Pith. sign in

REVIEW 4 major objections 4 minor 70 references

PATH-Bench: Path-Dependent Evaluation of Lifelong Agents

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read PATH-Bench traces how the order of prior tasks changes what a lifelong LLM agent transfers and retains, and shows that filtering experience at the moment of use cuts forgetting.

desk verdict Useful benchmark protocol, but the 'controlled' histories rely on an ICL proxy that the paper's own data show is inconsistent at the agent level. read the letter →

arxiv 2608.01149 v1 pith:IGGG3HOM submitted 2026-08-02 cs.AI

Xidong Yang , Xingyi Zhang , Wenhao Li , Wenyan Liu , Junjie Sheng , Yun Hua , Wei Yin , Tao Fang
show 2 more authors
Chuyun Shen Xiangfeng Wang
This is my paper
classification cs.AI
keywords pathdependencelifelonglearningLLMagentsexternalstateforwardtransferbackwardforgettingselectiveexperienceuse
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

PATH-Bench is a controlled evaluation framework for lifelong LLM agents that adapt by storing memories or skills outside the model. It estimates directed transfer relationships between tasks from one-shot in-context learning across three frozen LLMs, then builds probe-centered task sequences whose histories are dominated by either helpful or interfering tasks. Repeatedly scoring the same probe task lets it measure average performance, forward transfer, backward transfer, and forgetting under controlled paths. The paper's empirical claim is that experience utility is not intrinsic: it depends on both how experience is represented and on the task's interaction structure, and strong immediate transfer can coexist with substantial later forgetting. On the basis of these diagnostics, it introduces Selective Experience Use, a lightweight filter that reduces forgetting in every tested condition and improves forward transfer in most.

What carries the argument

The load-bearing object is the directed transfer-gain matrix $M\in\mathbb{R}^{D\times D}$ with entries $M_{i,j}=S(t_j\mid t_i)-S(t_j\mid\emptyset)$, estimated by three frozen LLMs and collapsed by sign vote into a relation matrix $V$. This matrix turns 'which tasks help or hurt which other tasks' into a controllable sampling prior. A probe-centered sequence sampler then builds histories of specified length and transfer composition, reinserting the probe at fixed intervals; average performance (AP), forward transfer (FWT), backward transfer (BWT), and forgetting (FGT) are computed from the probe trajectory and the whole sequence.

What would settle it

Concrete check: take a task pair that all three models voted as positive transfer, build a sequence where only that source task precedes the probe, and compare the probe score against its cold-start baseline across the eight agents. If the agent-level improvement is not systematically positive—or if the multi-task prefix gains reverse sign when the prefix is actually stored in agent memory rather than written into the prompt—then the in-context proxy is not controlling the path.

Watch

Extended reading notes

Core claim

PATH-Bench claims that path dependence in lifelong LLM agents can be measured efficiently and that, once measured, it overturns the assumption that accumulated experience simply helps. It treats probe-task performance as a function of an ordered intervening history, estimates a directed transfer-gain matrix $M_{i,j}=S(t_j\mid t_i)-S(t_j\mid\emptyset)$ for every task pair with three frozen LLMs, and collapses signs by majority vote. Probe-centered sequences are sampled from positive-, neutral-, and negative-transfer pools, with the probe reinserted at intervals, so forward transfer, backward transfer, and forgetting are read from one probe trajectory. Across eight agents on code generation an

Load-bearing premise

The design assumes that one-shot in-context transfer scores measured on frozen LLMs predict how a lifelong agent's stored memories or skills will actually help or hurt it on a later task; if that proxy fails, the controlled helpful and interfering histories stop being controlled.

Editorial extensions

If this is right

  • Benchmark rankings reported on a single fixed path are not reliable evidence about which lifelong agent is better; the same task set can order agents differently depending on history.
  • Transfer and retention must be reported as separate axes: a harness can show strong forward transfer while still forgetting a large share of what it learned.
  • Experience representations are not universally useful; task-specific memory helps on single-turn code tasks while compact abstractions help on multi-turn tool use.
  • A lightweight gating step over retrieved experience, applied at inference time, reduces forgetting across all tested histories and improves forward transfer in most settings.

