pith. machine review for the scientific record. sign in

arxiv: 2605.09315 · v1 · submitted 2026-05-10 · 💻 cs.AI · cs.CL

Recognition: 2 theorem links

· Lean Theorem

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

Ye Yu , Xiaopeng Yuan , Haibo Jin , Heming Liu , Yaoning Yu , Haohan Wang
Authors on Pith no claims yet

Pith reviewed 2026-05-12 04:07 UTC · model grok-4.3

classification 💻 cs.AI cs.CL
keywords self-evolving agentsLLM agentscapability degradationlifelong adaptationcontinual learningcapability preservationworkflow evolutionagent forgetting
0
0 comments X

The pith

Self-evolving LLM agents lose old capabilities as they adapt to new tasks across workflow, skill, model, and memory channels.

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

The paper establishes that self-evolution in LLM agents is non-monotonic: adapting to new task distributions progressively degrades previously acquired capabilities in all major evolution channels. This erosion appears consistently whether agents refine workflows, accumulate skills, update underlying models, or maintain memory. The authors introduce Capability-Preserving Evolution as a general stabilization principle that constrains destructive drift while still permitting adaptation gains. A sympathetic reader would care because reliable lifelong agents cannot be built on acquisition alone; they also require explicit mechanisms to retain what they already know.

Core claim

Self-evolution in LLM agents is often non-monotonic, with adaptation to new task distributions causing progressive degradation of prior capabilities across workflow, skill, model, and memory evolution channels. This phenomenon, called capability erosion under self-evolution, is addressed by Capability-Preserving Evolution (CPE), which improves retained capability stability while preserving adaptation performance. For instance, in workflow evolution under GPT-5.1 optimization, CPE raises retained simple-task performance from 41.8% to 52.8% while strengthening complex-task adaptation.

What carries the argument

Capability-Preserving Evolution (CPE), a stabilization principle that constrains destructive capability drift during continual adaptation of self-evolving LLM agents.

If this is right

  • Long-horizon self-evolving agents require explicit preservation of old capabilities in addition to acquiring new ones.
  • Without stabilization, adaptation to new tasks will progressively reduce performance on previously mastered tasks.
  • CPE applies uniformly across workflow, skill, model, and memory evolution to maintain stability during continual adaptation.
  • Practical agent systems must incorporate constraints against capability drift to achieve reliable lifelong operation.

Where Pith is reading between the lines

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

  • The erosion pattern may resemble catastrophic forgetting in continual learning, suggesting agent designers could borrow regularization techniques from that literature.
  • Developers should evaluate agents on multi-distribution benchmarks that track both new-task gains and old-task retention over extended horizons.
  • If CPE generalizes, it could enable safer deployment of autonomous agents in environments where forgetting carries real costs, such as personal assistants or scientific workflows.
  • The effect might intensify with longer adaptation sequences or more divergent task distributions, pointing to the need for scale and diversity experiments.

Load-bearing premise

The observed degradation stems from the self-evolution process itself rather than from particular implementation choices, task distributions, or model scale.

What would settle it

Running the same self-evolution protocols on a different model family or unrelated task distribution and finding no progressive loss of old capabilities would falsify the claim that erosion is inherent to self-evolution.

Figures

Figures reproduced from arXiv: 2605.09315 by Haibo Jin, Haohan Wang, Heming Liu, Xiaopeng Yuan, Yaoning Yu, Ye Yu.

Figure 1
Figure 1. Figure 1: Self-evolution under task distribution shift. Under unconstrained self-evolution, repeated [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Behavioral policy drift during complex-task workflow self-evolution. Starting from the [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Generalization erosion during skill self-evolution. As the task distribution shifts from [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Generalization erosion during skill self-evolution. Vanilla self-evolution and CPE undergo [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sequential self-evolution per￾formance on the Anatomy evaluation set using Qwen3-0.6B. Setup. We study model self-evolution using STaR self￾training [13] on MedMCQA [29] with Qwen3-0.6B [30] and Llama3.2-3B [31]. The mutable repository is the trainable parameter space of the evolving language model adapter. At each stage, the model generates rationale￾answer traces on a newly introduced medical domain, fil… view at source ↗
Figure 6
Figure 6. Figure 6: Memory-level interference under sequential memory evolution. We compare retained old-task performance before and after exposing the memory repository to new-task updates. Although the underlying model remains fixed, naïve memory evolution consistently degrades old-task retention after continual memory updates. CPE mitigates this interference by stabilizing historically reliable memories during long-horizon… view at source ↗
Figure 7
Figure 7. Figure 7: Full workflow evolution results across backbones. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Full skill/tool evolution results across backbones. [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Full Model evolution results across backbones. [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
read the original abstract

