pith. sign in

arxiv: 2606.30639 · v1 · pith:VK7HDGWCnew · submitted 2026-06-29 · 💻 cs.AI · cs.CL

Self-Evolving World Models for LLM Agent Planning

Xuan Zhang , Wenxuan Zhang , See-Kiong Ng , Yang Deng This is my paper

Pith reviewed 2026-06-30 05:46 UTC · model grok-4.3

classification 💻 cs.AI cs.CL
keywords world modelsLLM agentsagent planningepisodic memorysemantic memorytest-time revisionforesight
0
0 comments X

The pith

A self-evolving world model revises its deployment-time context through memory modules to raise prediction accuracy and LLM agent planning success without retraining.

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

The paper tries to establish that world models can become more reliable for long-horizon LLM agents by revising their own context at test time instead of relying on static predictions. WorldEvolver adds episodic memory drawn from real action transitions, semantic memory that turns prediction mismatches into heuristic rules, and selective foresight that drops low-confidence outputs before they reach the agent. If this holds, agents would receive usable foresight that improves both the accuracy of action predictions and the success rate of downstream plans. A sympathetic reader would care because unreliable world models often get ignored or cause worse decisions, and this method offers adaptation without changing the agent or any model parameters.

Core claim

WorldEvolver integrates three modules while keeping the downstream agent and all model parameters frozen: Episodic Memory exploits real action transitions through retrieval-based simulation, Semantic Memory extracts persistent heuristic rules from prediction-observation mismatches, and Selective Foresight filters low-confidence predictions before they enter the agent reasoning context. Experiments on ALFWorld and ScienceWorld, evaluated via Word2World for prediction accuracy and AgentBoard for agent success rate, show that WorldEvolver achieves the highest prediction accuracy across three backbones and leads other world model baselines on downstream agent success rate.

What carries the argument

WorldEvolver framework with its Episodic Memory, Semantic Memory, and Selective Foresight modules that enable test-time context revision.

If this is right

  • World model prediction accuracy reaches the highest levels across three different LLM backbones on Word2World.
  • Downstream agent success rate exceeds other world model baselines on AgentBoard tasks in ALFWorld and ScienceWorld.
  • Test-time memory revision improves both predictive fidelity and planning performance while parameters stay frozen.
  • The approach outperforms static world models by incorporating real transitions and error-derived rules selectively.

Where Pith is reading between the lines

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

  • The separation of world model revision from agent reasoning could reduce context overload in longer-horizon tasks.
  • This test-time approach might extend to environments where new observations arrive continuously without requiring full model updates.
  • If the modules scale, agents could maintain performance in changing settings by evolving only the memory components.

Load-bearing premise

The three modules integrate into the agent's reasoning context without causing the agent to ignore or misuse the updated information, and the benchmarks represent real planning challenges.

What would settle it

If WorldEvolver produces no improvement or lower prediction accuracy on Word2World and no gain in agent success rate on AgentBoard compared to non-evolving world model baselines across the same backbones.

Figures

Figures reproduced from arXiv: 2606.30639 by See-kiong Ng, Wenxuan Zhang, Xuan Zhang, Yang Deng.