Reading between the lines

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

  • Beyond the paper: the probe-centered protocol is portable, since any domain with a task pool and a base LLM could reuse the same transfer-gain matrix and sequence sampler at low marginal cost.
  • Beyond the paper: the SEU result implies a general architectural principle—treat retrieved experience as untrusted context and gate it before injection—which could be tested as a standalone module inside agents that do not otherwise filter memory.
  • Beyond the paper: the ICL proxy is untested when intervening tasks actually update the agent's external state; running the Appendix B prefix validation with real memory updates would either confirm or bound the benchmark's core assumption.
  • Beyond the paper: because only one backbone drives the agent evaluations, the path-dependence findings may be backbone-specific; repeating on other base models could separate harness-level effects from model-level priors.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. PATH-Bench proposes a benchmark for path-dependent evaluation of lifelong LLM agents. The pipeline selects task pools with performance headroom, estimates a directed task-transfer matrix using one-shot in-context learning with three frozen LLMs, and then constructs probe-centered learning sequences with positive-, neutral-, and negative-transfer intervening tasks. The evaluation protocol tracks a repeated probe task to compute average performance, forward transfer, backward transfer, and forgetting. The paper evaluates eight lifelong agents on BigCodeBench (single-turn code generation) and WildToolBench (multi-turn tool use) under positive- and negative-dominant histories, reports that experience utility depends jointly on representation and interaction structure, that transfer and retention dissociate, and that later experience can reshape earlier gains. It also introduces Selective Experience Use (SEU), a lightweight harness that filters retrieved experience, and reports that SEU reduces forgetting in all evaluated settings and improves forward transfer in a majority.

Significance. The paper addresses a real and under-studied problem: existing lifelong-agent benchmarks rarely control the order and composition of task histories, so path-dependence is often ignored. The probe-centered longitudinal design is a useful and efficient alternative to exhaustive full-set re-evaluation, and the multi-model voting approach is a reasonable attempt to reduce noise in transfer estimation. If the ICL-based transfer labels were validated at the agent level, PATH-Bench would be a valuable benchmarking infrastructure. The empirical study across eight agents is also a useful contribution, and SEU, although simple, provides actionable guidance with reported reductions in forgetting. The strengths are the clear protocolization of sequence sampling, the explicit metrics, and the inclusion of a validation appendix. However, as detailed below, the central claim of 'controlled helpful and interfering histories' is not established because the transfer labels come from an unvalidated proxy and some of the paper's own data contradict the expected control.

