Pith. sign in

REVIEW 5 major objections 7 minor 26 references

Pairing fast pattern-matching with slow, atomic fixes lets an LLM agent improve drifting ML models cheaply — while accepting only updates that raise measured accuracy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review

2026-08-04 08:49 UTC pith:AZ7JNTQX

load-bearing objection The dual-process + memory architecture is a genuine new entry for automated model maintenance, but the headline results are in-sample because the same drift sample is used both to select atomic changes and to compute final accuracy. the 5 major comments →

arxiv 2608.02351 v1 pith:AZ7JNTQX submitted 2026-08-03 cs.AI

KC-Agent: A Dual-Process Cognitive Architecture for Efficient ML Model Improvement

Gusseppe Bravo-Rocca , Jordi Guitart , Ajay Dholakia , David Ellison , Puneet Jain This is my paper
classification cs.AI
keywords data driftmodel maintenancedual-process architectureLLM agentsknowledge consolidationatomic changesmonotonic improvementautomated machine learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper argues that automated machine-learning model maintenance under data drift is best handled by a dual-process agent: a quick path that reuses previously successful fixes when the new drift resembles past scenarios, and a deliberate path that applies small, individually evaluated changes and rolls back failures. The central claim is that this combination beats single-strategy cognitive agents on accuracy (76.8% average across five datasets) while using an order of magnitude less computation and tokens than slow deliberation alone. A knowledge-consolidation mechanism transfers successful deliberate-path solutions into the fast path's memory, producing a 91% speedup over the slow-only variant. The paper also proves a monotonic-improvement property: because only changes that increase the measured score are kept, the agent's performance on the evaluation data never decreases. A sympathetic reader would care because it suggests a practical, verifiable way to automate model retraining in production without risking degradation.

Core claim

KC-Agent's central discovery is that the classic fast/slow split of human cognition maps cleanly onto the model-improvement problem: rapid pattern-based responses handle routine drift, and slow incremental reasoning handles hard cases — with a memory bridge that turns the deliberate path's successful strategies into the fast path's automatic reflexes. Implemented as an LLM agent over training code, the architecture selects improvements via a proxy threshold, applies atomic changes (small, targeted modifications that can be evaluated and reverted), and stores outcomes in semantic and episodic memory. On five datasets, including real turbofan degradation data, KC-Agent attains the highest aver

What carries the argument

The architecture pairs a fast graph (System 1) that matches incoming drift signatures against stored episodic memories using cosine similarity and replays known solutions, with a slow graph (System 2) that performs iterative atomic improvements — one hyperparameter, model choice, or preprocessing step at a time — evaluating each change and reverting failures. A threshold determines when a fast solution is accepted; otherwise the slow graph takes over. Three memory stores (semantic, episodic, working) carry the knowledge transfer, and an 'atomic change' norm constraint (each delta bounded) underlies the proof that accepted changes cannot decrease performance.

Load-bearing premise

The paper assumes that the accuracy measured on the new-data sample D_new — the same sample used to guide and evaluate every update — reflects genuine improvement on future drift; if D_new is not representative or is overfit, the performance and monotonicity claims can hold while real-world gains do not.

What would settle it

Split D_new chronologically or randomly into a decision set and a hold-out test set, then rerun the agent using only the decision set to guide atomic changes and the hold-out set to score final models. If KC-Agent's accuracy advantage over the strongest baselines shrinks or disappears on the hold-out, the conclusion that the architecture generalizes rather than overfits would be falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Automated model retraining can be both fast and safe: routine drift is handled in seconds, while hard cases trigger deliberate but still bounded effort.
  • Knowledge consolidation lets an agent get cheaper over time, as the fast path absorbs the slow path's successful strategies; the reported 91% speedup over slow-only is the direct consequence.
  • The monotonic-improvement guarantee, if trusted, means autonomous update loops can run without a human in the loop for every change, since failed changes are reverted by construction.
  • Architectures that reuse successful strategies across sessions (memory) outperform session-stateless agents, as shown by the gap over general-purpose code-action and other baselines.
  • A small LLM (around 8B parameters) suffices for the full loop, suggesting the approach is deployable on commodity hardware.

