Pith.

Artificial Intelligence · arXiv:2610.16078 v1 · submitted 12 Oct 2026 · 37 references

Heterogeneous Agentic Coevolution

Wenqian Ye, Ziwei Guan, Zi Yang, Shivani Modi, Cici Cong Wu, Yunlong He, Yuxin Chen, Zhengyu Zhang

show 9 more authors

Liang Luo, Venkatesh Ranganathan, Jianhui Sun, Yantao Yao, Wenlin Chen, Buyun Zhang, Ellie Dingqiao Wen, Henry Kautz, Aidong Zhang

The paper says

“Recent advances in self-improving LLM agents have demonstrated the potential of AI systems that drive their own intelligence. However, many domain-specific applications rely on specialized foundation models that possess strong domain expertise but remain dependent on human-designed improvements.”

more of the abstract

Extending recursive self-improvement to these models presents a key challenge: as a specialist improves, the strategies used to advance it must also adapt. To address this challenge, we propose Heterogeneous Agentic CoEvolution (HACE), a framework that jointly evolves specialized foundation models and LLM agents through reciprocal learning. HACE improves the specialist through supervised learning while training the agent with on-policy reinforced self-training, guided by verified performance gains on the current specialist. The agent's exploration, in turn, informs subsequent specialist training, enabling both models to adapt to each other's progress. To sustain this reciprocal learning, HACE further employs evolutionary search to optimize their interaction harness, adapting how the models exchange information and select training experiences as their capabilities evolve. We evaluate HACE across three challenging domains: recommendation systems, protein function prediction in biology, and crystal stability prediction in material discovery. We show that HACE outperforms the evaluated self-improving methods on the primary task metrics across all three domains. These results shed light on heterogeneous coevolution as a universal approach to improving specialized foundation models through reciprocal interaction with a learning agent.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 14 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A solid empirical paper on reciprocal coevolution of LLM agents and specialized models, with a weak theory and a minor blind spot about validation noise. This line is wrong

1 major objection · 4 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 37 references say.

The paper says, at Section 3.2, Eq. (4); Appendix B.3

The review did not quote this passage.

Pith says: Objection 1 of 1 · major

The acceptance criterion ΔJ_t > 0 is evaluated on a single validation batch of at most 256 instances. The paper provides no confidence bound, error analysis, or adaptive threshold to control false acceptances arising from sampling noise. Validation data are reused across rounds (Appendix B.3), compounding the risk of overfitting to noise over many rounds.

more of objection 1

Because the reciprocal loop depends on the reliability of this gate, uncharacterized noise in the acceptance decision is a load-bearing concern. The authors should quantify the false-positive rate (e.g., by reporting sensitivity to different validation-split seeds or by adopting a statistically grounded selection rule such as a one-sided test with a noise calibration step).

4 minor objections

The paper says, at Abstract

Pith says: minor objection 1 of 4

The sentence 'HACE further employs evolutionary search to optimize their interaction harness' would benefit from a brief indication of what the harness controls (information exchange and training experience selection), as is done in Section 3.1.

The paper says, at Figure 2

Pith says: minor objection 2 of 4

The figure is dense and the flow is not immediately clear from the caption alone. Consider adding a short step-by-step annotation in the caption or a numeric callout in the figure panels to guide the reader.

show 2 more

The paper says, at Appendix B.3, Eq. (4) text

Pith says: minor objection 3 of 4

The definition of acceptance as 'strictly positive improvement' does not account for the discretization of the validation metric. If the true ΔJ is zero but the estimated ΔJ is positive due to sampling noise, a non-improving pair could be accepted. Minor rewording to clarify that the condition is evaluated on the empirical estimate would help.

The paper says, at Theorem 1

Pith says: minor objection 4 of 4

The bound depends on the size of the specialist class |F|, which is fixed before exploration. The paper does not discuss how to choose or bound |F| in practice, nor whether the bound degrades if F is large. A brief remark on this limitation would be helpful.

