pith. sign in

arxiv: 2606.05661 · v1 · pith:C5CXNYZRnew · submitted 2026-06-04 · 💻 cs.AI · cs.CL

Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments

Parth Asawa , Christopher M. Glaze , Gabriel Orlanski , Ramya Ramakrishnan , Benji Xu , Asim Biswal , Vincent Sunn Chen , Frederic Sala
show 2 more authors
Matei Zaharia Joseph E. Gonzalez
This is my paper

Pith reviewed 2026-06-28 01:43 UTC · model grok-4.3

classification 💻 cs.AI cs.CL
keywords continual learningLLM agentsbenchmarkin-context learningmemory systemsstateful environmentslatent structureexpert-validated tasks
0
0 comments X

The pith

Dedicated memory systems do not improve continual learning in LLM agents; naive in-context learning outperforms them on a new multi-domain benchmark.

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

The paper introduces CL-Bench, the first expert-validated benchmark for measuring genuine improvement through sequential experience across six real-world domains where tasks share discoverable latent structures. It tests frontier models in agent setups ranging from simple in-context learning to dedicated memory modules, using a gain metric that separates learning from base capabilities. The central finding is that agents tend to overfit to recent observations or fail to carry knowledge forward, and adding memory systems does not solve this—in fact, plain in-context learning performs better. This indicates current architectures leave substantial room for systems that can reliably extract and reuse latent patterns online.

Core claim

CL-Bench shows that frontier LLM-based agents do not reliably improve through sequential experience even when tasks contain shared latent structures such as codebase layouts or opponent strategies; agents overfit to immediate observations or fail to reuse prior knowledge, and dedicated memory systems do not remedy this limitation—instead, naive in-context learning yields higher gains.

What carries the argument

CL-Bench benchmark with its gain metric, which measures performance improvement across task instances after isolating prior model capability, applied to domains with shared learnable latent structures.

If this is right

  • Current agent architectures leave measurable headroom for better continual learning.
  • Adding dedicated memory modules does not address overfitting or knowledge-reuse failures.
  • Naive in-context learning remains the strongest performer among the tested setups.
  • Future systems need mechanisms to discover and apply latent structures across instances rather than relying on raw memory storage.

Where Pith is reading between the lines

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

  • The results suggest that improvements may require changes in how experience is encoded during inference rather than just expanding storage capacity.
  • Benchmarks of this type could be extended to measure whether models can be prompted or trained to explicitly search for cross-instance patterns.
  • In practical deployments, simpler prompting strategies might currently deliver more reliable adaptation than complex memory architectures.
  • The gap between naive and memory-based approaches points to a need for evaluation methods that test transfer even when surface features differ between instances.

Load-bearing premise

The benchmark tasks are constructed so they share a discoverable latent structure that a stateful system can learn online but a stateless one cannot.

What would settle it

If memory-augmented agents consistently produce higher gain scores than naive in-context learning across the six domains while showing clear reuse of earlier task knowledge rather than overfitting.

Figures

Figures reproduced from arXiv: 2606.05661 by Asim Biswal, Benji Xu, Christopher M. Glaze, Frederic Sala, Gabriel Orlanski, Joseph E. Gonzalez, Matei Zaharia, Parth Asawa, Ramya Ramakrishnan, Vincent Sunn Chen.

