karthik Narasimhan , title =. Advances in Neural Information Processing Systems , volume =

  39. [39]

    arXiv preprint arXiv:2505.10978 , year=

    Group-in-group policy optimization for llm agent training , author=. arXiv preprint arXiv:2505.10978 , year=

    Pith/arXiv arXiv

  40. [40]

    arXiv preprint arXiv:2505.20732 , year=

    SPA-RL: Reinforcing LLM Agents via Stepwise Progress Attribution , author=. arXiv preprint arXiv:2505.20732 , year=

    arXiv

  41. [41]

    International conference on machine learning , pages=

    Asynchronous methods for deep reinforcement learning , author=. International conference on machine learning , pages=. 2016 , organization=

    2016

  42. [42]

    arXiv preprint arXiv:2310.05915 , year=

    Fireact: Toward language agent fine-tuning , author=. arXiv preprint arXiv:2310.05915 , year=

    arXiv

  43. [43]

    S table T ool B ench: Towards Stable Large-Scale Benchmarking on Tool Learning of Large Language Models

    Guo, Zhicheng and Cheng, Sijie and Wang, Hao and Liang, Shihao and Qin, Yujia and Li, Peng and Liu, Zhiyuan and Sun, Maosong and Liu, Yang. S table T ool B ench: Towards Stable Large-Scale Benchmarking on Tool Learning of Large Language Models. Findings of the Association for Computational Linguistics: ACL 2024. 2024. doi:10.18653/v1/2024.findings-acl.664

    work page doi:10.18653/v1/2024.findings-acl.664 2024

  44. [44]

    Griffiths and Yuan Cao and Karthik Narasimhan , title =

    Shunyu Yao and Dian Yu and Jeffrey Zhao and Izhak Shafran and Thomas L. Griffiths and Yuan Cao and Karthik Narasimhan , title =. Advances in Neural Information Processing Systems , editor =

  45. [45]

    Bandit Based Monte-Carlo Planning

    Levente Kocsis and Csaba Szepesv. Bandit Based Monte-Carlo Planning , booktitle =. doi:10.1007/11871842_29 , year =

    work page doi:10.1007/11871842_29

  46. [46]

    arXiv preprint arXiv:2601.08955 , year =

    Imagine-then-Plan: Agent Learning from Adaptive Lookahead with World Models , author =. arXiv preprint arXiv:2601.08955 , year =

    arXiv

  47. [47]

    2025 , eprint=

    From Word to World: Can Large Language Models be Implicit Text-based World Models? , author=. 2025 , eprint=

    2025

  48. [48]

    Frontiers of Computer Science , volume=

    A survey on large language model based autonomous agents , author=. Frontiers of Computer Science , volume=. 2024 , publisher=

    2024

  49. [49]

    Advances in Neural Information Processing Systems , volume=

    Embodied agent interface: Benchmarking llms for embodied decision making , author=. Advances in Neural Information Processing Systems , volume=

  50. [50]

    International Conference on Learning Representations , volume=

    Web agents with world models: Learning and leveraging environment dynamics in web navigation , author=. International Conference on Learning Representations , volume=

  51. [51]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    Enhancing agent learning through world dynamics modeling , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    2024

  52. [52]

    2026 , month = mar, day =

    Introducing GPT-5.4 , author =. 2026 , month = mar, day =

    2026

  53. [53]

    arXiv preprint arXiv:2604.00626 , year=

    A Survey of On-Policy Distillation for Large Language Models , author=. arXiv preprint arXiv:2604.00626 , year=

    Pith/arXiv arXiv

  54. [54]

    arXiv preprint arXiv:2602.10885 , year =

    Reinforcing Chain-of-Thought Reasoning with Self-Evolving Rubrics , author =. arXiv preprint arXiv:2602.10885 , year =

    arXiv

  55. [55]

    2026 , howpublished =

    DeepSeek-V4 Preview Release , author =. 2026 , howpublished =

    2026

  56. [56]

    2310.10134 , archivePrefix =

    Majumder, Bodhisattwa Prasad and Mishra, Bhavana Dalvi and Jansen, Peter and Tafjord, Oyvind and Tandon, Niket and Zhang, Li and Callison-Burch, Chris and Clark, Peter , year =. 2310.10134 , archivePrefix =

    arXiv

  57. [57]

    2024 , eprint =

    Investigate-Consolidate-Exploit: A General Strategy for Inter-Task Agent Self-Evolution , author =. 2024 , eprint =

    2024

  58. [58]

    2406.04151 , archivePrefix =

    Xi, Zhiheng and Ding, Yiwen and Chen, Wenxiang and Hong, Boyang and Guo, Honglin and Wang, Junzhe and Yang, Dingwen and Liao, Chenyang and Guo, Xin and He, Wei and Gao, Songyang and Chen, Lu and Zheng, Rui and Zou, Yicheng and Gui, Tao and Zhang, Qi and Qiu, Xipeng and Huang, Xuanjing and Wu, Zuxuan and Jiang, Yu-Gang , year =. 2406.04151 , archivePrefix =

    arXiv

  59. [59]

    2310.00533 , archivePrefix =

    Lu, Jianqiao and Zhong, Wanjun and Huang, Wenyong and Wang, Yufei and Zhu, Qi and Mi, Fei and Wang, Baojun and Wang, Weichao and Zeng, Xingshan and Shang, Lifeng and Jiang, Xin and Liu, Qun , year =. 2310.00533 , archivePrefix =

    arXiv

  60. [60]

    2024 , eprint =

    Symbolic Learning Enables Self-Evolving Agents , author =. 2024 , eprint =

    2024

  61. [61]

    2024 , eprint =

    Automated Design of Agentic Systems , author =. 2024 , eprint =

    2024

  62. [62]

    2025 , url =

    Yuan, Siyu and Song, Kaitao and Chen, Jiangjie and Tan, Xu and Li, Dongsheng and Yang, Deqing , booktitle =. 2025 , url =

    2025

  63. [63]

    Co-Evolving

    Wu, Xiyang and Li, Zongxia and Shi, Guangyao and Duffy, Alexander and Marques, Tyler and Olson, Matthew Lyle and Zhou, Tianyi and Manocha, Dinesh , year =. Co-Evolving. 2604.20987 , archivePrefix =

    Pith/arXiv arXiv

  64. [64]

    , year =

    Xu, Wujiang and Han, Jiaojiao and Guo, Minghao and Mei, Kai and Zhu, Xi and Zhang, Han and Metaxas, Dimitris N. , year =. 2604.21725 , archivePrefix =

    Pith/arXiv arXiv

  65. [65]

    2021 , url =

    Mohit Shridhar and Xingdi Yuan and Marc-Alexandre C\^ot\'e and Yonatan Bisk and Adam Trischler and Matthew Hausknecht , booktitle =. 2021 , url =

    2021