Where Pith is reading between the lines

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

  • The headline accuracy is measured on the same new-data sample (D_new) that guides each update; without a held-out split, the reported gains partly reflect fit to that specific drift sample, and the true generalization to future drift is likely smaller.
  • The monotonic guarantee is relative to the evaluation dataset, not to unseen data; it certifies that the agent never hurts the score it is optimizing, which is weaker than a guarantee about real-world performance.
  • An inexpensive next test would be a temporal holdout: use only the first portion of D_new for atomic-change decisions and the last portion for scoring, to see whether the 2.4% edge over the strongest baseline survives.
  • The proxy threshold is fixed across domains; making it self-tuning or dataset-adaptive would likely improve both acceptance quality and safety.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 7 minor

Summary. The paper presents KC-Agent, a dual-process (System 1/System 2) LLM agent architecture for automated ML model improvement under data drift. System 1 performs fast pattern-based responses by reusing stored successful strategies from semantic/episodic memory; System 2 performs deliberate incremental atomic changes with rollback. The paper claims state-of-the-art accuracy (76.8% average) and efficiency (13.2s execution time) over five datasets, outperforming CodeAct, ToT, ReAct, Reflexion, and others; reports a 91% speedup over the slow variant; provides an LLM-as-a-judge qualitative evaluation; and states a 'monotonic improvement' theorem. The paper also includes ablations (KC-fast, KC-slow) and dataset-difficulty analysis.

Significance. If the empirical evaluation were sound, the paper would offer a practically relevant contribution: a domain-specialized LLM agent that combines memory-augmented fast responses with deliberate incremental improvement, with unusually low latency and token costs, and ablations supporting the dual-process design. The architecture idea is well-aligned with prior dual-process agent work (SwiftSage, Talker-Reasoner) and is sensibly adapted to the ML maintenance setting. The reported knowledge-consolidation speedup (91%) and the 68% System-1 resolution rate are plausible and interesting. However, the central empirical claims are seriously compromised by the evaluation protocol: D_new is used both to select every accepted atomic change and as the test set for the headline accuracy numbers, so the reported SOTA advantage may largely reflect selection on the test sample rather than generalizable improvement. In addition, Theorem 1 is a tautological restatement of the acceptance rule, not a meaningful theoretical guarantee. The LLM-as-a-judge results (Table III) are weakened by the omission of a human-calibration check and by the fact that the 'Smartness' scores are claimed to 'confirm' quanti