Figure 1
Figure 1. Figure 1: The CL-BENCH evaluation framework, illustrated via a simplified Database Ex￾ploration task. In each CL-BENCH task, an agent completes a sequence of instances in a shared environment, accumulating past experience and feedback to improve over time. Tasks contain ex￾ploitable latent structure not known a priori—here, obfuscated schema conventions—which a learning agent needs to discover to improve reward (e.g… view at source ↗
Figure 2
Figure 2. Figure 2: CL-BENCH task construction and human validation pipeline. Each task is constructed against the 3 design criteria and reviewed by at least two authors, then independently reviewed by 2–3 domain experts on three axes (realism, reusable knowledge, and learning improvement) with tasks accepted into the benchmark only after confirming validity on all axes. After a task is proposed and meets the above criteria, … view at source ↗
Figure 3
Figure 3. Figure 3: summarizes the task sets introduced in CL-BENCH. Full descriptions, instance examples, variant definitions, and further curation details are provided in Appendix A. The tasks span six domains—software engineering, database analytics, epidemiological forecasting, RF spectrum monitoring, sales forecasting, and strategic gaming—with reward types ranging from step efficiency and prediction accuracy to signal q… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the gain metric. We compute the difference between a system’s stateful performance on task instances vs. its performance if it had seen each instance independently. This allows us to isolate the performance improvement due to learning from prior experience. For each instance t, we define gain as: gt = r sf t −r sl t where r sf t is the reward achieved by the stateful system and r sl t is the re… view at source ↗
Figure 5
Figure 5. Figure 5: Learning curves across all six tasks for the top-ranked system (ICL + Claude Sonnet 4.6). Solid lines show mean reward per instance when stateful; dashed lines show the same system run statelessly. The gap between lines isolates the contribution of learning from experience to performance. Sales Prediction and BSM show the largest and earliest-forming gaps; whereas Database Exploration is notable for the fr… view at source ↗
Figure 6
Figure 6. Figure 6: Aggregate Pareto views of the public CL-BENCH leaderboard. We compare the normalized reward against the rollout cost and the normalized gain against the rollout cost. Black lines connect configurations not dominated by another configuration on the plotted axes. Model names that start with solely G are shorthand for Gemini models. reward frequently collapses to zero while stateful maintains a floor, so gain… view at source ↗
Figure 7
Figure 7. Figure 7: Decomposition of normalized gain into stability and plasticity terms. Normalized gain for each system and agent decomposed into (1) a stability term that measures how effective the agent is at re-using previously acquired information and (2) a plasticity term that measures how effective the agent is at incorporating new information to perform tasks. with Claude Code providing the one clear exception: it ac… view at source ↗
Figure 8
Figure 8. Figure 8: shows per-system normalized reward across all six tasks under the same normalization as [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Blind Spectrum Monitoring: reward and gain vs. cost. Each point is one system configuration; black lines trace the Pareto frontier. Model name abbreviations follow the same convention as [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Blind Spectrum Monitoring: learning curve for the best system on this task (ICL + GPT-5.4). Solid line shows mean reward per instance in stateful mode; dashed line shows the same system run statelessly on identical instances. B.6 Codebase Adaptation [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Codebase Adaptation: reward and gain vs. cost. Each point is one system configuration; black lines trace the Pareto frontier. Model name abbreviations follow the same convention as [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Codebase Adaptation: learning curve for the best system on this task (ACE + GPT￾5.4). Solid line shows mean reward per instance in stateful mode; dashed line shows the same system run statelessly on identical instances. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Cohort Studies: reward and gain vs. cost. Each point is one system configuration; black lines trace the Pareto frontier. Model name abbreviations follow the same convention as [PITH_FULL_IMAGE:figures/full_fig_p025_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Cohort Studies: learning curve for the best system on this task (ICL + GPT-5.4). Solid line shows mean reward per instance in stateful mode; dashed line shows the same system run statelessly on identical instances. B.8 Database Exploration [PITH_FULL_IMAGE:figures/full_fig_p026_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Database Exploration: reward and gain vs. cost. Each point is one system configuration; black lines trace the Pareto frontier. Model name abbreviations follow the same convention as [PITH_FULL_IMAGE:figures/full_fig_p027_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Database Exploration: learning curve for the best system on this task (Claude Code + Sonnet 4.6). Solid line shows mean reward per instance in stateful mode; dashed line shows the same system run statelessly on identical instances. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Exploitable Poker: reward and gain vs. cost. Each point is one system configuration; black lines trace the Pareto frontier. Model name abbreviations follow the same convention as [PITH_FULL_IMAGE:figures/full_fig_p028_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Exploitable Poker: learning curve for the best system on this task (Claude Code + Sonnet 4.6). Solid line shows mean reward per instance in stateful mode; dashed line shows the same system run statelessly on identical instances. B.10 Sales Prediction [PITH_FULL_IMAGE:figures/full_fig_p029_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Sales Prediction: reward and gain vs. cost. Each point is one system configuration; black lines trace the Pareto frontier. Model name abbreviations follow the same convention as [PITH_FULL_IMAGE:figures/full_fig_p030_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Sales Prediction: learning curve for the best system on this task (ICL Notepad + Claude Sonnet 4.6). Solid line shows mean reward per instance in stateful mode; dashed line shows the same system run statelessly on identical instances. C Stability–Plasticity Decomposition of Gain The normalized gain gb in Section 4.3 captures how much a system learns over a run. We decomposed the measure into two terms rel… view at source ↗
Figure 21
Figure 21. Figure 21: Example of two learners with differing amounts of learning stability. Normalized gain over the blind spectrum monitoring task instances for two systems, one with 13.5% stability (Claude Opus 4.7 with ICL) and one with significantly lower stability at 9% of learning headroom (GPT-5.4 with Codex). Stability failure — Sales Prediction (Round 1, year 2035) Environment feedback after the previous round (step 4… view at source ↗
read the original abstract