major comments (4)
  1. [Section 3.2, Eq. (2); Appendix B; Appendix C; Figure 5] The 'controlled helpful and interfering histories' claim rests entirely on the one-shot ICL transfer-gain matrix V defined in Eq. (2). Appendix B validates only sign preservation under multi-task ICL prefixes on frozen backbone models, not transfer through agent memory/skill states. The paper's own data contradict the proxy's agent-level validity: Figure 5 shows that an initial negative interval improves probe performance for Clin, MemRL, AutoSkill, and SkillClaw; Appendix C shows WildToolBench negative-dominant curves at or above positive-dominant curves for Clin, AWM, and SimpleMem; Table 1 shows negative-dominant FWT exceeding positive-dominant FWT for several agents. Unless V is re-estimated per agent or supplemented with agent-level transfer validation, the positive/negative histories are not actually controlled, and the measured FWT/BWT/FGT differences cannot be attributed to path-
  2. [Section 4.1, Table 1] All primary results in Table 1 are point estimates without standard deviations, confidence intervals, or significance tests. Claims such as AutoSkill 'dominating' BigCodeBench and the cross-dataset reversal in Finding 2 rest on differences of 1–3 percentage points. The sequences and runs are repeated, and the comparison is paired across agents for the same sequences, so paired significance tests or bootstrap confidence intervals should be reported. Without them, the rankings and the 'majority of settings' conclusions are not statistically supported. This issue also affects Table 2 and Figure 5, where error bars are only partially provided.
  3. [Section 4.1, Finding 2] Finding 2 attributes the observed reversal in representation efficacy to interaction structure (single-turn vs. multi-turn). However, BigCodeBench and WildToolBench also differ in domain (code generation vs. tool use), task semantics, and evaluation protocol. The two dimensions are fully confounded. To support the claim that interaction structure is the driver, the comparison would need to hold domain fixed (e.g., single-turn vs. multi-turn tool use, or code tasks with multi-turn interaction) or otherwise control for domain. As written, the evidence supports only an interaction between representation and dataset, not specifically interaction structure.
  4. [Section 4.2, Figure 5; Section 4.3, Table 2] The path-transition analysis in Figure 5 is phrased as a confirmation that the transfer-gain matrix identifies genuinely helpful experience ('A positive-transfer interval consistently improves all eight agents'). Yet the same figure shows that negative transfer intervals also improve half the agents, and the paper dismisses this as agent-dependent. This is not a minor caveat; it is evidence that the sign labels are not behaviorally valid for a large fraction of agent/task combinations. The benchmark conclusions should be reframed as characterizing ICL-labeled sequences, with the discrepancy reported as a limitation, or the benchmark should incorporate agent-specific transfer labels. Similarly, Table 2 reports SEU effects on a 20-sequence subset without statistical significance tests; the claim of 'consistently reduces forgetting' should be qualified accordingly.
minor comments (4)
  1. [Section 3.3, Metrics] The metric definitions use p_probe^i but the notation is not formally introduced; the paragraph before 'Evaluation Metrics' defines p_probe_i. Aligning notation would improve readability. Also, the BWT and FGT definitions split the signed difference relative to pwarm into two one-sided terms; this is non-standard and should be explicitly justified.
  2. [Appendix D, Figure 11] The expected output format says 'Return exactly one of the following forms: KEEP, IGNORE, SUMMARIZE <the summarized memory>'. The 'SUMMARIZE' form is ambiguous: should the word SUMMARIZE appear alone if the summary is empty, or must it always be followed by text? Clarifying the output grammar would make the method reproducible.
  3. [Appendix B, Table 3] Cross-model agreement rates are reported as percentages, but the number of non-mixed pairs differs between datasets (12,279 vs. 12,038). The percentage alone does not make it easy to see the scale of disagreement; reporting counts and the distribution of mixed labels would be helpful.
  4. [Figure 1] The caption says 'Each cell shows the transition from Path 1 to Path 2' but the reader has to infer what the arrow notation and color scale mean. A short explanation of the rank-improvement encoding would make the figure self-contained.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PATH-Bench's empirical findings are not forced by its definitions or by self-citation.

full rationale

The only potentially circular-looking element is the use of the ICL transfer-gain matrix V to construct 'positive' and 'negative' histories. But V is estimated independently of the evaluated agents, from three frozen LLMs (DeepSeek-V4-Flash, GPT-5.4-mini, GLM-5.0), and is not fitted to the FWT/BWT/FGT values reported later. The measured probe-score changes can and do diverge from V's sign: in Figure 5, negative-labeled intervals still produce positive probe changes for Clin, MemRL, AutoSkill, and SkillClaw; in Table 1, BGE-M3 RAG has higher FWT under negative-dominant than positive-dominant histories on BigCodeBench. Such mismatches would be impossible if the results were definitionally determined by the sampling labels. The statement that a positive-transfer interval improves all eight agents is an empirical validation of the proxy, not a tautology, because the improvement is measured at the agent level and could have failed. The FWT/BWT/FGT metrics are standard lifelong-learning quantities and are not constructed to force the paper's findings; FWT and FGT can move independently. SEU is an intervention evaluated by paired comparisons on held-out sequences, not derived from the benchmark definitions. The only self-citations (e.g., Yang et al. 2026a for external learning states) appear in background statements and are not load-bearing for the central results. No step reduces to its own input by construction, and no fitted parameter is renamed as a prediction; the paper's claims remain externally falsifiable even though the ICL proxy's construct validity can be questioned as a correctness risk rather than circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central results depend on the ICL transfer proxy, the headroom filtering assumption, and several protocol hyperparameters chosen by hand. No new ontological entities are introduced; SEU is a method, not an entity. The lack of sensitivity analysis for the hyperparameters is a gap.