What would settle it

On the recommendation task (Toys/SASRec), run HACE with 10 different random seeds for the validation batch selection (holding training and test fixed). For each round t, record the estimated ΔJ and compute its standard deviation over seeds. If the standard deviation is comparable to or larger than the mean ΔJ of accepted proposals (e.g., > 0.001), then false acceptances are likely. Additionally, repeat the full HACE run with a second independent validation set (same size, different random split) and compare the final test HR@10; a significant drop (> 1 s.d.) would indicate overfitting to the original validation set through the acceptance rule.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 14 Oct 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic partly agrees 45 s · 26,898 tokens read, 4,558 written
  3. Editor wrote the verdict above 14 s · 27,832 tokens read, 1,248 written
  4. Referee major revision: 1 major objection, 4 minor 21 s · 27,250 tokens read, 2,063 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

You should know two things about this paper. First, the empirical story is genuinely strong: HACE pushes already-converged specialist models further in recommendation, protein function prediction, and crystal stability, across multiple architectures, with consistent gains over all compared self-improving methods. The ablation studies show each component pulls weight. Second, the theory (Theorem 1) is largely decorative — it assumes a fixed finite specialist class, fixed source probabilities, and an analysis loss independent of the harness, none of which match practice. The practical guarantee comes from the empirical evaluation, not the math.

What's new: The key idea — coupling supervised specialist updates with on-policy agent reinforcement via an evolving harness that controls information flow and instance selection — is original and well executed. Prior work updated either the model or the harness, not both reciprocally. The experiments cover three domains with enough rigor: standard deviations, multiple seeds, scaling experiments, and case studies.

Soft spots: The main one is the validation noise issue. The acceptance decision in Eq. 4 uses a single batch of size min(256,|Dval|). The paper provides no analysis of false-positive acceptance rates, the validation set is reused across rounds, and the theory implicitly assumes accepted proposals are true improvements. The stress-test concern is real — though the empirical trajectories show net gains, the sensitivity to batch composition remains untested. Second, Theorem 1's conditions are too idealized to bind in practice; the paper acknowledges this indirectly (fixed class, fixed source), but doesn't discuss how far the real setting departs from them. Third, no code or data released, which limits reproducibility. The authors list appropriate limitations (single specialist/agent, no multi-agent study), so they are not overselling.

Who is this for? Researchers working on self-improving AI systems, especially those combining LLM agents with domain-specific models. The empirical recipe is valuable even if the theory is thin.

Recommendation: Deserves a serious referee. I'd recommend conditional acceptance: ask the authors to (a) discuss the validation noise risk, even if only to bound it empirically, and (b) clarify the gap between Theorem 1's assumptions and the actual experimental setup. Code release would be a bonus.

The strongest objection

The acceptance criterion in Eq. 4 (ΔJ > 0) is evaluated on a single validation batch of size |Bt| = min(256, |Dval|). This deterministic threshold ignores sampling noise. If the true ΔJ is negative but the estimated ΔJ is positive (a false positive due to a small or unrepresentative batch), a harmful (S', H') pair is retained, potentially degrading specialist performance and corrupting the training data for both models in subsequent rounds. The paper provides no confidence bound, error analysis, or adaptive threshold to control false acceptances across rounds. The validation set is reused across rounds (Appendix B.3), compounding the risk: repeated selection on the same validation instances can lead to overfitting to noise. The theoretical guarantee (Theorem 1) conditions on the accepted training weights but does not model selection errors; it implicitly assumes that accepted proposals are true improvements. Because the entire reciprocal loop depends on the reliability of this gate, uncharacterized noise in the acceptance decision is a load-bearing concern for the claimed robustness of HACE. The empirical trajectories (Figure 3) show net gains, but the sensitivity to batch composition and noise level remains untested. Without quantifying the false-positive rate, it is unclear whether the method would still improve under a different validation-split seed, a smaller validation set, or a task with higher evaluation variance.

