pith. sign in

arxiv: 2606.02461 · v2 · pith:MKRF6HTTnew · submitted 2026-06-01 · 💻 cs.AI · cs.CL

AgentCL: Toward Rigorous Evaluation of Continual Learning in Language Agents

Yiheng Shu , Bernal Jim\'enez Guti\'errez , Saisri Padmaja Jonnalagedda , Yuguang Yao , Huan Sun , Yu Su This is my paper

Pith reviewed 2026-06-28 14:50 UTC · model grok-4.3

classification 💻 cs.AI cs.CL
keywords continual learninglanguage agentsevaluation frameworkmemory designstask streamstransfer gainscompositional streams
0
0 comments X

The pith

Controlled compositional streams distinguish memory plasticity in language agents where naive streams fail.

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

The paper introduces AgentCL as an evaluation framework that builds controlled task streams in which earlier sub-solutions, evidence, or workflows are intentionally reusable in later tasks. It contrasts these streams with naive ones where reusability is not guaranteed, then applies the framework to test non-parametric memory designs on coding, research, and reasoning tasks. The central finding is that controlled streams more clearly separate designs by their ability to achieve transfer gains without degradation, while naive and held-out settings frequently produce limited gains or expose interference. This matters because current benchmarks leave it unclear whether agents are actually accumulating and reusing experience across episodes rather than simply overwriting prior knowledge.

Core claim

AgentCL constructs compositional streams where sub-solutions from earlier tasks are reusable in later ones and shows through empirical analysis that these streams distinguish memory designs' plasticity and stable reuse more effectively than naive streams, which often yield limited gains and can reveal memory-induced degradation.

What carries the argument

Compositional task streams that intentionally link reusable sub-solutions across tasks, together with the MemProbe method that stores interactions, insights, and skills while filtering unreliable experiences during consolidation.

If this is right

  • Memory designs must balance the capacity to incorporate new experiences with the stable retention of reusable prior workflows.
  • Non-parametric memories can be diagnosed for consolidation failures by tracking which experiences are retained or filtered.
  • Benchmarks that rely only on naive or held-out task sequences are likely to underestimate both the potential and the risks of continual learning.
  • Stronger memory mechanisms are needed that maintain reuse without interference when tasks share substructure.

Where Pith is reading between the lines

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

  • Adoption of controlled streams could shift agent training toward explicit modeling of workflow reuse rather than isolated task performance.
  • The same construction method might be extended to multi-turn conversations or embodied environments where sub-skills recur.
  • If the streams generalize, they could serve as a diagnostic for whether current large language models already exhibit implicit continual learning or require external memory.

Load-bearing premise

The intentionally constructed compositional streams accurately capture the cross-task reusability relationships that arise in real agent usage.

What would settle it

A memory design that shows identical transfer gains and degradation rates on both controlled compositional streams and naive streams would falsify the claim that controlled streams are required to distinguish plasticity.

Figures

Figures reproduced from arXiv: 2606.02461 by Bernal Jim\'enez Guti\'errez, Huan Sun, Saisri Padmaja Jonnalagedda, Yiheng Shu, Yuguang Yao, Yu Su.

Figure 1
Figure 1. Figure 1: Task-stream construction and its impact on evaluation. Compositional relationship refers to [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Evaluation protocol and metrics of AGENTCL. Middle: memoryless baseline. Right: two passes are performed on the same stream. The first pass allows both reading from and writing to memory, whereas the second pass (or hold-out) is read-only. Left: held-out samples are not used to build memory. For task i, Bi , Fi , Si , and Hi denote the performance of the memoryless method, the first-pass performance, the s… view at source ↗
Figure 3
Figure 3. Figure 3: Performance on CodeEval-Pro complex tasks. The BigCodeBench-Lite-Pro subset is [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance on the selected 100 original BrowseComp+ tasks. another form of the same issue: it achieves positive compositional-stream PG (+3.8) and the largest naive-stream SG on CodeEval-Pro (+4.2), but suffers the largest GG drop after the compositional stream (−11.7), suggesting that memory useful within a stream can become harmful when the test distribution changes. BrowseComp+ in Appendix [PITH_FULL_… view at source ↗
Figure 5
Figure 5. Figure 5: Example of a CodeEval-Pro task pair. The subtask computes character frequencies for [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Example of a BrowseComp+ task and our synthesized subtasks. The subtasks derived [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Memory update prompt in the coding setting. The interactive trajectory is stored as [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Retrieved memory context in the coding setting. For each retrieved task, the solver is [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
read the original abstract