major comments (5)
  1. [§IV.B and §V.A, Algorithm 1 (lines 8–14), Table II] The evaluation protocol is in-sample: in each improvement run, r_t = score(f_temp, D_new) − score(f*, D_new) is used to decide which atomic changes to accept, and the final accuracy in Table II is then reported on that same D_new. Because the agent's choices are directly optimized against the reported test sample, the headline numbers (0.768 avg, +2.4% over CodeAct) quantify selection on the drift sample, not generalization to future drift. The paper must split D_new into a development set (used for acceptance decisions and memory retrieval) and a held-out test set (used only for final metrics), and report both. As it stands, the central empirical claim is not established.
  2. [§III.E, Theorem 1] Theorem 1 states that score(f_T, D) ≥ score(f_0, D), with proof by construction: only changes with r_t > 0 are accepted. This is exactly the definition of the acceptance rule, so the theorem provides no nontrivial bound, convergence result, or generalization guarantee. It should be relabeled as a property of the algorithm, and the paper should avoid implying that it supports predictive reliability. If a substantive guarantee is intended (e.g., under Lipschitz continuity of score w.r.t. ||Δ_t||), it should be stated and proved.
  3. [§V.A, Table II] No statistical analysis supports the claimed advantages. Standard deviations across datasets straddle the differences: KC-agent (0.768±0.116) vs CodeAct (0.744±0.108) and ToT (0.741±0.103) differ by 0.024 and 0.027, well within one standard deviation; there are no per-dataset paired results, no significance tests, and no confidence intervals. The claim of 'consistently outperforms' requires per-dataset comparisons or paired tests.
  4. [§V.E and §V.D] The knowledge-consolidation evaluation is also in-sample. Episodic memory is populated on D_new during improvement runs and retrieved for the same D_new scenarios; the 91% speedup and '73% vs 45% System-1 resolution' figures therefore reflect the agent's ability to re-identify the exact scenarios it just solved, not transfer to new drift scenarios. The paper should evaluate consolidation on held-out drift scenarios (e.g., new datasets or held-out engines/time windows) to support a learning-transfer claim.
  5. [§V.B, Table III] The LLM-as-a-judge evaluation uses the same contaminated final code (selected on D_new) and lacks a human-calibration check. The favorable 'Smartness' score cannot rescue the quantitative claims, and the statement that low judge variance 'validates' the methodology is over-strong: LLM judges can agree with each other while both being biased. At minimum, a small human expert evaluation on a random subset should be included.
minor comments (7)
  1. [§IV.C] The list of baselines is inconsistent: §IV.C says 'eight established approaches' but names seven (ReAct, Reflexion, ToT, Self-Discovery, Plan-and-Execute, CodeAct, Standard), and the text refers to 'Plan-and-Solve' while Table II uses 'Plan-Execute'. Please correct the count and unify terminology.
  2. [§IV.A] NASA FD001/FD002 are typically used for remaining-useful-life regression, not classification at fixed accuracy thresholds. Please clarify the prediction target, the class construction, and how old/new distributions are split; the phrase '10,000 samples' without a time-indexed train/test description leaves the setup ambiguous.
  3. [§III.D] The proxy threshold μ is described in III.B as an accuracy improvement threshold, but in III.D as a 'cosine similarity score' threshold with the same symbol (μ=0.05). The two usages of μ conflate distinct quantities; define separate symbols.
  4. [§IV.D] The experimental parameter details are incomplete: the number of experimental runs per dataset, variance decomposition (across seeds vs datasets), max iterations T, and temperature (mentioned as 1.0 in §IV.C) are not fully specified. Reporting them is necessary for reproducibility.
  5. [Abstract/§V] The abstract states 'five datasets including... three synthetic datasets' but the synthetic dataset description (§IV.A) is terse. Provide the generation code/scripts and drift seeds as supplementary material to support the reproducibility claim.
  6. [§V.F, Figs 4–6] Figures 4-6 are not referenced in the main results narrative and are only described in a short paragraph; consider moving them to supplementary material or expanding the interpretation.
  7. [§I and §VI] Minor wording: the related work claim that 'few frameworks provide theoretical guarantees' is contradicted by the paper's own Theorem 1 being a tautology; rephrase to avoid overstating the theoretical contribution.

Circularity Check

3 steps flagged

The 76.8% SOTA claim is an in-sample optimization result: Algorithm 1 accepts only changes that raise accuracy on D_new, and Table II reports accuracy on that same D_new; Theorem 1 restates the acceptance rule.

specific steps
  1. self definitional [Section III.E, Theorem 1 and its proof]
    "Theorem 1 (Monotonic improvement property). Under the atomic change constraint ∥∆t∥ ≤ η and immediate reversion of unsuccessful modifications, KC-Agent maintains: score(fT ,D) ≥ score(f0,D)... Proof. By construction, KC-Agent only accepts changes ∆t where r t = score(f t+1,D)−score(f t,D)>0. Changes with rt ≤0 are immediately reverted, ensuring f t+1 = f t. Therefore, score(f t+1,D) ≥ score(f t,D) for all t."

    The proof says 'by construction'; the theorem's conclusion is exactly Algorithm 1's acceptance rule (lines 10-12: accept iff r_t > 0, otherwise revert). No independent bound or external fact is derived. The guarantee is a restatement of the decision rule, so it cannot serve as a separate theoretical foundation for the architecture.

  2. fitted input called prediction [Section III.C (δ and proxy) / Algorithm 1 lines 9-12; Section IV.B; Section V.A Table II]
    "Algorithm 1: 'r t ← evaluate change; if r t > 0 then f ∗ ← f temp' ... Section III.B: 'δ(update) =score(f ′,D new)−score(f,D new)' ... Section IV.B: 'We measure performance across four complementary metrics: accuracy on old data (preservation of historical performance), accuracy on new data (adaptation effectiveness on shifted distributions), execution time (computational efficiency for production deployment), and token consumption (resource utilization and computational cost).'"