Figure 1
Figure 1. Figure 1: Contrast of different world models. Frozen (a) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Preliminary oracle study with Gemma-4-26B [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of WORLDEVOLVER. A frozen world model produces action-conditioned predictions using Episodic Memory for exploitation through retrieval-based simulation over previous action transitions and Semantic Memory for exploration through persistent heuristic-rule discovery from prediction-observation mismatches. Selec￾tive Foresight filters the prediction before it conditions the frozen agent. candidate ac… view at source ↗
Figure 4
Figure 4. Figure 4: Cumulative best-of-L Agent Planning Success Rate on ALFWorld and ScienceWorld. -2 -1 0 Δ EM (%) -25 -15 -5 0 -2 -1 0 Δ EM (%) -25 -15 -5 0 1 2 3 4 5 kMS -2 -1 0 Δ EM (%) 1 2 3 4 5 kME -25 -15 -5 0 Gemma-4-26B Qwen3.5-9B Gemma-4-31B ALFWorld ScienceWorld [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Relative gains from memory hyperparameters [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Selective foresight confidence calibration [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Trajectory-level world model prediction Exact Match (%) along the Word2World deployment order. We [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Agent Planning Success Rate (%) by Agent [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Heatmap of agent planning success rates (%) for GPT-5.4-mini across different world models, reported as [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Heatmap of agent planning success rates (%) for Gemma-4-26B-A4B across different world models, [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Selective foresight confidence calibration, measured as Exact Match on the top-confidence prefix. [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: ReAct agent prompt for ALFWorld, with placeholders for available actions and the response format. [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: ReflAct agent prompt for ALFWorld, replacing the ReAct [PITH_FULL_IMAGE:figures/full_fig_p016_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: ReAct agent prompt for ScienceWorld, enumerating six command groups (Manipulation, Inspection, [PITH_FULL_IMAGE:figures/full_fig_p017_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: ReflAct agent prompt for ScienceWorld, inheriting the six ScienceWorld command groups. [PITH_FULL_IMAGE:figures/full_fig_p017_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Zero-Shot world model prompt. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: RAWM-ϕ (Yang et al., 2025) world model prompt, augmented with top-k transitions retrieved by cosine similarity over Qwen3-Embedding-8B embeddings of the live (st, at) query against the fixed Word2World corpus. Prompt Template for ITP-I World Model You are a world model for the {env_name} environment. Given an action/observation history, imagine the next few steps, describing likely observations and key ob… view at source ↗
Figure 18
Figure 18. Figure 18: ITP-I world model prompt following the Imagine-then-Plan ( [PITH_FULL_IMAGE:figures/full_fig_p018_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Episodic Memory world model prompt, augmented with top- [PITH_FULL_IMAGE:figures/full_fig_p018_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Semantic Memory world model prompt, grounded by mismatch-derived persistence rules from [PITH_FULL_IMAGE:figures/full_fig_p019_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: WORLDEVOLVER world model prompt combining Episodic and Semantic Memory grounding blocks, with foresight-based confidence filtering applied post-generation without modifying the prompt. Prompt Template for Observation Factorizer System message: You extract compact world-state triples from text observations. Return only valid JSON; do not explain. User message: Extract (subject, predicate, object) triples t… view at source ↗
Figure 22
Figure 22. Figure 22: Observation factorizer prompt, converting predicted and gold observations into factorized triples whose [PITH_FULL_IMAGE:figures/full_fig_p019_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Preservation-rule extractor prompt, processing a batch of [PITH_FULL_IMAGE:figures/full_fig_p020_23.png] view at source ↗
read the original abstract

World models offer a principled way to equip long-horizon LLM agents with foresight: predictions of action consequences before execution. However, unreliable foresight can be ignored, misused, or even degrade downstream decision-making. In this paper, we introduce WorldEvolver, a self-evolving world model framework that revises its deployment-time context while keeping the downstream agent and all model parameters frozen. WorldEvolver integrates three modules: (i) Episodic Memory, which exploits real action transitions through retrieval-based simulation; (ii) Semantic Memory, which extracts persistent heuristic rules from prediction-observation mismatches; and (iii) Selective Foresight, which filters low-confidence predictions before integrating them into agent reasoning context. We evaluate WorldEvolver on ALFWorld and ScienceWorld, measuring world model prediction accuracy on Word2World and downstream agent success rate on AgentBoard. Extensive experiments show that WorldEvolver achieves the highest prediction accuracy across three backbones and leads other world model baselines on downstream agent success rate, demonstrating that test-time memory revision enhances both predictive fidelity and planning performance.

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

3 major / 2 minor

Summary. The paper introduces WorldEvolver, a self-evolving world model for LLM agents consisting of Episodic Memory (retrieval-based simulation of action transitions), Semantic Memory (extraction of heuristic rules from prediction-observation mismatches), and Selective Foresight (filtering low-confidence predictions). With the downstream agent and all parameters kept frozen, the framework revises deployment-time context and is evaluated on ALFWorld and ScienceWorld using Word2World for prediction accuracy and AgentBoard for agent success rate. The central claim is that this test-time memory revision yields the highest prediction accuracy across three backbones and superior downstream planning performance relative to other world-model baselines.