Recent advances in LLM agents enable systems that autonomously refine workflows, accumulate reusable skills, self-train their underlying models, and maintain persistent memory. However, we show that such self-evolution is often non-monotonic: adapting to new task distributions can progressively degrade previously acquired capabilities across all major evolution channels. We identify this phenomenon as \emph{capability erosion under self-evolution} and show that it consistently emerges across workflow, skill, model, and memory evolution. To mitigate this issue, we propose \emph{Capability-Preserving Evolution} (CPE), a general stabilization principle that constrains destructive capability drift during continual adaptation. Across all four evolution dimensions, CPE consistently improves retained capability stability while preserving adaptation performance. For example, in workflow evolution, CPE improves retained simple-task performance from 41.8\% to 52.8\% under GPT-5.1 optimization while simultaneously achieving stronger complex-task adaptation. Our findings suggest that stable long-horizon self-evolving agents require not only acquiring new capabilities, but also explicitly preserving previously learned ones during continual adaptation.

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 claims that self-evolving LLM agents, which autonomously refine workflows, accumulate skills, self-train models, and update memory, exhibit non-monotonic capability degradation (termed 'capability erosion under self-evolution') when adapting to new task distributions. This degradation occurs across workflow, skill, model, and memory channels. The authors propose Capability-Preserving Evolution (CPE) as a stabilization principle that constrains destructive drift, reporting consistent improvements in retained capability (e.g., simple-task performance rising from 41.8% to 52.8% in workflow evolution under GPT-5.1) while preserving adaptation gains.

Significance. If the central empirical observations hold after addressing controls, the work identifies a practically important obstacle for long-horizon autonomous agents and supplies a concrete mitigation principle applicable across evolution channels. The cross-channel consistency and the specific numeric example of retention improvement provide a useful starting point for future agent design, though the absence of isolating baselines limits immediate impact.

major comments (2)
  1. [Results on workflow evolution (and parallel sections for other channels)] The attribution of degradation specifically to self-evolution mechanisms (workflow refinement, skill accumulation, etc.) rather than generic continual adaptation effects is not load-bearing supported. The reported 41.8% simple-task retention (improving to 52.8% with CPE) is shown only within the self-evolving GPT-5.1 setup; no non-self-evolving baseline receiving identical task sequences is presented to test whether the non-monotonic loss occurs under standard fine-tuning or prompting on the same distributions.
  2. [CPE definition and evaluation sections] CPE is presented as a 'general stabilization principle,' yet all quantitative evidence is confined to the GPT-5.1 workflow case. No ablation or transfer results are given for other models, scales, or the skill/model/memory channels, undermining the claim of consistent improvement 'across all four evolution dimensions.'
minor comments (2)
  1. [Abstract] The abstract states high-level results and one numeric example but omits experimental details such as number of runs, statistical tests, task distribution descriptions, or exact measurement protocols for 'retained capability.'
  2. [Related work] References to the broader continual learning and catastrophic forgetting literature are sparse; explicit positioning against standard mitigation techniques (e.g., replay buffers, regularization) would clarify novelty.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive review. We address each major comment below, clarifying our experimental design and outlining revisions to strengthen the empirical support for our claims.

read point-by-point responses

  1. Referee: [Results on workflow evolution (and parallel sections for other channels)] The attribution of degradation specifically to self-evolution mechanisms (workflow refinement, skill accumulation, etc.) rather than generic continual adaptation effects is not load-bearing supported. The reported 41.8% simple-task retention (improving to 52.8% with CPE) is shown only within the self-evolving GPT-5.1 setup; no non-self-evolving baseline receiving identical task sequences is presented to test whether the non-monotonic loss occurs under standard fine-tuning or prompting on the same distributions.

    Authors: We agree that including non-self-evolving baselines on identical task sequences would provide stronger isolation of the effect. Our experiments are centered on the self-evolving paradigm because that is the novel setting under study, where autonomous mechanisms such as workflow refinement and skill accumulation introduce distinct dynamics not present in standard fine-tuning or prompting. To address the concern directly, we will add comparative baselines using standard continual adaptation methods on the same distributions in the revised manuscript, allowing explicit measurement of whether non-monotonic degradation is amplified under self-evolution. revision: yes

  2. Referee: [CPE definition and evaluation sections] CPE is presented as a 'general stabilization principle,' yet all quantitative evidence is confined to the GPT-5.1 workflow case. No ablation or transfer results are given for other models, scales, or the skill/model/memory channels, undermining the claim of consistent improvement 'across all four evolution dimensions.'