free parameters (5)
  • Warm-up length = 5 tasks
    Chosen by hand as the short positive-transfer block before the post-warm-up reference; no sensitivity analysis reported.
  • Probe interval = 6-12 intervening tasks
    Uniform range chosen to interleave probes; no sensitivity analysis.
  • Dominance ratio = 70%
    Fraction of intervening tasks drawn from the dominant transfer set; chosen by hand, not swept.
  • Sequence count = 50 sequences per condition, 5 runs
    Evaluation budget; no power analysis to support significance of observed differences.
  • Task pool size = 120 per dataset
    Sampled after headroom filter; no analysis of stability of the transfer matrix with pool size.
assumptions (4)
  • domain assumption One-shot ICL transfer gain is a valid proxy for inter-task transfer in lifelong agent learning.
    Section 3.2: 'we use one-shot in-context learning effects as a practical proxy for inter-task transferability'. The entire sequence construction depends on it.
  • domain assumption Majority voting over three model-specific transfer sign matrices yields reliable task relations.
    Section 3.2: entries fixed when at least two models agree; mixed pairs excluded. Reliability of this consensus is assumed, with validation in Appendix B.
  • domain assumption Headroom filtering by base-LLM baseline scores yields a meaningful task pool.
    Section 3.2 Task selection: tasks with saturated or negligible baseline performance are excluded. This assumes that headroom predicts learning signal.
  • ad hoc to paper Differences between BigCodeBench and WildToolBench are driven by interaction structure (single-turn vs multi-turn), not by domain.
    Finding 2 in Section 4.1 attributes representation preferences to interaction structure, but the datasets differ in domain as well. This unstated premise is required for the claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PATH-Bench: Path-Dependent Evaluation of Lifelong Agents." pith.science (2026). https://pith.science/paper/IGGG3HOM

@misc{pith2026260801149,
  author       = {Pith},
  title        = {Pith review of: PATH-Bench: Path-Dependent Evaluation of Lifelong Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IGGG3HOM}},
  note         = {Machine review of arXiv:2608.01149}
}
read the original abstract

Lifelong LLM agents increasingly adapt through external learning states that store past interactions as retrievable memories or reusable skills, yet existing benchmarks rarely account for how the path of accumulated experience shapes what agents transfer and retain. In this work, we establish PATH-Bench, a benchmark for path-dependent evaluation of lifelong agents. PATH-Bench estimates directed task relationships via multi-model in-context learning, constructs probe-centered sequences with controlled helpful and interfering histories, and repeatedly evaluates probe tasks to measure average performance, forward transfer, backward transfer, and forgetting. We evaluate eight representative agents on single-turn code generation and multi-turn tool-use tasks under positive- and negative-dominant histories. Benchmark results show that experience utility depends jointly on how experience is represented and on the task's interaction structure, that strong transfer does not ensure retention, and that later experience can reshape gains acquired earlier in the learning path. Based on these findings, we propose Selective Experience Use (SEU), an agent harness that regulates how path-accumulated experience influences each new task, admitting helpful items while filtering out potential interference. SEU consistently reduces forgetting while improving forward transfer in the majority of settings. The PATH-Bench provides both a controlled evaluation framework and actionable guidance for designing more selective and robust lifelong agents.

Figures

Figures reproduced from arXiv: 2608.01149 by the authors.

Figure 1
Figure 1. Agent-rank changes across two task paths on Life [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of PATH-Bench. The framework filters both task pools for measurable headroom, estimates directed task [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Probe reappearance dynamics on BigCodeBench under positive- and negative-dominant histories. Curves report the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Change in average performance relative to the [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Controlled path-transition effects on BigCodeBench. Points and error bars denote means and 95% confidence intervals. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Probe reappearance dynamics on WildToolBench under positive- and negative-dominant histories. Curves report the [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 8
Figure 8. Figure 8: Task composition of the BigCodeBench (BCB) [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: Fifteen most frequently required libraries in the [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Distribution of instruction transition types across [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Prompt for Selective Experience Use [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

70 extracted references · 51 canonical work pages

  1. [1]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

  2. [2]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

  3. [3]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    2016

  4. [4]

    Lin, Wenye and Roberts, Jonathan and Yang, Yunhan and Albanie, Samuel and Lu, Zongqing and Han, Kai , booktitle=

  5. [5]

    ACM Transactions on Software Engineering and Methodology , volume=

    Large language models for software engineering: A systematic literature review , author=. ACM Transactions on Software Engineering and Methodology , volume=

    work page

  6. [6]