Significance. If the results hold, the work would demonstrate a practical mechanism for improving foresight reliability in long-horizon LLM agents without retraining. Credit is due for grounding the evaluation in the standard external benchmarks ALFWorld and ScienceWorld, testing across multiple backbones, and focusing on the interaction between world-model revision and frozen-agent planning.

major comments (3)
  1. [Experiments / Evaluation] The central claim that inserting the outputs of the three modules into the agent's context produces the observed gains requires evidence that the agent actually conditions on the revised memory. No ablation that removes the memory content while preserving other prompt changes, and no analysis (e.g., attention maps or forced-reference prompts) showing the agent attends to the inserted information, is described.
  2. [Abstract / Experiments] The abstract states that WorldEvolver 'achieves the highest prediction accuracy across three backbones and leads other world model baselines on downstream agent success rate,' yet the provided text supplies no quantitative numbers, error bars, statistical significance tests, or details on baseline implementations and hyper-parameters, making it impossible to verify the magnitude or robustness of the reported improvements.
  3. [Method / Evaluation] The assumption that the three modules can be integrated into the agent's reasoning context without causing the agent to ignore or misuse the updated information is load-bearing for the performance claims, but the manuscript provides no diagnostic experiments or failure-case analysis addressing this integration risk.
minor comments (2)
  1. [Abstract] The abstract references 'Word2World' and 'AgentBoard' without defining them or citing their sources.
  2. [Method] Notation for the three modules (Episodic Memory, Semantic Memory, Selective Foresight) is introduced without an accompanying diagram or pseudocode showing data flow and insertion points into the frozen agent's prompt.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback emphasizing the need for direct evidence of agent conditioning on memory and clearer quantitative reporting. We address each major comment below and will incorporate revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Experiments / Evaluation] The central claim that inserting the outputs of the three modules into the agent's context produces the observed gains requires evidence that the agent actually conditions on the revised memory. No ablation that removes the memory content while preserving other prompt changes, and no analysis (e.g., attention maps or forced-reference prompts) showing the agent attends to the inserted information, is described.

    Authors: We agree that isolating the contribution of the memory content itself is important for validating the central claim. In the revised version, we will add an ablation that substitutes neutral placeholder text for the actual memory outputs while preserving prompt length, structure, and all other elements. We will also include qualitative analysis of agent reasoning traces demonstrating explicit references to the inserted memory content. revision: yes

  2. Referee: [Abstract / Experiments] The abstract states that WorldEvolver 'achieves the highest prediction accuracy across three backbones and leads other world model baselines on downstream agent success rate,' yet the provided text supplies no quantitative numbers, error bars, statistical significance tests, or details on baseline implementations and hyper-parameters, making it impossible to verify the magnitude or robustness of the reported improvements.

    Authors: The results section contains tables reporting quantitative metrics with means and standard deviations across multiple runs, along with baseline details. To address the concern directly, we will revise the abstract to include key numerical improvements and ensure the main text and appendix explicitly list hyper-parameters, baseline implementations, and any statistical tests performed. revision: yes

  3. Referee: [Method / Evaluation] The assumption that the three modules can be integrated into the agent's reasoning context without causing the agent to ignore or misuse the updated information is load-bearing for the performance claims, but the manuscript provides no diagnostic experiments or failure-case analysis addressing this integration risk.

    Authors: We acknowledge that the integration assumption requires explicit support. We will add a dedicated subsection with diagnostic metrics (e.g., frequency of memory references in generated plans) and a failure-case analysis covering instances of ignoring or misusing the memory, including examples and discussion of observed patterns. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical framework on external benchmarks

full rationale

The paper introduces WorldEvolver as an empirical framework with three described modules (Episodic Memory, Semantic Memory, Selective Foresight) whose outputs are inserted into a frozen agent's context. Claims rest on prediction accuracy and success-rate measurements on the independent external benchmarks ALFWorld, ScienceWorld, Word2World and AgentBoard. No equations, fitted parameters renamed as predictions, self-definitional relations, or load-bearing self-citations appear in the provided text; the derivation chain is therefore self-contained against external data.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 3 invented entities

The central claim rests on the effectiveness of these three newly introduced modules. No free parameters are mentioned. Axioms are standard assumptions in the field of LLM agents. Since only the abstract is available, the ledger is based on high-level descriptions.

axioms (2)
  • domain assumption World models are useful for equipping LLM agents with foresight in long-horizon tasks.
    Opening statement of the abstract.
  • domain assumption Real action transitions and prediction-observation mismatches provide useful information for improving the world model.
    Basis for the episodic and semantic memory modules.
invented entities (3)
  • Episodic Memory no independent evidence
    purpose: Exploits real action transitions through retrieval-based simulation
    New module introduced in the framework.
  • Semantic Memory no independent evidence
    purpose: Extracts persistent heuristic rules from prediction-observation mismatches
    New module for extracting rules.
  • Selective Foresight no independent evidence
    purpose: Filters low-confidence predictions before integrating them into agent reasoning context
    New filtering component.