Continual learning, the ability of AI systems to improve through sequential experience, has attracted substantial interest, but no high-quality benchmark exists to evaluate it. We introduce Continual Learning Bench (CL-Bench), the first difficult, expert-validated benchmark designed to measure whether LLM-based systems genuinely improve with experience. CL-Bench spans six diverse domains (software engineering, signal processing, disease outbreak forecasting, database querying, strategic game-playing, and demand forecasting), each validated by domain experts and designed so that tasks share a learnable latent structure (codebase layout, disease outbreak dynamics, opponent strategies) that a stateful system can discover online but a stateless one cannot. We evaluate frontier models across several agent architectures, from naive in-context learning (ICL) to dedicated memory systems, introducing a gain metric to isolate learning from prior capabilities. We find that these systems leave headroom for improved continual learning: agents frequently overfit to immediate observations or fail to reuse knowledge across instances, and dedicated memory systems do not fix this -- in fact, naive ICL outperforms systems dedicated to memory management. CL-Bench is the first benchmark to evaluate continual learning across diverse real-world domains with expert-validated tasks and isolate online learning from underlying model capability, showing a need for better continual learning systems.

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

0 major / 2 minor

Summary. The paper introduces Continual Learning Bench (CL-Bench), the first expert-validated benchmark spanning six real-world domains (software engineering, signal processing, disease outbreak forecasting, database querying, strategic game-playing, demand forecasting) designed to test whether LLM-based agents improve via sequential experience on tasks sharing discoverable latent structures. It compares agent architectures from naive in-context learning (ICL) to dedicated memory systems, introduces a gain metric to separate online learning from base capabilities, and reports that current systems overfit or fail to reuse knowledge, with dedicated memory systems underperforming naive ICL.

Significance. If the experimental controls and task constructions hold, the benchmark would provide a valuable, reproducible tool for measuring genuine continual learning progress in stateful settings, and the counterintuitive result that ICL outperforms specialized memory architectures would usefully redirect research attention toward architectures that better exploit latent structure across episodes.

minor comments (2)
  1. Clarify the exact formula and statistical controls for the gain metric (mentioned in the abstract) so readers can verify it isolates improvement from base model capability.
  2. Add an appendix or table listing the specific latent structures per domain and the expert validation protocol to support reproducibility claims.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive summary of the paper and for recommending minor revision. The referee's description accurately reflects the benchmark design, the gain metric, and the main empirical finding that naive ICL outperforms dedicated memory systems on CL-Bench. No major comments were enumerated in the report, so we have no point-by-point rebuttals to provide. We remain available to address any minor suggestions or clarifications the editor may request.