    Shojaee, Parshin and Nguyen, Ngoc-Hieu and Meidani, Kazem and Farimani, Amir Barati and Doan, Khoa D and Reddy, Chandan K , booktitle=

    work page

  7. [7]

    arXiv preprint arXiv:2506.18096 , year=

    Deep research agents: A systematic examination and roadmap , author=. arXiv preprint arXiv:2506.18096 , year=

    arXiv

  8. [8]

    arXiv preprint arXiv:2512.13564 , year=

    Memory in the Age of AI Agents , author=. arXiv preprint arXiv:2512.13564 , year=

    arXiv

Show all 70 references
  1. [9]

    Evaluating very long-term conversational memory of

    Maharana, Adyasha and Lee, Dong-Ho and Tulyakov, Sergey and Bansal, Mohit and Barbieri, Francesco and Fang, Yuwei , booktitle=. Evaluating very long-term conversational memory of

  2. [10]

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

    Beyond prompts: Dynamic conversational benchmarking of large language models , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  3. [11]

    Wan, Luanbo and Ma, Weizhi , journal=

  4. [12]

    Wu, Cheng-Kuang and Tam, Zhi R and Lin, Chieh-Yen and Chen, Yun-Nung and Lee, Hung-yi , booktitle=

  5. [13]

    Zheng, Junhao and Cai, Xidi and Li, Qiuke and Zhang, Duzhen and Li, ZhongZhi and Zhang, Yingying and Song, Le and Ma, Qianli , journal=

  6. [14]

    Ai, Qingyao and Tang, Yichen and Wang, Changyue and Long, Jianming and Su, Weihang and Liu, Yiqun , journal=

  7. [15]

    arXiv preprint arXiv:2508.19005 , year=

    Building self-evolving agents via experience-driven lifelong learning: A framework and benchmark , author=. arXiv preprint arXiv:2508.19005 , year=

  8. [16]

    ACM Transactions on Information Systems , year=

    A survey on the memory mechanism of large language model based agents , author=. ACM Transactions on Information Systems , year=

  9. [17]

    Wei, Tianxin and Sachdeva, Noveen and Coleman, Benjamin and He, Zhankui and Bei, Yuanchen and Ning, Xuying and Ai, Mengting and Li, Yunzhe and He, Jingrui and Chi, Ed H and others , journal=

  10. [18]

    arXiv preprint arXiv:2506.21872 , year=

    A Survey of Continual Reinforcement Learning , author=. arXiv preprint arXiv:2506.21872 , year=

    arXiv

  11. [19]

    arXiv preprint arXiv:2506.21899 , year=

    Advancements and Challenges in Continual Reinforcement Learning: A Comprehensive Review , author=. arXiv preprint arXiv:2506.21899 , year=

    arXiv

  12. [20]

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

    Continual world: A robotic benchmark for continual reinforcement learning , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  13. [21]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Lifelong learning of large language model based agents: A roadmap , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

  14. [22]

    Conference on Empirical Methods in Natural Language Processing (EMNLP) , year=

    Revisiting catastrophic forgetting in large language model tuning , author=. Conference on Empirical Methods in Natural Language Processing (EMNLP) , year=

  15. [23]

    ACM Computing Surveys , volume=

    Towards lifelong learning of large language models: A survey , author=. ACM Computing Surveys , volume=

  16. [24]

    ACM Computing Surveys , volume=

    Continual learning of large language models: A comprehensive survey , author=. ACM Computing Surveys , volume=

  17. [25]

    Zhang, Shengtao and Wang, Jiaqian and Zhou, Ruiwen and Liao, Junwei and Feng, Yuchen and Zhang, Weinan and Wen, Ying and Li, Zhiyu and Xiong, Feiyu and Qi, Yutao and Tang, Bo and Wen, Muning , journal=

  18. [26]

    He, Junda and Treude, Christoph and Lo, David , journal=

  19. [27]

    International Conference on Learning Representations (ICLR) , year=

    Self-evolving multi-agent collaboration networks for software development , author=. International Conference on Learning Representations (ICLR) , year=

  20. [28]

    Wang, Zhenting and Chang, Qi and Patel, Hemani and Biju, Shashank and Wu, Cheng-En and Liu, Quan and Ding, Aolin and Rezazadeh, Alireza and Shah, Ankit and Bao, Yujia and others , journal=