The reader's weakest assumption focuses on the existence of positive responses from the agent. While important, that concern is partially addressed by the empirical positive results across three domains, and the paper acknowledges the possibility of collapse (Appendix B.1). My concern addresses a different failure mode: even if verified responses exist, the joint selection gate can be noisy and accept harmful updates, degrading the specialist and breaking the loop. The paper does not discuss this noise or provide safeguards, making it a load-bearing yet unaddressed risk. The empirical results are strong, so I recommend keeping the CONDITIONAL verdict (unchanged), as the concern does not invalidate the paper but highlights a need for additional analysis or a robustness check. The agreement with the reader is partial because their focus was on the agent's output, while mine is on the selection mechanism's reliability.

The referee's report

If the results hold, HACE provides a practical mechanism for extending autonomous self-improvement to specialized foundation models that have plateaued under supervised training. The empirical gains are substantial: +12.1% average HR@10 on recommendation, +0.046–0.050 NDCG on protein function prediction, and consistent improvements on material discovery. The framework is architecture-agnostic and the ablations show each component contributes. The theoretical analysis offers a formal lens on learning from adaptively selected experience, though its assumptions are idealized.

Pith's reading of the paper (a machine's summary, not the authors')

HACE is a framework for jointly evolving a specialized foundation model (the specialist) and a heterogeneous LLM agent through a closed-loop reciprocal learning process. The agent explores candidate predictions and selects training instances for the specialist, whose subsequent supervised update in turn changes which agent contributions count as improvements. An evolutionary harness adapts how the two models interact as their capabilities shift. Across recommendation, protein function prediction, and crystal stability prediction, HACE starts from converged specialists and significantly outperforms all evaluated self-improving baselines, achieving average HR@10 gains of 12.1% and NDCG boosts of up to 0.05 in biology. The paper argues that this kind of heterogeneous coevolution can serve as a universal approach to autonomously improving domain-specific AI models.

What carries the argument. The central mechanism is the reciprocal learning loop plus an evolutionary interaction harness. In each round, the harness controls what the agent sees and which training instances the specialist receives. Agent responses are verified against the current specialist’s predictions; only responses with positive contribution (ri(τ) > 0) enter agent training, and the harness’s curriculum selects which instances train the specialist via supervised learning. Joint validation retains both the specialist and harness only when the pair improves a held-out task score with the current agent fixed. The harness itself is mutated from an archive of accepted harnesses, so it adapts to the evolving capabilities of both learners.

Core claim. The paper claims that specialized foundation models that have already reached a supervised-convergence plateau can be substantially improved through reciprocal interaction with a learning LLM agent. The central discovery is that an agent whose overall domain accuracy is lower than the specialist can still provide per-instance contributions that, when used to guide the selection of supervised training experience, enable the specialist to reach new performance levels. The improvement is sustained across rounds because the harness governing information exchange and experience selection also evolves, preventing the kind of stale feedback that would occur if either learner were fixed.

The premise it rests on. The framework assumes that for a non-negligible fraction of training instances, the LLM agent can produce verified responses (responses with positive contribution ri(τ) > 0) under the current specialist and harness, and that these responses provide useful signal for specialist retraining; if no positive responses ever arise, the reciprocal loop collapses and no improvement occurs.

If the paper is right (Pith's inferences, not the authors')

Specialized models that have stopped improving under standard supervised schedules can be further advanced without human intervention, simply by pairing them with an LLM agent and starting the coevolution loop.

Because the approach is agnostic to specialist architecture (transformers, graph networks, factorization machines) and works with frozen backbones, it can be retrofitted to existing scientific and industrial models.

The learned agent and harness together encode a reusable improvement strategy: the trained agent can suggest new training curricula or view configurations for similar tasks, reducing future human effort.

The theoretical bound (Theorem 1) implies that as long as the agent can generate a non-zero fraction of verified responses over the source distribution, the specialist risk can be driven to zero, providing a formal guarantee for the framework’s effectiveness under mild conditions.