    The only criterion for retaining a change is that it increases score on D_new (Algorithm 1 lines 10-12; δ defined on D_new). The reported headline accuracy (Table II, 0.768) is the same 'accuracy on new data' metric defined in Section IV.B. D_new is never split into development and held-out test sets, so the agent's greedy search is evaluated on the exact objective it optimizes. The 76.8% value and the +2.4% margin over CodeAct are therefore in-sample selection results, not predictions on unseen drift, which is the 'fitted input called prediction' pattern.

  3. fitted input called prediction [Section III.D / Section V.D (Knowledge consolidation effectiveness)]
    "Episodic memory records specific improvement scenarios, including dataset characteristics, model states, and applied solutions... when System 2 successfully improves a model, three elements are stored: the drift signature ..., the improvement strategy ..., and the performance outcome... For instance, on financial datasets, System 1 successfully resolved 73% of improvement scenarios by the third iteration, compared to 45% in the first iteration."

    The 'improvement scenarios' are the same five datasets used in the evaluation. Memory is populated from successful System 2 changes made on D_new for those datasets, and the learning-curve statistic is then measured on the same financial dataset. The speedup/accuracy gain attributed to knowledge consolidation is thus an in-sample reuse measurement: System 1 is credited for recognizing scenarios it has already encountered during the experiment. No held-out scenario set is used to test transfer, so this evidence does not establish generalization to new drift.

full rationale

The two headline results of the paper reduce to the algorithm's own acceptance loop. Theorem 1's 'monotonic improvement guarantee' is a tautology: the algorithm keeps only changes with positive r_t, so the proof is literally 'by construction.' More importantly, the central empirical claim (76.8% SOTA, outperforming CodeAct/ToT/ReAct/Reflexion) is not an independent prediction. Algorithm 1 accepts only modifications that increase score on D_new, and the final accuracy is reported on that same D_new via the 'accuracy on new data' metric; no held-out split of D_new is described anywhere. The knowledge-consolidation speedup is likewise measured on the same datasets that populate episodic memory, making it an in-sample learning curve rather than evidence of transfer. There are no self-citation chains or imported uniqueness theorems here; the circularity is in the evaluation/acceptance construction itself. Some components (architecture design, LLM-judge qualitative assessment, reproducibility logging) are independent, which is why the score is not 10, but the SOTA claim and the theoretical guarantee are forced by definition/selection, giving a score of 8.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical or mathematical entities; its free parameters and assumptions are concentrated in the evaluation protocol (data-split), the LLM-judge methodology, and the tautological improvement theorem.

free parameters (2)
  • µ (proxy decision threshold / retrieval calibration threshold) = 0.05
    Set by hand across all experiments; used in Section III.B as the minimum accepted performance delta and in Section III.D as the cosine-similarity threshold for memory retrieval. No sensitivity analysis is provided.
  • η (atomic change bound) = not specified
    Algorithm 1 imposes ||Δ_t|| ≤ η; the paper only says it is small and limits each iteration to a single targeted modification, but no numeric value or sensitivity analysis is given.
axioms (5)
  • domain assumption Using D_new both to select changes and to compute final accuracy is a valid way to measure future-drift improvement.
    Section III Algorithm 1 evaluates each change on D_new, and Section V.A reports final accuracy on the same data without a held-out split. This is the load-bearing data-split assumption.
  • domain assumption LLM judges' consensus scores are a reliable proxy for strategic quality.
    Section V.B uses three LLMs as judges; no calibration against human judgment is provided, and standard deviations up to 2.7 are described as 'strong inter-rater agreement'.
  • domain assumption Cosine similarity between dataset meta-features is a sufficient measure for memory transfer.
    Section III.D: System 1 applies stored strategies when cosine similarity exceeds a threshold; no analysis shows that this similarity correlates with actual transfer success.
  • ad hoc to paper Unsuccessful changes can be independently evaluated without interaction effects.
    Theorem 1 relies on r_t reflecting only change Δ_t; in a sequence of atomic changes, earlier accepted changes alter the model and can confound the evaluation of later changes.
  • domain assumption A RandomForest baseline on tabular data is sufficient to validate the cognitive architecture's claims.
    Section VII acknowledges the baseline is simple, but the central empirical claims are only tested on this controlled setup; generalization to deep learning pipelines is assumed, not shown.