Circularity Check

0 steps flagged

No significant circularity in benchmark design or empirical claims

full rationale

The paper introduces CL-Bench as a new empirical benchmark across six domains with expert-validated tasks sharing latent structures. It evaluates agent architectures (including naive ICL vs. dedicated memory) using a gain metric to isolate improvement. No equations, derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The central findings (ICL outperforming memory systems, overfitting issues) follow directly from the task constructions and comparisons under identical backbones, without reducing to self-definition or prior author results by construction. This is a standard benchmark paper whose claims rest on new data collection and measurement rather than internal reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The benchmark's validity rests on the domain assumption that tasks contain discoverable latent structures; no free parameters or invented entities are described in the abstract.

axioms (1)
  • domain assumption Tasks in each domain share a learnable latent structure that stateful systems can discover online but stateless ones cannot.
    This premise is required for the benchmark to distinguish genuine continual learning from baseline capabilities.

pith-pipeline@v0.9.1-grok · 5792 in / 1216 out tokens · 40909 ms · 2026-06-28T01:43:18.952080+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

43 extracted references · 2 canonical work pages

  1. [1]

    Longhealth: A question answering benchmark with long clinical documents.arXiv preprint arXiv:2401.14490, 2024

    Lisa Adams, Felix Busch, Tianyu Han, Jean-Baptiste Excoffier, Matthieu Ortala, Alexander Löser, Hugo JWL Aerts, Jakob Nikolas Kather, Daniel Truhn, and Keno Bressem. Longhealth: A question answering benchmark with long clinical documents.arXiv preprint arXiv:2401.14490, 2024

    arXiv 2024

  2. [2]

    Memorybench: A benchmark for memory and continual learning in llm systems, 2026

    Qingyao Ai, Yichen Tang, Changyue Wang, Jianming Long, Weihang Su, and Yiqun Liu. Memorybench: A benchmark for memory and continual learning in llm systems, 2026. URL https://arxiv.org/abs/2510.17281

    Pith/arXiv arXiv 2026

  3. [3]

    Claude code.https://www.anthropic.com/product/claude-code, 2025

    Anthropic. Claude code.https://www.anthropic.com/product/claude-code, 2025

    2025

  4. [4]

    Introducing Claude Opus 4.7

    Anthropic. Introducing Claude Opus 4.7. https://www.anthropic.com/news/ claude-opus-4-7, April 2026

    2026

  5. [5]

    Introducing Claude Sonnet 4.6

    Anthropic. Introducing Claude Sonnet 4.6. https://www.anthropic.com/news/ claude-sonnet-4-6, February 2026

    2026

  6. [6]

    Dimakis, and Matei Zaharia

    Parth Asawa, Alexandros G. Dimakis, and Matei Zaharia. Sieve: Sample-efficient parametric learning from natural language, 2026. URLhttps://arxiv.org/abs/2604.02339

    Pith/arXiv arXiv 2026

  7. [7]

    Finance agent benchmark: Benchmarking llms on real-world financial research tasks.arXiv preprint arXiv:2508.00828, 2025

    Antoine Bigeard, Langston Nashold, Rayan Krishnan, and Shirley Wu. Finance agent benchmark: Benchmarking llms on real-world financial research tasks.arXiv preprint arXiv:2508.00828, 2025

    arXiv 2025

  8. [8]

    Mem0: Building production-ready ai agents with scalable long-term memory.arXiv preprint arXiv:2504.19413, 2025

    Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. Mem0: Building production-ready ai agents with scalable long-term memory.arXiv preprint arXiv:2504.19413, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    A continual learning survey: Defying forgetting in classification tasks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022

    Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. URL https://arxiv.org/abs/1909.08383. Comprehensive survey with stability-plasticity fram...

    arXiv 2022

  10. [10]

    van de Ven, and Tinne Tuytelaars

    Matthias De Lange, Gido M. van de Ven, and Tinne Tuytelaars. Continual evaluation for lifelong learning: Identifying the stability gap. InInternational Conference on Learning Representations (ICLR), 2023. URL https://openreview.net/forum?id=Zy350cRstc6. Per-iteration evaluation reveals transient forgetting invisible to task-boundary-only metrics

    2023

  11. [11]

    Swe-bench pro: Can ai agents solve long-horizon software engineering tasks?, 2025

    Xiang Deng, Jeff Da, Edwin Pan, Yannis Yiming He, Charles Ide, Kanak Garg, Niklas Lauffer, Andrew Park, Nitin Pasari, Chetan Rane, Karmini Sampath, Maya Krishnan, Srivatsa Kundurthy, Sean Hendryx, Zifan Wang, Vijay Bharadwaj, Jeff Holm, Raja Aluri, Chen Bo Calvin Zhang, Noah Jacobson, Bing Liu, and Brad Kenstler. Swe-bench pro: Can ai agents solve long-ho...