  21. [29]

    Neural Networks , volume=

    A domain-agnostic approach for characterization of lifelong learning systems , author=. Neural Networks , volume=

  22. [30]

    Johnson, Erik C and Nguyen, Eric Q and Schreurs, Blake and Ewulum, Chigozie S and Ashcraft, Chace and Fendley, Neil M and Baker, Megan M and New, Alexander and Vallabha, Gautam K , journal=

  23. [31]

    Neural networks , volume=

    Continual lifelong learning with neural networks: A review , author=. Neural networks , volume=

  24. [32]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    A comprehensive survey of continual learning: Theory, method and application , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

  25. [33]

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

    Gradient episodic memory for continual learning , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  26. [34]

    Proceedings of the National Academy of Sciences , volume=

    Overcoming catastrophic forgetting in neural networks , author=. Proceedings of the National Academy of Sciences , volume=

  27. [35]

    AAAI Conference on Artificial Intelligence (AAAI) , year=

    Continual variational autoencoder via continual generative knowledge distillation , author=. AAAI Conference on Artificial Intelligence (AAAI) , year=

  28. [36]

    AAAI Conference on Artificial Intelligence (AAAI) , year=

    Lifelong generative modelling using dynamic expansion graph model , author=. AAAI Conference on Artificial Intelligence (AAAI) , year=

  29. [37]

    IEEE Transactions on Audio, Speech and Language Processing , volume=

    An empirical study of catastrophic forgetting in large language models during continual fine-tuning , author=. IEEE Transactions on Audio, Speech and Language Processing , volume=

  30. [38]

    ACM Symposium on User Interface Software and Technoloy (UIST) , year=

    Generative agents: Interactive simulacra of human behavior , author=. ACM Symposium on User Interface Software and Technoloy (UIST) , year=

  31. [39]

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

  32. [40]

    Transactions on Machine Learning Research , volume=

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

  33. [41]

    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 , journal=

  34. [42]

    arXiv preprint arXiv:2409.07429 , year=

    Agent workflow memory , author=. arXiv preprint arXiv:2409.07429 , year=

    arXiv

  35. [43]

    arXiv preprint arXiv:2502.14802 , year=

    Guti. arXiv preprint arXiv:2502.14802 , year=

    arXiv

  36. [44]

    Learn to memorize: Optimizing

    Zhang, Zeyu and Dai, Quanyu and Li, Rui and Bo, Xiaohe and Chen, Xu and Dong, Zhenhua , journal=. Learn to memorize: Optimizing

  37. [45]

    arXiv preprint arXiv:2508.17862 , year=

    Retrieval feedback memory enhancement large model retrieval generation method , author=. arXiv preprint arXiv:2508.17862 , year=

    arXiv

  38. [46]

    Du, Xingbo and Li, Loka and Zhang, Duzhen and Song, Le , journal=

  39. [47]

    Cai, Zhicheng and Guo, Xinyuan and Pei, Yu and Feng, Jiangtao and Su, Jinsong and Chen, Jiangjie and Zhang, Ya-Qin and Ma, Wei-Ying and Wang, Mingxuan and Zhou, Hao , journal=

  40. [48]

    Liu, Jiaqi and Su, Yaofeng and Xia, Peng and Han, Siwei and Zheng, Zeyu and Xie, Cihang and Ding, Mingyu and Yao, Huaxiu , journal=

  41. [49]

    International Conference on Learning Representations (ICLR) , year=

    Benchmark Reduction with In-Context Transferability , author=. International Conference on Learning Representations (ICLR) , year=

  42. [50]

    Yang, Yutao and Li, Junsong and Pan, Qianjun and Zhan, Bihao and Cai, Yuxuan and Du, Lin and Zhou, Jie and Chen, Kai and Chen, Qin and Li, Xin and others , journal=

  43. [51]

    Ma, Ziyu and Yang, Shidong and Ji, Yuxiang and Wang, Xucong and Wang, Yong and Hu, Yiming and Huang, Tongwen and Chu, Xiangxiang , journal=

  44. [52]

    Annual Meeting of the Association for Computational Linguistics (ACL) , year=

    Agentic episodic control , author=. Annual Meeting of the Association for Computational Linguistics (ACL) , year=