pith-pipeline@v1.3.0-daily-deepseek · 12667 in / 12426 out tokens · 106843 ms · 2026-08-04T08:49:06.984609+00:00 · methodology

0 comments
read the original abstract

Data drift poses significant challenges for machine learning systems in production, requiring continuous model updates to maintain performance. We present KC-Agent, a dual-process cognitive architecture for automated ML model improvement that combines fast pattern recognition (System 1) with deliberate incremental updates (System 2). Our approach implements structured memory systems enabling System 1 to leverage successful solutions previously discovered by System 2, achieving efficient pattern-based responses without costly re-computation. KC-Agent incorporates atomic change principles and rollback capabilities to ensure reliable, verifiable updates in production environments. We evaluate our method on five datasets including real-world NASA turbofan data with authentic temporal degradation and synthetic datasets with controlled drift scenarios. KC-Agent achieves state-of-the-art performance (76.8% accuracy) while maintaining optimal efficiency (13.2s execution time), outperforming established cognitive architectures: CodeAct (+2.4%), Tree of Thoughts (+3.6%), ReAct (+8.0%), and Reflexion (+8.9%). Consensus evaluation by a panel of state-of-the-art LLMs confirms superior strategic efficacy (8.33/10 Smartness score), significantly outperforming baseline agents. The knowledge consolidation mechanism delivers 91% speedup over the slow variant while maintaining higher accuracy. Our approach demonstrates both theoretical foundations and practical viability for cognitive-inspired automated ML improvement systems capable of handling complex real-world data drift scenarios.

Figures

Figures reproduced from arXiv: 2608.02351 by Ajay Dholakia, David Ellison, Gusseppe Bravo-Rocca, Jordi Guitart, Puneet Jain.

Figure 1
Figure 1. Figure 1: KC-Agent dual-process architecture showing the numbered workflow [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Memory architecture with three components: Semantic memory (SM) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Example improvement flow showing KC-Agent processing a drift [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Agent adaptability analysis showing performance consistency vs. mean [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Dataset difficulty ranking showing average agent performance, ease [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

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

26 extracted references · 7 linked inside Pith

  1. [1]

    Monitoring machine learning models: Online detection of relevant deviations,

    F. Heinrichs, “Monitoring machine learning models: Online detection of relevant deviations,” 2023. [Online]. Available: https://arxiv.org/abs/2309.15187

    Pith/arXiv arXiv 2023

  2. [2]

    A monitor- ing framework for deployed machine learning models with supply chain examples,

    B. Eck, D. Kabakci-Zorlu, Y . Chen, F. Savard, and X. Bao, “A monitor- ing framework for deployed machine learning models with supply chain examples,” inProceedings of the 2022 IEEE International Conference on Big Data, Osaka, Japan, 2022, pp. 2231–2238

    2022

  3. [3]

    Detecting and correcting for label shift with black box predictors,

    Z. C. Lipton, Y . Wang, and A. J. Smola, “Detecting and correcting for label shift with black box predictors,” inProceedings of the 35th International Conference on Machine Learning, ICML 2018, Stock- holmsm¨assan, Stockholm, Sweden, July 10-15, 2018, ser. Proceedings of Machine Learning Research, vol. 80. PMLR, 2018, pp. 3128–3136

    2018

  4. [4]

    Evaluating large language models trained on code,

    M. Chen, J. Tworek, H. Jun, Q. Yuan, H. P. de Oliveira Pinto, J. Kaplan, H. Edwards, Y . Burda, N. Joseph, G. Brockman, A. Ray, R. Puri, G. Krueger, M. Petrov, H. Khlaaf, G. Sastry, P. Mishkin, B. Chan, S. Gray, N. Ryder, M. Pavlov, A. Power, L. Kaiser, M. Bavarian, C. Winter, P. Tillet, F. P. Such, D. Cummings, M. Plappert, F. Chantzis, E. Barnes, A. Her...