    Authors: The primary quantitative evaluation of CPE is presented in depth for the workflow channel under GPT-5.1 as a representative case study. The manuscript applies the same stabilization principle to the skill, model, and memory channels and reports consistent directional improvements, though these are less extensively quantified. We acknowledge that broader ablations across models, scales, and dedicated transfer experiments would better substantiate the generality claim. In the revision we will expand the evaluation sections to include additional quantitative results and ablations for the skill and model channels. revision: yes

Circularity Check

0 steps flagged

Empirical observation with no circular derivation chain

full rationale

The paper reports experimental observations of capability degradation across workflow, skill, model, and memory evolution channels in LLM agents, then proposes CPE as a mitigation principle. No equations, derivations, or first-principles claims appear; results are presented as direct measurements (e.g., retention percentages under GPT-5.1) rather than predictions derived from fitted parameters or self-referential definitions. The work is self-contained against external benchmarks via controlled experiments and does not rely on load-bearing self-citations or uniqueness theorems. This is the expected non-circular outcome for an empirical study.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Abstract-only review limits visibility into exact assumptions; the work rests on the premise that self-evolution occurs along the four listed channels and that degradation is measurable and mitigable.

axioms (1)
  • domain assumption LLM agents evolve autonomously along workflow, skill, model, and memory dimensions
    Stated as the setup for self-evolving agents in the abstract.
invented entities (1)
  • Capability-Preserving Evolution (CPE) no independent evidence
    purpose: Constrain destructive capability drift during continual adaptation
    Newly proposed stabilization principle without external validation shown in abstract.

pith-pipeline@v0.9.0 · 5501 in / 1120 out tokens · 42186 ms · 2026-05-12T04:07:30.024950+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages · 13 internal anchors

  1. [1]

    A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence

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

    work page internal anchor Pith review arXiv 2026

  2. [2]

    Automated Design of Agentic Systems

    Shengran Hu, Cong Lu, and Jeff Clune. Automated design of agentic systems, 2025. URL https://arxiv.org/abs/2408.08435

    work page internal anchor Pith review arXiv 2025

  3. [3]

    AFlow: Automating Agentic Workflow Generation

    Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xionghui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, Bingnan Zheng, Bang Liu, Yuyu Luo, and Chenglin Wu. Aflow: Automating agentic workflow generation, 2025. URL https://arxiv.org/abs/ 2410.10762

    work page internal anchor Pith review arXiv 2025

  4. [4]

    Evoagentx: An automated frame- work for evolving agentic workflows, 2025

    Yingxu Wang, Siwei Liu, Jinyuan Fang, and Zaiqiao Meng. Evoagentx: An automated frame- work for evolving agentic workflows, 2025. URLhttps://arxiv.org/abs/2507.03616

    work page arXiv 2025

  5. [5]

    arXiv preprint arXiv:2502.07373 , year=

    Guibin Zhang, Kaijie Chen, Guancheng Wan, Heng Chang, Hong Cheng, Kun Wang, Shuyue Hu, and Lei Bai. Evoflow: Evolving diverse agentic workflows on the fly, 2025. URL https://arxiv.org/abs/2502.07373

    work page arXiv 2025

  6. [6]

    SEW: Self-Evolving Agentic Workflows for Automated Code Generation

    Siwei Liu, Jinyuan Fang, Han Zhou, Yingxu Wang, and Zaiqiao Meng. Sew: Self-evolving agentic workflows for automated code generation, 2026. URL https://arxiv.org/abs/ 2505.18646

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  7. [7]

    Rossi, Handong Zhao, Ruiyi Zhang, Puneet Mathur, Nedim Lipka, Yu Wang, Trung Bui, Franck Dernoncourt, and Tianyi Zhou

    Dang Nguyen, Viet Dac Lai, Seunghyun Yoon, Ryan A. Rossi, Handong Zhao, Ruiyi Zhang, Puneet Mathur, Nedim Lipka, Yu Wang, Trung Bui, Franck Dernoncourt, and Tianyi Zhou. Dynasaur: Large language agents beyond predefined actions, 2025. URL https://arxiv. org/abs/2411.01747

    work page arXiv 2025

  8. [8]

    MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents

    Haozhen Zhang, Quanyu Long, Jianzhu Bao, Tao Feng, Weizhi Zhang, Haodong Yue, and Wenya Wang. Memskill: Learning and evolving memory skills for self-evolving agents, 2026. URLhttps://arxiv.org/abs/2602.02474

    work page internal anchor Pith review arXiv 2026

  9. [9]

    V oyager: An open-ended embodied agent with large language models,

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. V oyager: An open-ended embodied agent with large language models,

    work page

  10. [10]

    URLhttps://arxiv.org/abs/2305.16291

    work page internal anchor Pith review Pith/arXiv arXiv

  11. [11]

    Expel: Llm agents are experiential learners

    Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. Expel: Llm agents are experiential learners, 2024. URLhttps://arxiv.org/abs/2308.10144

    work page arXiv 2024

  12. [12]

    Tool-R0: Self-evolving LLM agents for tool-learning from zero data.arXiv, 2026

    Emre Can Acikgoz, Cheng Qian, Jonas Hübotter, Heng Ji, Dilek Hakkani-Tür, and Gokhan Tur. Tool-r0: Self-evolving llm agents for tool-learning from zero data, 2026. URL https: //arxiv.org/abs/2602.21320

    work page arXiv 2026

  13. [13]

    Skillcraft: Can llm agents learn to use tools skillfully?,

    Shiqi Chen, Jingze Gai, Ruochen Zhou, Jinghan Zhang, Tongyao Zhu, Junlong Li, Kangrui Wang, Zihan Wang, Zhengyu Chen, Klara Kaleb, Ning Miao, Siyang Gao, Cong Lu, Manling Li, Junxian He, and Yee Whye Teh. Skillcraft: Can llm agents learn to use tools skillfully?,

    work page

  14. [14]

    URLhttps://arxiv.org/abs/2603.00718

    work page arXiv

  15. [15]

    Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman. Star: Bootstrapping reasoning with reasoning, 2022. URLhttps://arxiv.org/abs/2203.14465

    work page arXiv 2022

  16. [16]

    B-STaR: Monitoring and Balancing Exploration and Exploitation in Self-Taught Reasoners,

    Weihao Zeng, Yuzhen Huang, Lulu Zhao, Yijun Wang, Zifei Shan, and Junxian He. B-star: Monitoring and balancing exploration and exploitation in self-taught reasoners, 2025. URL https://arxiv.org/abs/2412.17256

    work page arXiv 2025

  17. [17]

    arXiv preprint arXiv:2404.12253 , year=

    Ye Tian, Baolin Peng, Linfeng Song, Lifeng Jin, Dian Yu, Haitao Mi, and Dong Yu. Toward self-improvement of llms via imagination, searching, and criticizing, 2024. URL https: //arxiv.org/abs/2404.12253. 10

    work page arXiv 2024

  18. [18]

    Dynamic Cheatsheet: Test-Time Learning with Adaptive Memory , journal =

    Mirac Suzgun, Mert Yuksekgonul, Federico Bianchi, Dan Jurafsky, and James Zou. Dynamic cheatsheet: Test-time learning with adaptive memory, 2025. URL https://arxiv.org/abs/ 2504.07952

    work page arXiv 2025

  19. [19]

    Evo-Memory: Benchmarking LLM Agent Test-time Learning with Self-Evolving Memory

    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, 2025. URLhttps://arxiv.org/abs/2511.20857

    work page internal anchor Pith review arXiv 2025

  20. [20]

    Overcoming catastrophic forgetting in neural networks.Proceedings of the National Academy of Sciences, 114(13): 3521–3526, 2017

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catas- trophic forgetting in neural networks.Proceedings of the National Academy of Sciences, 114(13):352...

    work page doi:10.1073/pnas.1611835114 2017

  21. [21]

    A comprehensive survey of continual learning: Theory, method and application, 2024

    Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application, 2024. URL https://arxiv.org/abs/2302. 00487

    work page 2024

  22. [22]

    Gao et al

    Omer Marcus et al. SWE-Bench-CL: Continual learning for coding agents.arXiv preprint arXiv:2507.00014, 2025

    work page arXiv 2025

  23. [23]

    SuRe: Surprise-driven prioritised replay for continual LLM learning

    Zafeirios Fountas et al. SuRe: Surprise-driven prioritised replay for continual LLM learning. arXiv preprint arXiv:2511.22367, 2025

    work page arXiv 2025

  24. [24]

    Analyzing and reducing catastrophic forgetting in parameter efficient tuning.arXiv preprint arXiv:2402.18865, 2024

    Ming Chen et al. Analyzing and reducing catastrophic forgetting in parameter efficient tuning. arXiv preprint arXiv:2402.18865, 2024

    work page arXiv 2024

  25. [25]

    Mechanistic analysis of catastrophic forgetting in large language models during continual fine-tuning.arXiv preprint arXiv:2601.18699, 2026

    Wei Zhang et al. Mechanistic analysis of catastrophic forgetting in large language models during continual fine-tuning.arXiv preprint arXiv:2601.18699, 2026

    work page arXiv 2026

  26. [26]

    An empirical study of catastrophic forgetting in large language models during continual fine- tuning, 2025

    Yuhao Luo et al. An empirical study of catastrophic forgetting in large language models during continual fine-tuning.arXiv preprint arXiv:2308.08747, 2024

    work page arXiv 2024

  27. [27]

    $\tau^2$-Bench: Evaluating Conversational Agents in a Dual-Control Environment

    Victor Barres, Honghua Dong, Soham Ray, Xujie Si, and Karthik Narasimhan. τ 2-bench: Evaluating conversational agents in a dual-control environment, 2025. URL https://arxiv. org/abs/2506.07982

    work page internal anchor Pith review arXiv 2025

  28. [28]

    Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, Akshay Nathan, Alan Luo, Alec Helyar, Aleksander Madry, Aleksandr Efremov, Aleksandra Spyra, Alex Baker-Whitcomb, Alex Beutel, Alex Karpenko, Alex Makelov, Alex Neitz, Alex Wei, Alexandra Barr, Alexandre Kirchmeyer, Ale...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  29. [29]

    OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Flo- rencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, 12 Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgum, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher B...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  30. [30]

    Measuring mathematical problem solving with the math dataset,

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset,

    work page

  31. [31]

    URLhttps://arxiv.org/abs/2103.03874

    work page internal anchor Pith review Pith/arXiv arXiv

  32. [32]

    Medmcqa : A large-scale multi-subject multi-choice dataset for medical domain question answering

    Ankit Pal, Logesh Kumar Umapathi, and Malaikannan Sankarasubbu. Medmcqa : A large- scale multi-subject multi-choice dataset for medical domain question answering, 2022. URL https://arxiv.org/abs/2203.14371

    work page arXiv 2022

  33. [33]

    Qwen3 Technical Report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, ...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [34]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ah- mad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  35. [35]

    Revisiting weight regularization for low-rank continual learning, 2026

    Yaoyue Zheng, Yin Zhang, Joost van de Weijer, Gido M van de Ven, Shaoyi Du, Xuetao Zhang, and Zhiqiang Tian. Revisiting weight regularization for low-rank continual learning, 2026. URL https://arxiv.org/abs/2602.17559

    work page arXiv 2026

  36. [36]

    and Peng, Y

    Jinyuan Fang, Yanwen Peng, Xi Zhang, Yingxu Wang, Xinhao Yi, Guibin Zhang, Yi Xu, Bin Wu, Siwei Liu, Zihao Li, Zhaochun Ren, Nikos Aletras, Xi Wang, Han Zhou, and Zaiqiao Meng. A comprehensive survey of self-evolving ai agents: A new paradigm bridging foundation models and lifelong agentic systems, 2025. URLhttps://arxiv.org/abs/2508.07407

    work page arXiv 2025

  37. [37]

    Your agent may misevolve: Emergent risks in self-evolving llm agents.arXiv preprint arXiv:2509.26354, 2025

    Shuai Shao, Qihan Ren, Chen Qian, Boyi Wei, Dadi Guo, Jingyi Yang, Xinhao Song, Linfeng Zhang, Weinan Zhang, Dongrui Liu, and Jing Shao. Your agent may misevolve: Emergent risks in self-evolving llm agents, 2026. URLhttps://arxiv.org/abs/2509.26354

    work page arXiv 2026

  38. [38]

    decisions

    Yujun Zhou, Yue Huang, Han Bao, Kehan Guo, Zhenwen Liang, Pin-Yu Chen, Tian Gao, Werner Geyer, Nuno Moniz, Nitesh V Chawla, and Xiangliang Zhang. Capability-oriented training induced alignment risk, 2026. URLhttps://arxiv.org/abs/2602.12124. 15 A Proofs for the Local CPE Analysis Proof of Proposition 1.By Taylor expansion ofL <t aroundR t−1, for a local u...

    work page arXiv 2026