    Pith/arXiv arXiv 2025

  12. [12]

    Cartridges: Lightweight and general-purpose long context representations via self-study, 2025

    Sabri Eyuboglu, Ryan Ehrlich, Simran Arora, Neel Guha, Dylan Zinsley, Emily Liu, Will Tennien, Atri Rudra, James Zou, Azalia Mirhoseini, and Christopher Re. Cartridges: Lightweight and general-purpose long context representations via self-study, 2025. URL https://arxiv.org/abs/2506.06266

    arXiv 2025

  13. [13]

    Arc-agi-3: A new challenge for frontier agentic intelligence, 2026

    ARC Prize Foundation. Arc-agi-3: A new challenge for frontier agentic intelligence, 2026. URLhttps://arxiv.org/abs/2603.24621

    Pith/arXiv arXiv 2026

  14. [14]

    Introducing Gemini 3 Flash: Frontier intelligence built for speed

    Google DeepMind. Introducing Gemini 3 Flash: Frontier intelligence built for speed. https: //blog.google/products/gemini/gemini-3-flash/, December 2025

    2025

  15. [15]

    Gemini 3.1 Pro: A smarter model for your most com- plex tasks

    Google DeepMind. Gemini 3.1 Pro: A smarter model for your most com- plex tasks. https://blog.google/innovation-and-ai/models-and-research/ gemini-models/gemini-3-1-pro/, March 2026

    2026

  16. [16]

    Richard Hake. Interactive-engagement versus traditional methods: A six-thousand-student survey of mechanics test data for introductory physics courses.American Journal of Physics - AMER J PHYS, 66, 01 1998. doi: 10.1119/1.18809

    work page doi:10.1119/1.18809 1998

  17. [17]

    Bridging language and items for retrieval and recommendation.arXiv preprint arXiv:2403.03952, 2024

    Yupeng Hou, Jiacheng Li, Zhankui He, An Yan, Xiusi Chen, and Julian McAuley. Bridging language and items for retrieval and recommendation.arXiv preprint arXiv:2403.03952, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Reinforcement learning via self-distillation, 2026

    Jonas Hübotter, Frederike Lübeck, Lejs Behric, Anton Baumann, Marco Bagatella, Daniel Marta, Ido Hakimi, Idan Shenfeld, Thomas Kleine Buening, Carlos Guestrin, and Andreas Krause. Reinforcement learning via self-distillation, 2026. URL https://arxiv.org/abs/ 2601.20802

    Pith/arXiv arXiv 2026

  19. [19]

    Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan

    Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. Swe-bench: Can language models resolve real-world github issues?, 2024. URLhttps://arxiv.org/abs/2310.06770

    Pith/arXiv arXiv 2024

  20. [20]

    Swe-bench-cl: Continual learning for coding agents.arXiv preprint arXiv:2507.00014, 2025

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

    arXiv 2025

  21. [21]

    Continual learning via sparse memory finetuning, 2025

    Jessy Lin, Luke Zettlemoyer, Gargi Ghosh, Wen-Tau Yih, Aram Markosyan, Vincent-Pierre Berges, and Barlas O ˘guz. Continual learning via sparse memory finetuning, 2025. URL https://arxiv.org/abs/2510.15103

    arXiv 2025

  22. [22]

    Evaluating very long-term conversational memory of llm agents, 2024

    Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. Evaluating very long-term conversational memory of llm agents, 2024. URL https://arxiv.org/abs/2402.17753

    Pith/arXiv arXiv 2024

  23. [23]

    Merrill, Alexander G

    Mike A. Merrill, Alexander G. Shaw, Nicholas Carlini, Boxuan Li, Harsh Raj, Ivan Bercovich, Lin Shi, Jeong Yeon Shin, Thomas Walshe, E. Kelly Buchanan, Junhong Shen, Guanghao Ye, Haowei Lin, Jason Poulos, Maoyu Wang, Marianna Nezhurina, Jenia Jitsev, Di Lu, Orfeas Menis Mastromichalakis, Zhiwei Xu, Zizhao Chen, Yue Liu, Robert Zhang, Leon Liangyu Chen, An...

    Pith/arXiv arXiv 2026

  24. [24]

    Codex: A cloud-based software engineering agent

    OpenAI. Codex: A cloud-based software engineering agent. https://openai.com/index/ introducing-codex/, 2025

    2025

  25. [25]

    Introducing GPT-5.4

    OpenAI. Introducing GPT-5.4. https://openai.com/index/introducing-gpt-5-4/ , March 2026

    2026

  26. [26]

    Slopcodebench: Benchmarking how coding agents degrade over long-horizon iterative tasks.arXiv preprint arXiv:2603.24755, 2026

    Gabriel Orlanski, Devjeet Roy, Alexander Yun, Changho Shin, Alex Gu, Albert Ge, Dyah Adila, Nicholas Roberts, Frederic Sala, and Aws Albarghouthi. Slopcodebench: Benchmarking how coding agents degrade over long-horizon iterative tasks.arXiv preprint arXiv:2603.24755, 2026

    Pith/arXiv arXiv 2026

  27. [27]

    Patil, Ion Stoica, and Joseph E

    Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. Memgpt: Towards llms as operating systems, 2024. URL https: //arxiv.org/abs/2310.08560

    Pith/arXiv arXiv 2024

  28. [28]

    Posttrainbench: Can llm agents automate llm post-training? arXiv preprint arXiv:2603.08640, 2026

    Ben Rank, Hardik Bhatnagar, Ameya Prabhu, Shira Eisenberg, Karina Nguyen, Matthias Bethge, and Maksym Andriushchenko. Posttrainbench: Can llm agents automate llm post-training? arXiv preprint arXiv:2603.08640, 2026

    arXiv 2026

  29. [29]

    Self-distillation enables continual learning, 2026

    Idan Shenfeld, Mehul Damani, Jonas Hübotter, and Pulkit Agrawal. Self-distillation enables continual learning, 2026. URLhttps://arxiv.org/abs/2601.19897

    Pith/arXiv arXiv 2026

  30. [30]

    Learning to (learn at test time): Rnns with expressive hidden states, 2025

    Yu Sun, Xinhao Li, Karan Dalal, Jiarui Xu, Arjun Vikram, Genghan Zhang, Yann Dubois, Xinlei Chen, Xiaolong Wang, Sanmi Koyejo, Tatsunori Hashimoto, and Carlos Guestrin. Learning to (learn at test time): Rnns with expressive hidden states, 2025. URL https: //arxiv.org/abs/2407.04620

    Pith/arXiv arXiv 2025

  31. [31]

    End-to-end test-time training for long context, 2025

    Arnuv Tandon, Karan Dalal, Xinhao Li, Daniel Koceja, Marcel Rød, Sam Buchanan, Xiaolong Wang, Jure Leskovec, Sanmi Koyejo, Tatsunori Hashimoto, Carlos Guestrin, Jed McCaleb, Yejin Choi, and Yu Sun. End-to-end test-time training for long context, 2025. URL https: //arxiv.org/abs/2512.23675

    arXiv 2025

  32. [32]

    van de Ven, Tinne Tuytelaars, and Andreas S

    Gido M. van de Ven, Tinne Tuytelaars, and Andreas S. Tolias. Three types of incremental learning.Nature Machine Intelligence, 2022. URL https://www.nature.com/articles/ s42256-022-00568-3 . Canonical taxonomy: task-incremental, domain-incremental, class- incremental

    2022

  33. [33]

    van de Ven, Nicholas Soures, and Dhireesha Kudithipudi.Continual learn- ing and catastrophic forgetting, page 153–168

    Gido M. van de Ven, Nicholas Soures, and Dhireesha Kudithipudi.Continual learn- ing and catastrophic forgetting, page 153–168. Elsevier, 2025. ISBN 9780443157554. doi: 10.1016/b978-0-443-15754-7.00073-0. URL http://dx.doi.org/10.1016/ B978-0-443-15754-7.00073-0

    work page doi:10.1016/b978-0-443-15754-7.00073-0 2025

  34. [34]

    Browsecomp: A simple yet challenging benchmark for browsing agents.arXiv preprint arXiv: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.arXiv preprint arXiv:2504.12516, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Continual world: A robotic benchmark for continual reinforcement learning, 2021

    Maciej Wołczyk, Michał Zaj ˛ ac, Razvan Pascanu, Łukasz Kuci´nski, and Piotr Miło´s. Continual world: A robotic benchmark for continual reinforcement learning, 2021. URL https://arxiv. org/abs/2105.10919

    arXiv 2021

  36. [36]

    Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments, 2024

    Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, Yitao Liu, Yiheng Xu, Shuyan Zhou, Silvio Savarese, Caiming Xiong, Victor Zhong, and Tao Yu. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments, 2024. URL https: //arxiv.org/abs/2...

    Pith/arXiv arXiv 2024

  37. [37]

    Xu, Yufan Song, Boxuan Li, Yuxuan Tang, Kritanjali Jain, Mengxue Bao, Zora Z

    Frank F. Xu, Yufan Song, Boxuan Li, Yuxuan Tang, Kritanjali Jain, Mengxue Bao, Zora Z. Wang, Xuhui Zhou, Zhitong Guo, Murong Cao, Mingyang Yang, Hao Yang Lu, Amaad Martin, Zhe Su, Leander Maben, Raj Mehta, Wayne Chi, Lawrence Jang, Yiqing Xie, Shuyan Zhou, and Graham Neubig. Theagentcompany: Benchmarking llm agents on consequential real world tasks, 2025....

    Pith/arXiv arXiv 2025

  38. [38]

    τ-bench: A benchmark for tool-agent-user interaction in real-world domains.ArXiv, abs/2406.12045, 2024

    Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. τ-bench: A benchmark for tool-agent-user interaction in real-world domains.ArXiv, abs/2406.12045, 2024. URL https://api.semanticscholar.org/CorpusID:270562578

    Pith/arXiv arXiv 2024

  39. [39]

    Agentic context engineering: Evolving contexts for self-improving language models, 2026

    Qizheng Zhang, Changran Hu, Shubhangi Upasani, Boyuan Ma, Fenglu Hong, Vamsidhar Kamanuru, Jay Rainton, Chen Wu, Mengmeng Ji, Hanchen Li, Urmish Thakker, James Zou, and Kunle Olukotun. Agentic context engineering: Evolving contexts for self-improving language models, 2026. URLhttps://arxiv.org/abs/2510.04618

    Pith/arXiv arXiv 2026

  40. [40]

    Lifelongagentbench: Evaluating llm agents as lifelong learners, 2025

    Junhao Zheng, Xidi Cai, Qiuke Li, Duzhen Zhang, ZhongZhi Li, Yingying Zhang, Le Song, and Qianli Ma. Lifelongagentbench: Evaluating llm agents as lifelong learners, 2025. URL https://arxiv.org/abs/2505.11942

    arXiv 2025

  41. [41]

    Shanshan Zhong, Yi Lu, Jingjie Ning, Yibing Wan, Lihan Feng, Yuyi Ao, Leonardo F. R. Ribeiro, Markus Dreyer, Sean Ammirati, and Chenyan Xiong. Skilllearnbench: Benchmarking continual learning methods for agent skill generation on real-world tasks, 2026. URL https: //arxiv.org/abs/2604.20087

    Pith/arXiv arXiv 2026

  42. [42]

    Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig

    Shuyan Zhou, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig. Webarena: A realistic web environment for building autonomous agents, 2024. URL https://arxiv.org/abs/ 2307.13854

    Pith/arXiv arXiv 2024

  43. [43]

    transmitters

    Adam Zweiger, Jyothish Pari, Han Guo, Ekin Akyürek, Yoon Kim, and Pulkit Agrawal. Self- adapting language models, 2025. URLhttps://arxiv.org/abs/2506.10943. A Task Details This appendix describes each CL-BENCHtask in detail: its continual-learning framing, the informa- tion available to the agent each instance, an illustrative prompt–response exchange, th...

    arXiv 2025