    Pith/arXiv arXiv 2021

  5. [5]

    AgentBench: Evaluating LLMs as Agents,

    X. Liu, H. Yu, H. Zhang, Y . Xu, X. Lei, H. Lai, Y . Gu, H. Ding, K. Men, K. Yang, S. Zhang, X. Deng, A. Zeng, Z. Du, C. Zhang, S. Shen, T. Zhang, Y . Su, H. Sun, M. Huang, Y . Dong, and J. Tang, “AgentBench: Evaluating LLMs as Agents,” 2023. [Online]. Available: https://arxiv.org/abs/2308.03688

    Pith/arXiv arXiv 2023

  6. [6]

    MIND2WEB: towards a generalist agent for the web,

    X. Deng, Y . Gu, B. Zheng, S. Chen, S. Stevens, B. Wang, H. Sun, and Y . Su, “MIND2WEB: towards a generalist agent for the web,” in Proceedings of the 37th International Conference on Neural Information Processing Systems, ser. NIPS’23. Curran Associates Inc., 2023

    2023

  7. [7]

    Reflexion: language agents with verbal reinforcement learning,

    N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, and S. Yao, “Reflexion: language agents with verbal reinforcement learning,” in Proceedings of the 37th International Conference on Neural Information Processing Systems, ser. NIPS’23. Curran Associates Inc., 2023

    2023

  8. [8]

    Large language models empowered agent-based modeling and simulation: A survey and perspectives,

    C. Gao, X. Lan, N. Li, Y . Yuan, J. Ding, Z. Zhou, F. Xu, and Y . Li, “Large language models empowered agent-based modeling and simulation: A survey and perspectives,”Humanities and Social Sciences Communications, vol. 11, 2024

    2024

  9. [9]

    Generative Agents: Interactive Simulacra of Human Behavior,

    J. S. Park, J. O’Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bern- stein, “Generative Agents: Interactive Simulacra of Human Behavior,” inProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, ser. UIST’23. Association for Computing Machinery, 2023

    2023

  10. [10]

    Kahneman,Thinking, fast & slow

    D. Kahneman,Thinking, fast & slow. Farrar, Straus & Giroux, 2011

    2011

  11. [11]

    Clear,Atomic Habits: An Easy & Proven Way to Build Good Habits & Break Bad Ones

    J. Clear,Atomic Habits: An Easy & Proven Way to Build Good Habits & Break Bad Ones. Penguin Publishing Group, 2018. [Online]. Available: https://jamesclear.com/atomic-habits

    2018

  12. [12]

    ReAct: Synergizing Reasoning and Acting in Language Models,

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. R. Narasimhan, and Y . Cao, “ReAct: Synergizing Reasoning and Acting in Language Models,” in Proceedings of the Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023

    2023

  13. [13]

    Tree of thoughts: deliberate problem solving with large language models,

    S. Yao, D. Yu, J. Zhao, I. Shafran, T. L. Griffiths, Y . Cao, and K. Narasimhan, “Tree of thoughts: deliberate problem solving with large language models,” inProceedings of the 37th International Conference on Neural Information Processing Systems, ser. NIPS’23. Curran Associates Inc., 2023

    2023

  14. [14]

    Self-discover: large language models self-compose reasoning structures,

    P. Zhou, J. Pujara, X. Ren, X. Chen, H.-T. Cheng, Q. V . Le, E. H. Chi, D. Zhou, S. Mishra, and H. S. Zheng, “Self-discover: large language models self-compose reasoning structures,” inProceedings of the 38th International Conference on Neural Information Processing Systems, ser. NIPS’24. Curran Associates Inc., 2024

    2024

  15. [15]

    Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models,

    L. Wang, W. Xu, Y . Lan, Z. Hu, Y . Lan, R. K.-W. Lee, and E.-P. Lim, “Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models,” inProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Toronto, Canada: Association for Computational Linguistics, Jul. 2023, ...

    2023

  16. [16]

    Executable code actions elicit better llm agents,

    X. Wang, Y . Chen, L. Yuan, Y . Zhang, Y . Li, H. Peng, and H. Ji, “Executable code actions elicit better llm agents,” inProceedings of the 41st International Conference on Machine Learning, ser. ICML’24. JMLR.org, 2024

    2024

  17. [17]

    SWIFTSAGE: a generative agent with fast and slow thinking for complex interactive tasks,

    B. Y . Lin, Y . Fu, K. Yang, F. Brahman, S. Huang, C. Bhagavatula, P. Ammanabrolu, Y . Choi, and X. Ren, “SWIFTSAGE: a generative agent with fast and slow thinking for complex interactive tasks,” in Proceedings of the 37th International Conference on Neural Information Processing Systems, ser. NIPS’23. Curran Associates Inc., 2023

    2023

  18. [18]

    Agents thinking fast and slow: A talker-reasoner architecture,

    K. Christakopoulou, S. Mourad, and M. Matari ´c, “Agents thinking fast and slow: A talker-reasoner architecture,” 2024. [Online]. Available: https://arxiv.org/abs/2410.08328

    Pith/arXiv arXiv 2024

  19. [19]

    Overcoming catastrophic forgetting in neural networks,

    J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, D. Hassabis, C. Clopath, D. Kumaran, and R. Hadsell, “Overcoming catastrophic forgetting in neural networks,”Proceedings of the National Academy of Sciences, vol. 114, no. 13, p. 3521–3526, Mar. 2017

    2017

  20. [20]

    Learning to learn without forgetting by maximizing transfer and minimizing interference,

    M. Riemer, I. Cases, R. Ajemian, M. Liu, I. Rish, Y . Tu, and G. Tesauro, “Learning to learn without forgetting by maximizing transfer and minimizing interference,” 2019. [Online]. Available: https://arxiv.org/abs/1810.11910

    Pith/arXiv arXiv 2019

  21. [21]

    Dark experience for general continual learning: a strong, simple baseline,

    P. Buzzega, M. Boschini, A. Porrello, D. Abati, and S. Calderara, “Dark experience for general continual learning: a strong, simple baseline,” in Proceedings of the 34th International Conference on Neural Information Processing Systems, ser. NIPS’20. Curran Associates Inc., 2020

    2020

  22. [22]

    Don’t forget, there is more than forgetting: new metrics for continual learning,

    N. D ´ıaz-Rodr´ıguez, V . Lomonaco, D. Filliat, and D. Maltoni, “Don’t forget, there is more than forgetting: new metrics for continual learning,” 2018. [Online]. Available: https://arxiv.org/abs/1810.13166

    Pith/arXiv arXiv 2018

  23. [23]

    Benchmark and survey of automated ma- chine learning frameworks,

    M.-A. Z ¨oller and M. F. Huber, “Benchmark and survey of automated ma- chine learning frameworks,”Journal of Artificial Intelligence Research (JAIR), vol. 70, pp. 409–472, 2021

    2021

  24. [24]

    AI Maintenance: A Robustness Perspective,

    P.-Y . Chen and P. Das, “AI Maintenance: A Robustness Perspective,” Computer, vol. 56, no. 2, pp. 48–56, 2023

    2023

  25. [25]

    Applying machine learning in self- adaptive systems: A systematic literature review,

    O. Gheibi, D. Weyns, and F. Quin, “Applying machine learning in self- adaptive systems: A systematic literature review,”ACM Trans. Auton. Adapt. Syst., vol. 15, no. 3, Aug. 2021

    2021

  26. [26]

    Towards a science of scaling agent systems,

    Y . Kim, K. Gu, C. Park, C. Park, S. Schmidgall, A. A. Heydari, Y . Yan, Z. Zhang, Y . Zhuang, M. Malhotra, P. P. Liang, H. W. Park, Y . Yang, X. Xu, Y . Du, S. Patel, T. Althoff, D. McDuff, and X. Liu, “Towards a science of scaling agent systems,” 2025. [Online]. Available: https://arxiv.org/abs/2512.08296

    Pith/arXiv arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.