  45. [53]

    Xia, Peng and Chen, Jianwen and Wang, Hanyang and Liu, Jiaqi and Zeng, Kaide and Wang, Yu and Han, Siwei and Zhou, Yiyang and Zhao, Xujiang and Chen, Haifeng and others , journal=

  46. [54]

    Alzubi, Salaheddin and Provenzano, Noah and Bingham, Jaydon and Chen, Weiyuan and Vu, Tu , journal=

  47. [55]

    Zhang, Haozhen and Long, Quanyu and Bao, Jianzhu and Feng, Tao and Zhang, Weizhi and Yue, Haodong and Wang, Wenya , journal=

  48. [56]

    Zhuo, Terry Yue and Chien, Vu Minh and Chim, Jenny and Hu, Han and Yu, Wenhao and Widyasari, Ratnadira and Yusuf, Imam Nur Bani and Zhan, Haolan and He, Junda and Paul, Indraneil and others , booktitle=

  49. [57]

    Benchmarking

    Yu, Peijie and Liu, Wei and Yang, Yifan and Li, Jinjian and Zhang, Zelong and Feng, Xiao and others , booktitle=. Benchmarking

  50. [58]

    Evaluating memory in

    Hu, Yuanzhe and Wang, Yu and McAuley, Julian , journal=. Evaluating memory in

  51. [59]

    arXiv preprint arXiv:2606.05661 , year=

    Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments , author=. arXiv preprint arXiv:2606.05661 , year=

    arXiv

  52. [60]

    Zhu, Deyao and Zhou, Xin and Qin, Shengling and Zhu, Xuekai and Ding, Hangliang and Zhong, Shu and Wen, Zixin and Xie, Zhonglin and Gou, Chenhui and Ren, Linxuan and others , journal=

  53. [61]

    Zhao, Yujie and Yuan, Boqin and Huang, Junbo and Yuan, Haocheng and Yu, Zhongming and Xu, Haozhou and Hu, Lanxiang and Shankarampeta, Abhilash and Huang, Zimeng and Ni, Wentao and Tian, Yuandong and Zhao, Jishen , journal=

  54. [62]

    Do Self-Evolving Agents Forget? Capability Degradation and Preservation in Lifelong

    Yu, Ye and Yuan, Xiaopeng and Jin, Haibo and Liu, Heming and Yu, Yaoning and Wang, Haohan , journal=. Do Self-Evolving Agents Forget? Capability Degradation and Preservation in Lifelong

  55. [63]

    Xu, Jundong and Li, Qingchuan and Wu, Jiaying and Lan, Yihuai and Li, Shuyue Stella and Zhou, Huichi and Jiang, Bowen and Wang, Lei and Wang, Jun and Luu, Anh Tuan and Xiong, Caiming and Park, Hae Won and Hooi, Bryan and Hu, Zhiyuan , journal=

  56. [64]

    Jianlv Chen and Shitao Xiao and Peitian Zhang and Kun Luo and Defu Lian and Zheng Liu , booktitle=

  57. [65]

    arXiv preprint arXiv:2606.19348 , year=

    Deepseek-v4: Towards highly efficient million-token context intelligence , author=. arXiv preprint arXiv:2606.19348 , year=

  58. [66]

    2026 , howpublished =

    Introducing. 2026 , howpublished =

    2026

  59. [67]

    Zeng, Aohan and Lv, Xin and Hou, Zhenyu and Du, Zhengxiao and Zheng, Qinkai and Chen, Bin and Yin, Da and Ge, Chendi and Huang, Chenghua and Xie, Chengxing and others , journal=

  60. [68]

    arXiv preprint arXiv:2205.13323 , year=

    The effect of task ordering in continual learning , author=. arXiv preprint arXiv:2205.13323 , year=

    arXiv

  61. [69]

    Conference on Lifelong Learning Agents (CoLLAs) , year=

    A theory for knowledge transfer in continual learning , author=. Conference on Lifelong Learning Agents (CoLLAs) , year=

  62. [70]

    International Conference on Machine Learning (ICML) , year=

    Reward translation via reward machine in semi-alignable MDPs , author=. International Conference on Machine Learning (ICML) , year=

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.