pith-pipeline@v0.9.1-grok · 5714 in / 1511 out tokens · 42991 ms · 2026-06-30T05:46:01.757159+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

63 extracted references · 4 canonical work pages · 1 internal anchor

  1. [1]

    2023 , url =

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

    2023

  2. [2]

    OpenAI Blog , year =

    Language Models are Unsupervised Multitask Learners , author =. OpenAI Blog , year =

  3. [3]

    arXiv preprint arXiv:1803.10122 , year =

    World Models , author =. arXiv preprint arXiv:1803.10122 , year =

    Pith/arXiv arXiv

  4. [4]

    Nature , volume =

    Mastering diverse control tasks through world models , author =. Nature , volume =. 2025 , doi =

    2025

  5. [5]

    2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages =

    Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World , author =. 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages =. 2017 , publisher =

    2017

  6. [6]

    Transactions on Machine Learning Research (TMLR) , year =

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

  7. [7]

    and Stoica, Ion and Gonzalez, Joseph E

    Packer, Charles and Wooders, Sarah and Lin, Kevin and Fang, Vivian and Patil, Shishir G. and Stoica, Ion and Gonzalez, Joseph E. , journal =. 2023 , url =

    2023

  8. [8]

    Transactions on Machine Learning Research , year =

    Cognitive Architectures for Language Agents , author =. Transactions on Machine Learning Research , year =

  9. [9]

    2025 , url =

    Kim, Jeonghye and Rhee, Sojeong and Kim, Minbeom and Kim, Dohyung and Lee, Sangmook and Sung, Youngchul and Jung, Kyomin , booktitle =. 2025 , url =

    2025

  10. [10]

    Proceedings of the 34th International Conference on Machine Learning (ICML) , year =

    Neural Episodic Control , author =. Proceedings of the 34th International Conference on Machine Learning (ICML) , year =

  11. [11]

    arXiv preprint arXiv:1606.04460 , year =

    Model-Free Episodic Control , author =. arXiv preprint arXiv:1606.04460 , year =

    Pith/arXiv arXiv

  12. [12]

    , journal =

    Kumaran, Dharshan and Hassabis, Demis and McClelland, James L. , journal =. What Learning Systems do Intelligent Agents Need?

  13. [13]

    Efficient Integration of External Knowledge to

    Yang, Chang and Wang, Xinrun and Zhang, Qinggang and Jiang, Qi and Huang, Xiao , booktitle =. Efficient Integration of External Knowledge to. 2025 , pages =

    2025

  14. [14]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Reflexion: Language Agents with Verbal Reinforcement Learning , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  15. [15]

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

    Shridhar, Mohit and Yuan, Xingdi and C. The Ninth International Conference on Learning Representations (ICLR) , year =

  16. [16]

    Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    Wang, Ruoyao and Jansen, Peter and C. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    2022

  17. [17]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    Reasoning with Language Model is Planning with World Model , author =. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP) , year =

    2023

  18. [18]

    2025 , url =

    Fang, Tianqing and Zhang, Hongming and Zhang, Zhisong and Ma, Kaixin and Yu, Wenhao and Mi, Haitao and Yu, Dong , booktitle =. 2025 , url =

    2025

  19. [19]

    Gu, Yu and Zhang, Kai and Ning, Yuting and Zheng, Boyuan and Gou, Boyu and Xue, Tianci and Chang, Cheng and Srivastava, Sanjari and Xie, Yanan and Qi, Peng and Sun, Huan and Su, Yu , journal =. Is Your. 2025 , url =

    2025

  20. [20]

    arXiv preprint arXiv:2601.03905 , year =

    Current Agents Fail to Leverage World Model as Tool for Foresight , author =. arXiv preprint arXiv:2601.03905 , year =

    arXiv

  21. [21]

    Xia, Peng and Zeng, Kaide and Liu, Jiaqi and Qin, Can and Wu, Fang and Zhou, Yiyang and Xiong, Caiming and Yao, Huaxiu , journal =

  22. [22]