Language agents spend substantial inference time solving individual tasks, yet the experience acquired in one episode is often underutilized in future episodes. Continual learning expects an agent to accumulate reusable experience across a stream of tasks, improve over time, and avoid interference from irrelevant experiences. Unfortunately, existing benchmarks struggle to evaluate continual learning in language agents rigorously. Most efforts focus on retrieval and reasoning over long-context conversations or documents, while recent lifelong-adaptation benchmarks often rely on naive task streams with limited analysis of cross-task relationships, making it difficult to understand what an agent learns and reuses over time. This paper presents an evaluation framework AgentCL for continual learning in agents, centered on controlled task streams and metrics for transfer gains. AgentCL constructs compositional streams where earlier sub-solutions, evidence, or workflows are intentionally reusable in later tasks, and contrasts them with naive streams where such reusability is not guaranteed. We use the benchmark to evaluate non-parametric memory designs for continual learning. To diagnose how memory design choices affect continual learning, we develop MemProbe, a probing method that stores interactions, insights, and skills, while filtering unreliable experiences during consolidation. Empirical analysis across coding, deep research, and language understanding/reasoning tasks shows that naive streams offer limited ability to distinguish memory designs, whereas controlled streams more clearly distinguish their plasticity. Meanwhile, naive and held-out settings often yield limited gains and can expose memory-induced degradation. These results highlight the need for stronger memory designs that balance plasticity and stable reuse.

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 / 1 minor

Summary. The paper introduces the AgentCL evaluation framework for continual learning in language agents. It constructs controlled compositional task streams in which earlier sub-solutions, evidence, or workflows are deliberately reusable in later tasks, contrasts these with naive streams lacking guaranteed reusability, and evaluates non-parametric memory designs via a new MemProbe probing method that stores interactions, insights, and skills while filtering unreliable experiences. Empirical results across coding, deep research, and language tasks indicate that controlled streams distinguish memory plasticity more clearly than naive streams, while naive and held-out settings often produce limited gains or memory-induced degradation, motivating stronger memory designs that balance plasticity and stable reuse.

Significance. If the engineered reusability relationships in the controlled streams correspond to cross-task structure that arises in deployed agents, the framework would supply a more diagnostic benchmark than existing naive lifelong-adaptation suites and could guide development of memory mechanisms that accumulate reusable experience without interference. The introduction of MemProbe as a diagnostic tool for memory consolidation is a concrete methodological contribution that could be adopted more broadly.

major comments (2)
  1. [Benchmark construction] Benchmark construction (Abstract and the section describing controlled compositional streams): the central claim that controlled streams 'more clearly distinguish' memory plasticity rests on the streams being deliberately engineered for reusability; without an independent validation (e.g., comparison against naturally occurring task sequences from real agent logs or a quantitative measure of how well the engineered overlaps match observed reuse patterns), the reported superiority of controlled streams over naive streams remains conditional on the benchmark's internal construction rather than demonstrating a more rigorous or realistic evaluation.
  2. [Empirical analysis] Empirical analysis (the section reporting results on naive vs. controlled streams): the abstract states that 'naive streams offer limited ability to distinguish memory designs' and that 'naive and held-out settings often yield limited gains and can expose memory-induced degradation,' yet the provided abstract supplies no quantitative metrics, error bars, statistical tests, or exclusion criteria; if the full manuscript likewise lacks these details or does not report effect sizes for the plasticity distinctions, the load-bearing empirical contrast cannot be verified.
minor comments (1)
  1. [Abstract] The abstract refers to 'non-parametric memory designs' and 'MemProbe' without a brief parenthetical definition or pointer to the methods section; adding one sentence would improve accessibility for readers unfamiliar with the specific memory variants under test.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major point below, clarifying our design choices and empirical reporting while outlining planned revisions.

read point-by-point responses

  1. Referee: [Benchmark construction] Benchmark construction (Abstract and the section describing controlled compositional streams): the central claim that controlled streams 'more clearly distinguish' memory plasticity rests on the streams being deliberately engineered for reusability; without an independent validation (e.g., comparison against naturally occurring task sequences from real agent logs or a quantitative measure of how well the engineered overlaps match observed reuse patterns), the reported superiority of controlled streams over naive streams remains conditional on the benchmark's internal construction rather than demonstrating a more rigorous or realistic evaluation.

    Authors: The core contribution of AgentCL is a controlled evaluation framework that deliberately engineers known reusability relationships to isolate and diagnose memory plasticity effects, which naive streams do not permit. This design choice enables clearer distinctions in our experiments precisely because overlaps are measurable and intentional, supporting rigorous analysis of transfer versus interference. We do not claim equivalence to real-world logs; rather, the framework provides a diagnostic benchmark for memory mechanisms. We will add explicit discussion of this motivation and its trade-offs in the introduction and a new limitations subsection. revision: partial

  2. Referee: [Empirical analysis] Empirical analysis (the section reporting results on naive vs. controlled streams): the abstract states that 'naive streams offer limited ability to distinguish memory designs' and that 'naive and held-out settings often yield limited gains and can expose memory-induced degradation,' yet the provided abstract supplies no quantitative metrics, error bars, statistical tests, or exclusion criteria; if the full manuscript likewise lacks these details or does not report effect sizes for the plasticity distinctions, the load-bearing empirical contrast cannot be verified.

    Authors: The full manuscript reports quantitative results across tasks, including performance deltas, error bars from multiple runs, and statistical comparisons in the empirical analysis section. The abstract summarizes findings at a high level due to space limits. We will revise the abstract to include representative quantitative metrics and effect sizes for the controlled versus naive distinctions. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical benchmark evaluation is self-contained

full rationale

The paper introduces AgentCL as an evaluation framework that explicitly constructs compositional task streams with engineered reusability and contrasts them against naive streams. It then reports empirical results from running memory designs on these streams, showing clearer distinctions in plasticity for controlled streams. No equations, fitted parameters, or derivations are present that reduce claims to self-definition or input renaming. The central distinction follows directly from the transparent benchmark construction and external experimental outcomes rather than any load-bearing self-citation or ansatz. This matches the default case of a non-circular empirical contribution.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review supplies no information on free parameters, background axioms, or new postulated entities.

pith-pipeline@v0.9.1-grok · 5824 in / 1018 out tokens · 23605 ms · 2026-06-28T14:50:42.237009+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

32 extracted references · 8 linked inside Pith

  1. [1]

    gpt-oss-120b & gpt-oss-20b model card.arXiv preprint arXiv:2508.10925, 2025

    Sandhini Agarwal et al. gpt-oss-120b & gpt-oss-20b model card.arXiv preprint arXiv:2508.10925, 2025

    Pith/arXiv arXiv 2025

  2. [2]

    Memorybench: A benchmark for memory and continual learning in LLM systems.CoRR, abs/2510.17281, 2025

    Qingyao Ai, Yichen Tang, Changyue Wang, Jianming Long, Weihang Su, and Yiqun Liu. Memorybench: A benchmark for memory and continual learning in LLM systems.CoRR, abs/2510.17281, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    Gonzalez

    Parth Asawa, Chris Glaze, Gabe Orlanski, Ramya Ramakrishnan, Benji Xu, Asim Biswal, Vincent Sunn Chen, Frederic Sala, Matei Zaharia, and Joseph E. Gonzalez. Continual learn- ing bench. https://continual-learning-bench.com/docs/, 2026. Documentation, accessed 2026-05-04

    2026

  4. [4]

    Browsecomp-plus: A more fair and transparent evaluation benchmark of deep-research agent.CoRR, abs/2508.06600, 2025

    Zijian Chen, Xueguang Ma, Shengyao Zhuang, Ping Nie, Kai Zou, Andrew Liu, Joshua Green, Kshama Patel, Ruoxi Meng, Mingyi Su, Sahel Sharifymoghaddam, Yanxi Li, Haoran Hong, Xinyu Shi, Xuye Liu, Nandan Thakur, Crystina Zhang, Luyu Gao, Wenhu Chen, and Jimmy Lin. Browsecomp-plus: A more fair and transparent evaluation benchmark of deep-research agent.CoRR, a...

    arXiv 2025

  5. [5]

    Mem0: Building production-ready AI agents with scalable long-term memory

    Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. Mem0: Building production-ready AI agents with scalable long-term memory. In Inês Lynce, Nello Murano, Mauro Vallati, Serena Villata, Federico Chesani, Michela Milano, Andrea Omicini, and Mehdi Dastani, editors,ECAI 2025 - 28th European Conference on Artificial Intelligence, 25-...