HACE opens the door to a new regime of autonomous model development where multiple domain specialists could be coevolved with shared or specialized agents, creating a self-improving ecosystem for scientific AI.

The evolutionary component of the harness suggests that the optimal interface between two models is itself a learned object; this principle could extend to human-AI collaboration, where the protocol for exchanging information is continuously optimized.

A direct testable extension: apply HACE to a domain where the agent is deliberately initialized to produce zero verified responses on every instance. The framework should collapse to the baseline (no specialist improvement), confirming the load-bearing role of the verification condition.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Recent advances in self-improving LLM agents have demonstrated the potential of AI systems that drive their own intelligence. However, many domain-specific applications rely on specialized foundation models that possess strong domain expertise but remain dependent on human-designed improvements. Extending recursive self-improvement to these models presents a key challenge: as a specialist improves, the strategies used to advance it must also adapt. To address this challenge, we propose Heterogeneous Agentic CoEvolution (HACE), a framework that jointly evolves specialized foundation models and LLM agents through reciprocal learning. HACE improves the specialist through supervised learning while training the agent with on-policy reinforced self-training, guided by verified performance gains on the current specialist. The agent's exploration, in turn, informs subsequent specialist training, enabling both models to adapt to each other's progress. To sustain this reciprocal learning, HACE further employs evolutionary search to optimize their interaction harness, adapting how the models exchange information and select training experiences as their capabilities evolve. We evaluate HACE across three challenging domains: recommendation systems, protein function prediction in biology, and crystal stability prediction in material discovery. We show that HACE outperforms the evaluated self-improving methods on the primary task metrics across all three domains. These results shed light on heterogeneous coevolution as a universal approach to improving specialized foundation models through reciprocal interaction with a learning agent.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-14

The authors' main claim, as the classifier picked it out 0.86

“These results shed light on heterogeneous coevolution as a universal approach to improving specialized foundation models through reciprocal interaction with a learning agent.”

How the main claim is stated: asserted 0.50

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.10 · human participants 0.05 · controlled comparison 0.63 · randomized assignment 0.02 · states quantitative result 0.10 · compares with prior results 0.94 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Recent advances in self-improving LLM agents have demonstrated the potential of AI systems that drive their own intelligence. background 1.00 0.04 0.02 asserted 0.98
However, many domain-specific applications rely on specialized foundation models that possess strong domain expertise but remain dependent on human-designed improvements. background 0.66 0.05 0.02 asserted 0.99
Extending recursive self-improvement to these models presents a key challenge: as a specialist improves, the strategies used to advance it must also adapt. gap 0.65 0.06 0.02 asserted 0.98
To address this challenge, we propose Heterogeneous Agentic CoEvolution (HACE), a framework that jointly evolves specialized foundation models and LLM agents through reciprocal learning. contribution 0.95 0.74 0.07 asserted 0.99
HACE improves the specialist through supervised learning while training the agent with on-policy reinforced self-training, guided by verified performance gains on the current specialist. method 0.99 0.79 0.05 asserted 1.00
The agent's exploration, in turn, informs subsequent specialist training, enabling both models to adapt to each other's progress. method 0.98 0.56 0.06 asserted 0.98
To sustain this reciprocal learning, HACE further employs evolutionary search to optimize their interaction harness, adapting how the models exchange information and select training experiences as their capabilities evolve. method 0.98 0.83 0.05 asserted 1.00
We evaluate HACE across three challenging domains: recommendation systems, protein function prediction in biology, and crystal stability prediction in material discovery. method 0.63 0.95 0.03 asserted 0.92
We show that HACE outperforms the evaluated self-improving methods on the primary task metrics across all three domains. result 1.00 0.96 0.56 asserted 1.00
These results shed light on heterogeneous coevolution as a universal approach to improving specialized foundation models through reciprocal interaction with a learning agent. interpretation 1.00 0.64 0.82 asserted 0.50