    2025 , url =

    Qi, Zehan and Liu, Xiao and Iong, Iat Long and Lai, Hanyu and Sun, Xueqiao and Zhao, Wenyi and Yang, Yu and Yang, Xinyue and Sun, Jiadai and Yao, Shuntian and Zhang, Tianjie and Xu, Wei and Tang, Jie and Dong, Yuxiao , booktitle =. 2025 , url =

    2025

  23. [23]

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

    Web Agents with World Models: Learning and Leveraging Environment Dynamics in Web Navigation , author =. The Thirteenth International Conference on Learning Representations (ICLR) , year =

  24. [24]

    2025 , url =

    Fu, Dayuan and Huang, Jianzhao and Lu, Siyuan and Dong, Guanting and Wang, Yejie and He, Keqing and Xu, Weiran , booktitle =. 2025 , url =

    2025

  25. [25]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Agent Planning with World Knowledge Model , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  26. [26]

    2025 , url =

    Zhou, Siyu and Zhou, Tianyi and Yang, Yijun and Long, Guodong and Ye, Deheng and Jiang, Jing and Zhang, Chengqi , journal =. 2025 , url =

    2025

  27. [27]

    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

  28. [28]

    Reinforcement World Model Learning for

    Yu, Xiao and Peng, Baolin and Xu, Ruize and Shen, Yelong and He, Pengcheng and Nath, Suman and Singh, Nikhil and Gao, Jiangfeng and Yu, Zhou , journal =. Reinforcement World Model Learning for

  29. [29]

    2026 , url =

    Ding, Hang and Liu, Peidong and Wang, Junqiao and Ji, Ziwei and Cao, Meng and Zhang, Rongzhao and Ai, Lynn and Yang, Eric and Shi, Tianyu and Yu, Lei , journal =. 2026 , url =

    2026

  30. [30]

    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

  31. [31]

    Zero: Self-Evolving Search Agents without Training Data , author =

    Dr. Zero: Self-Evolving Search Agents without Training Data , author =. arXiv preprint arXiv:2601.07055 , year =

    arXiv

  32. [32]

    Guo, Jiacheng and Yang, Ling and Chen, Peter and Xiao, Qixin and Wang, Yinjie and Juan, Xinzhe and Qiu, Jiahao and Shen, Ke and Wang, Mengdi , journal =

  33. [33]

    arXiv preprint arXiv:2604.22748 , year =

    Agentic World Modeling: Foundations, Capabilities, Laws, and Beyond , author =. arXiv preprint arXiv:2604.22748 , year =

    Pith/arXiv arXiv

  34. [34]

    2024 , doi =

    Ma, Chang and Zhang, Junlei and Zhu, Zhihao and Yang, Cheng and Yang, Yujiu and Jin, Yaohui and Lan, Zhenzhong and Kong, Lingpeng and He, Junxian , booktitle =. 2024 , doi =

    2024

  35. [35]

    arXiv preprint arXiv:2510.16732 , year=

    A comprehensive survey on world models for embodied ai , author=. arXiv preprint arXiv:2510.16732 , year=

    Pith/arXiv arXiv

  36. [36]

    ACM Computing Surveys , volume=

    Understanding world or predicting future? a comprehensive survey of world models , author=. ACM Computing Surveys , volume=. 2025 , publisher=. doi:10.1145/3746449 , url=

    work page doi:10.1145/3746449 2025

  37. [37]

    arXiv preprint arXiv:2602.06130 , year=

    Self-Improving World Modelling with Latent Actions , author=. arXiv preprint arXiv:2602.06130 , year=

    arXiv

  38. [38]

    2024 , doi=

    Zhao, Andrew and Huang, Daniel and Xu, Quentin and Lin, Matthieu and Liu, Yong-Jin and Huang, Gao , booktitle=. 2024 , doi=

    2024

  39. [39]

    Huang, Chengsong and Yu, Wenhao and Wang, Xiaoyang and Zhang, Hongming and Li, Zongxia and Li, Ruosen and Huang, Jiaxin and Mi, Haitao and Yu, Dong , journal=

  40. [40]

    arXiv preprint arXiv:2510.08558 , year=

    Agent Learning via Early Experience , author=. arXiv preprint arXiv:2510.08558 , year=

    Pith/arXiv arXiv

  41. [41]

    arXiv preprint arXiv:2511.03773 , year=

    Scaling Agent Learning via Experience Synthesis , author=. arXiv preprint arXiv:2511.03773 , year=

    arXiv

  42. [42]

    arXiv preprint arXiv:2511.22254 , year=

    Co-Evolving Agents: Learning from Failures as Hard Negatives , author=. arXiv preprint arXiv:2511.22254 , year=

    arXiv

  43. [43]