    2025

  6. [6]

    A survey of self-evolving agents: What, when, how, and where to evolve on the path to artificial super intelligence.Trans

    Huan-ang Gao, Jiayi Geng, Wenyue Hua, Mengkang Hu, Xinzhe Juan, Hongzhang Liu, Shilong Liu, Jiahao Qiu, Xuan Qi, Qihan Ren, Yiran Wu, Hongru Wang, Han Xiao, Yuhang Zhou, Shaokun Zhang, Jiayi Zhang, Jinyu Xiang, Yixiong Fang, Qiwen Zhao, Dongrui Liu, Cheng Qian, Zhenhailong Wang, Minda Hu, Huazheng Wang, Qingyun Wu, Heng Ji, and Mengdi Wang. A survey of se...

    2026

  7. [7]

    Yuanzhe Hu, Yu Wang, and Julian J. McAuley. Evaluating memory in LLM agents via incremental multi-turn interactions.CoRR, abs/2507.05257, 2025. 10

    Pith/arXiv arXiv 2025

  8. [8]

    Swe-bench-cl: Continual learning for coding agents.CoRR, abs/2507.00014, 2025

    Thomas Joshi, Shayan Chowdhury, and Fatih Uysal. Swe-bench-cl: Continual learning for coding agents.CoRR, abs/2507.00014, 2025

    arXiv 2025

  9. [9]

    Langmem, 2026

    LangChain AI. Langmem, 2026. GitHub repository

    2026

  10. [10]

    Agentboard: An analytical evaluation board of multi-turn LLM agents

    Chang Ma, Junlei Zhang, Zhihao Zhu, Cheng Yang, Yujiu Yang, Yaohui Jin, Zhenzhong Lan, Lingpeng Kong, and Junxian He. Agentboard: An analytical evaluation board of multi-turn LLM agents. In Amir Globersons, Lester Mackey, Danielle Belgrave, Angela Fan, Ulrich Paquet, Jakub M. Tomczak, and Cheng Zhang, editors,Advances in Neural Information Processing Syst...

    2024

  11. [11]

    Evaluating very long-term conversational memory of LLM agents

    Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. Evaluating very long-term conversational memory of LLM agents. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors,Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thaila...

    2024

  12. [12]

    Introducing gpt-5.2

    OpenAI. Introducing gpt-5.2. https://openai.com/index/introducing-gpt-5-2/ , De- cember 2025

    2025

  13. [13]

    Qwen3.5-35b-a3b

    Qwen Team. Qwen3.5-35b-a3b. Hugging Face model card, March 2026

    2026

  14. [14]

    Language agents: Foundations, prospects, and risks

    Yu Su, Diyi Yang, Shunyu Yao, and Tao Yu. Language agents: Foundations, prospects, and risks. In Junyi Jessy Li and Fei Liu, editors,Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: EMNLP 2024 - Tutorial Abstracts, Miami, Florida, USA, November 12-16, 2024, pages 17–24. Association for Computational Linguistics, 2024

    2024

  15. [15]

    Dynamic cheatsheet: Test-time learning with adaptive memory

    Mirac Suzgun, Mert Yüksekgönül, Federico Bianchi, Dan Jurafsky, and James Zou. Dynamic cheatsheet: Test-time learning with adaptive memory. In Vera Demberg, Kentaro Inui, and Lluís Marquez, editors,Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2026 - Volume 1: Long Papers, Rabat, Morocco,...

    2026

  16. [16]

    A comprehensive survey of continual learning: Theory, method and application.IEEE Trans

    Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application.IEEE Trans. Pattern Anal. Mach. Intell., 46(8):5362– 5383, 2024

    2024

  17. [17]

    Mmlu-pro: A more robust and challenging multi-task language understanding benchmark

    Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, Tianle Li, Max Ku, Kai Wang, Alex Zhuang, Rongqi Fan, Xiang Yue, and Wenhu Chen. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. In Amir Globersons, Lester Mackey, Danielle Belgrave, Angela ...

    2024

  18. [18]

    Agent workflow memory

    Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, and Graham Neubig. Agent workflow memory. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors,Forty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025, Proceedings of Machine L...

    2025

  19. [19]

    Browsecomp: A simple yet challenging benchmark for browsing agents.CoRR, abs/2504.12516, 2025

    Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, and Amelia Glaese. Browsecomp: A simple yet challenging benchmark for browsing agents.CoRR, abs/2504.12516, 2025

    Pith/arXiv arXiv 2025

  20. [20]

    Chi, Chi Wang, Shuo Chen, Fernando Pereira, Wang-Cheng Kang, and Derek Zhiyuan Cheng

    Tianxin Wei, Noveen Sachdeva, Benjamin Coleman, Zhankui He, Yuanchen Bei, Xuying Ning, Mengting Ai, Yunzhe Li, Jingrui He, Ed H. Chi, Chi Wang, Shuo Chen, Fernando Pereira, Wang-Cheng Kang, and Derek Zhiyuan Cheng. Evo-memory: Benchmarking LLM agent test-time learning with self-evolving memory.CoRR, abs/2511.20857, 2025. 11

    Pith/arXiv arXiv 2025

  21. [21]

    Stream- bench: Towards benchmarking continuous improvement of language agents

    Cheng-Kuang Wu, Zhi Rui Tam, Chieh-Yen Lin, Yun-Nung Chen, and Hung-yi Lee. Stream- bench: Towards benchmarking continuous improvement of language agents. In Amir Glober- sons, Lester Mackey, Danielle Belgrave, Angela Fan, Ulrich Paquet, Jakub M. Tomczak, and Cheng Zhang, editors,Advances in Neural Information Processing Systems 38: Annual Confer- ence on...

    2024

  22. [22]

    Long- memeval: Benchmarking chat assistants on long-term interactive memory

    Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. Long- memeval: Benchmarking chat assistants on long-term interactive memory. InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28,

    2025

  23. [23]

    OpenReview.net, 2025

    2025

  24. [24]

    Agent-dice: Disentangling knowledge updates via geometric consensus for agent continual learning.CoRR, abs/2601.03641, 2026

    Zheng Wu, Xingyu Lou, Xinbei Ma, Yansi Li, Weiwen Liu, Weinan Zhang, Jun Wang, and Zhuosheng Zhang. Agent-dice: Disentangling knowledge updates via geometric consensus for agent continual learning.CoRR, abs/2601.03641, 2026

    Pith/arXiv arXiv 2026

  25. [25]

    A-MEM: agentic memory for LLM agents.CoRR, abs/2502.12110, 2025

    Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. A-MEM: agentic memory for LLM agents.CoRR, abs/2502.12110, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    Autonomous continual learning of computer-use agents for environment adaptation, 2026

    Tianci Xue, Zeyi Liao, Tianneng Shi, Zilu Wang, Kai Zhang, Dawn Song, Yu Su, and Huan Sun. Autonomous continual learning of computer-use agents for environment adaptation, 2026

    2026

  27. [27]

    Narasimhan, and Yuan Cao

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5,

    2023

  28. [28]

    OpenReview.net, 2023

    2023

  29. [29]

    CGL: advancing continual GUI learning via reinforcement fine-tuning

    Zhenquan Yao, Zitong Huang, Yihan Zeng, Jianhua Han, Hang Xu, Chun-Mei Feng, Jianwei Ma, and Wangmeng Zuo. CGL: advancing continual GUI learning via reinforcement fine-tuning. CoRR, abs/2603.02951, 2026

    arXiv 2026

  30. [30]

    Humaneval pro and MBPP pro: Evaluating large language models on self-invoking code generation task

    Zhaojian Yu, Yilun Zhao, Arman Cohan, and Xiaoping Zhang. Humaneval pro and MBPP pro: Evaluating large language models on self-invoking code generation task. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar, editors,Findings of the Association for Computational Linguistics, ACL 2025, Vienna, Austria, July 27 - August 1, 2025,...

    2025

  31. [31]

    Qwen3 embed- ding: Advancing text embedding and reranking through foundation models.arXiv preprint arXiv:2506.05176, 2025

    Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, Fei Huang, and Jingren Zhou. Qwen3 embed- ding: Advancing text embedding and reranking through foundation models.arXiv preprint arXiv:2506.05176, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Lifelongagentbench: Evaluating LLM agents as lifelong learners.CoRR, abs/2505.11942, 2025

    Junhao Zheng, Xidi Cai, Qiuke Li, Duzhen Zhang, Zhong-Zhi Li, Yingying Zhang, Le Song, and Qianli Ma. Lifelongagentbench: Evaluating LLM agents as lifelong learners.CoRR, abs/2505.11942, 2025. 12 Appendix Contents A Limitations 14 B Experimental Setting and Details 14 B.1 Examples of Compositionality . . . . . . . . . . . . . . . . . . . . . . . . . . . ....

    arXiv 2025