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.81

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-14: 1 request, 7,501 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (4)
Figure 1 of the paper
Figure 1. From human-directed optimization to heterogeneous agentic coevolution. HACE brings together specialized foundation models and LLM agents into a shared learning process, where each helps determine how the other improves. Across rounds, both learners and their interaction harness evolve, creating new opportunities for reciprocal improvement. †These authors jointly supervised this work 1 arXiv:2610.16078v1 [cs.AI] 12 Oct 2026
Figure 2 of the paper
Figure 2. Overview of HACE. Agent exploration selects specialist training experience. Joint validation retains the specialist and harness, after which fresh verified responses train the agent and shape later exploration. The on-policy agent optimization step uses reinforced self-training. 3.1 Reciprocal Learning Let S denote the specialist, A the agent, and H the harness. Round t = 0, 1, . . . starts from (St, Ht, At). We propose a harness H′ by modifying a parent from an archive of accepted harnesses. Its fields control the task information and specialist context shown to the agent, how…
Figure 3 of the paper
Figure 3. Evolution dynamics of three domains. Best recorded scores during evolution, measured by HR@10, Recall@10, and mean NDCG@3 and NDCG@5. Dashed lines mark the frozen specialist. Improvement over rounds
Figure 4 of the paper
Figure 4. Illustrations on events during coevolution. Upper panels compare specialist and combined predictions. Lower panels show recorded selection scores and their running maxima from one run in each domain, with selected updates annotated. 9

Figures are the authors', reproduced from the paper.