    Co-Evolving

    Wang, Yinjie and Yang, Ling and Tian, Ye and Shen, Ke and Wang, Mengdi , booktitle =. Co-Evolving. 2025 , note =

    2025

  44. [44]

    2025 , url =

    Kim, Minsoo and Hwang, Seung-won , booktitle =. 2025 , url =

    2025

  45. [45]

    arXiv preprint arXiv:2602.10480 , year=

    Neuro-Symbolic Synergy for Interactive World Modeling , author=. arXiv preprint arXiv:2602.10480 , year=

    arXiv

  46. [46]

    arXiv preprint arXiv:2512.02472 , year=

    Guided self-evolving llms with minimal human supervision , author=. arXiv preprint arXiv:2512.02472 , year=

    arXiv

  47. [47]

    Maes, Lucas and Lidec, Quentin Le and Scieur, Damien and LeCun, Yann and Balestriero, Randall , journal=

  48. [48]

    Organization of Memory , editor=

    Episodic and Semantic Memory , author=. Organization of Memory , editor=

  49. [49]

    TextGrad: Automatic "Differentiation" via Text

    Y. arXiv preprint arXiv:2406.07496 , year=. doi:10.48550/arXiv.2406.07496 , eprint=

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2406.07496

  50. [50]

    arXiv preprint arXiv:2603.09400 , year=

    Reward Prediction with Factorized World States , author=. arXiv preprint arXiv:2603.09400 , year=. doi:10.48550/arXiv.2603.09400 , eprint=

    work page doi:10.48550/arxiv.2603.09400

  51. [51]

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

    Synapse: Trajectory-as-Exemplar Prompting with Memory for Computer Control , author=. The Twelfth International Conference on Learning Representations (ICLR) , year=

  52. [52]

    2024 , url=

    Zhou, Ruiwen and Yang, Yingxuan and Wen, Muning and Wen, Ying and Wang, Wenhao and Xi, Chunling and Xu, Guoqiang and Yu, Yong and Zhang, Weinan , booktitle=. 2024 , url=

    2024

  53. [53]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Contextual Experience Replay for Self-Improvement of Language Agents , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2025 , doi=

    2025

  54. [54]

    Deng, Yang and Zhang, Xuan and Zhang, Wenxuan and Yuan, Yifei and Ng, See Kiong and Chua, Tat-Seng , booktitle=. On the. 2024 , doi=

    2024

  55. [55]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    When to Trust Your Model: Model-Based Policy Optimization , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  56. [56]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Can We Edit Factual Knowledge by In-Context Learning? , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=. 2023 , doi=

    2023

  57. [57]

    arXiv preprint arXiv:2412.03624 , year=

    How to correctly do semantic backpropagation on language-based agentic systems , author=. arXiv preprint arXiv:2412.03624 , year=

    arXiv

  58. [58]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Editing Large Language Models: Problems, Methods, and Opportunities , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=. 2023 , address=. doi:10.18653/v1/2023.emnlp-main.632 , url=

    work page doi:10.18653/v1/2023.emnlp-main.632 2023

  59. [59]

    Aging with

    Hartvigsen, Thomas and Sankaranarayanan, Swami and Palangi, Hamid and Kim, Yoon and Ghassemi, Marzyeh , booktitle=. Aging with. 2023 , url=

    2023

  60. [60]

    2026 , url=

    Introducing. 2026 , url=

    2026

  61. [61]

    2025 , doi=

    Zhang, Yanzhao and Li, Mingxin and Long, Dingkun and Zhang, Xin and Lin, Huan and Yang, Baosong and Xie, Pengjun and Yang, An and Liu, Dayiheng and Lin, Junyang and Huang, Fei and Zhou, Jingren , journal=. 2025 , doi=

    2025

  62. [62]

    2024 , url=

    Zhong, Wanjun and Guo, Lianghong and Gao, Qiqi and Ye, He and Wang, Yanlin , booktitle=. 2024 , url=

    2024

  63. [63]

    2024 , url=

    Chen, Minghao and Li, Yihang and Yang, Yanting and Yu, Shiyu and Lin, Binbin and He, Xiaofei , booktitle=. 2024 , url=

    2024