References (37)
  1. [1] Zhang, Jenny and Hu, Shengran and Lu, Cong and Lange, Robert and Clune, Jeff , booktitle = on Pith
  2. [2] Meta-harness: End-to-end optimization of model harnesses , author=. arXiv preprint arXiv:2603.28052 , year= on PithPith's reviewarXiv
  3. [3] Self-Adapting Language Models , author =. Advances in Neural Information Processing Systems , volume =. 2025 , doi = on Pith
  4. [4] Learning to Discover at Test Time , author =. International Conference on Machine Learning , year = on Pith
  5. [5] Evolutionary-scale prediction of atomic-level protein structure with a language model , author =. Science , volume =. 2023 , doi = on Pith
  6. [6] Deng, Bowen and Zhong, Peichen and Jun, KyuJung and Riebesell, Janosh and Han, Kevin and Bartel, Christopher J. and Ceder, Gerbrand , journal =. 2023 , doi = on Pith
  7. [7] Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations , author =. Proceedings of the 41st International Conference on Machine Learning , series = on Pith
  8. [8] Ma, Yecheng Jason and Liang, William and Wang, Guanzhi and Huang, De-An and Bastani, Osbert and Jayaraman, Dinesh and Zhu, Yuke and Fan, Linxi and Anandkumar, Anima , booktitle = on Pith
  9. [9] Hyperagents , author =. 2026 , eprint = on Pith
  10. [10] Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models , author =. Transactions on Machine Learning Research , year = on Pith
  11. [11] J. Cumulated gain-based evaluation of. ACM Transactions on Information Systems , volume =. 2002 , doi = on Pith
  12. [12] Counterfactual Multi-Agent Policy Gradients , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2018 , doi = on Pith
  13. [13] Automated Design of Agentic Systems , author =. International Conference on Learning Representations , year = on Pith
  14. [14] AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design , author=. arXiv preprint arXiv:2608.13560 , year= on PithPith's reviewarXiv
  15. [15] Liu, Ziyang and Guo, Xinyan and Wei, Xuchen and Hao, Han and Yang, Liu , howpublished =. 2026 , eprint = on Pith
  16. [16] Wang, Rui and Lehman, Joel and Clune, Jeff and Stanley, Kenneth O. , booktitle =. 2019 , doi = on Pith
  17. [17] Emergent Complexity and Zero-shot Transfer via Unsupervised Environment Design , author =. Advances in Neural Information Processing Systems , volume = on Pith
  18. [18] Self-Attentive Sequential Recommendation , author =. 2018 IEEE International Conference on Data Mining (ICDM) , pages =. 2018 , doi = on Pith
  19. [19] Zhang, Buyun and Luo, Liang and Chen, Yuxin and Nie, Jade and Liu, Xi and Li, Shen and Zhao, Yanli and Hao, Yuchen and Yao, Yantao and Wen, Ellie Dingqiao and Park, Jongsoo and Naumov, Maxim and Chen, Wenlin , booktitle = on Pith
  20. [20] An expanded evaluation of protein function prediction methods shows an improvement in accuracy , author =. Genome Biology , volume =. 2016 , doi = on Pith
show 17 more
  1. [21] A framework to evaluate machine learning crystal stability predictions , author =. Nature Machine Intelligence , volume =. 2025 , doi = on Pith
  2. [22] Batatia, Ilyes and Kov. Advances in Neural Information Processing Systems , volume =. 2022 , doi = on Pith
  3. [23] Elnaggar, Ahmed and Heinzinger, Michael and Dallago, Christian and Rehawi, Ghalia and Wang, Yu and Jones, Llion and Gibbs, Tom and Feher, Tamas and Angerer, Christoph and Steinegger, Martin and Bhowmik, Debsindhu and Rost, Burkhard , journal =. 2022 , doi = on Pith
  4. [24] OpenEvolve: an open-source evolutionary coding agent , author =. 2025 , publisher = on Pith
  5. [25] Janzen, Daniel H. , title =. Evolution , volume =. 1980 , doi = on Pith
  6. [26] Adaevolve: Adaptive llm driven zeroth-order optimization , author=. arXiv preprint arXiv:2602.20133 , year= on PitharXiv
  7. [27] EvoX: Meta-Evolution for Automated Discovery , author=. Third Conference on Language Modeling , year= on Pith
  8. [28] Harper, F. Maxwell and Konstan, Joseph A. , journal =. The. 2015 , doi = on Pith
  9. [29] Image-Based Recommendations on Styles and Substitutes , author =. Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages =. 2015 , doi = on Pith
  10. [30] Predicting stable crystalline compounds using chemical similarity , author =. npj Computational Materials , volume =. 2021 , doi = on Pith
  11. [31] Structured Agent Distillation for Large Language Model Agents , author =. 2026 , eprint = on Pith
  12. [32] Kang, Minki and Jeong, Jongwon and Lee, Seanie and Cho, Jaewoong and Hwang, Sung Ju , year =. Distilling. 2505.17612 , archivePrefix = on PitharXiv
  13. [33] Wang, Jianze and Liu, Ying and Chen, Jinlong and Hu, Xuchun and Zhang, Qilong and Cao, Yu and Wang, Jun and Yang, Hua and Xie, Yong and Chen, Qianglong , year =. 2605.01347 , archivePrefix = on PithPith's reviewarXiv
  14. [34] Yang, Shidong and Ma, Ziyu and Huang, Tongwen and Hu, Yiming and Wang, Yong and Chu, Xiangxiang. C o E volve: Training LLM Agents via Agent-Data Mutual Evolution. Proceedings of the 64th Annual Meeting of the A ssociation for C omputational L inguistics (Volume 1: Long Papers)... on Pithdoi:10.18653/v1/2026.acl-long.1055
  15. [35] Co-Harness: Co-Evolving Harnesses and Model for Agents , author=. COLM 2026 The 2nd Workshop on Lifelong Agents: Learning, Aligning, and Evolving , year= on Pith
  16. [36] Time-uniform, nonparametric, nonasymptotic confidence sequences , author=. The Annals of Statistics , volume=. 2021 , doi= on Pith
  17. [37] Heterogeneous Scientific Foundation Model Collaboration , author=. arXiv preprint arXiv:2604.27351 , year= on PithPith's reviewarXiv
Cited by 1 Pith paper

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment