OpenLoopEvolve: A Verifiable Self-Evolution Framework for Loop Policies in Long-Horizon Complex Tasks

Li Li; Siqi Wang; Xinlin Li; Zhenglin Li

REVIEW 5 major objections 7 minor 20 references

OpenLoopEvolve: A Verifiable Self-Evolution Framework for Loop Policies in Long-Horizon Complex Tasks

T0 review · 5 major / 7 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read OpenLoopEvolve claims that externalizing an agent's full execution loop as a versioned, gated Loop Policy raises long-horizon performance, increasing mean final funds on YC-Bench by 140% and 166% over a fixed policy.

desk verdict A well-specified framework for externalized loop policies with a real contribution in asset/release machinery, but the empirical evidence is thin and the offline mode needs a documented data split before the headline gains can be believed. read the letter →

arxiv 2608.09380 v1 pith:CSSABSVV submitted 2026-08-10 cs.AI

Siqi Wang , Xinlin Li , Zhenglin Li , Li Li This is my paper · ORCID

classification cs.AI

keywords LoopPolicyself-evolutionlong-horizontaskslargelanguagemodelagentsChampion-ChallengerassetversioningreleasegateYC-Bench

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper's central claim is that the complete control logic of a long-horizon agent—how it observes, plans, uses memory, acts, verifies, recovers, stops, and budgets—can be lifted out of the host program and treated as a single versioned asset, called a Loop Policy, that accumulates experience across runs. If that claim holds, the same machinery that governs software releases—candidate generation, paired evaluation against a current best version, a release gate, and rollback—can govern agent behavior, so recurring failures in a year-long task become reusable evidence rather than lost traces. On YC-Bench, a simulated year of business management, both an online mode that evolves from live feedback and an offline mode that searches archived traces improve on keeping the initial policy fixed: mean final funds rise from about 366,000 to 879,000 and 975,000 simulated units, task success from 73.9% to 87.9% and 91.8%, and annual survival from one of three seeds to two and three of three. The paper frames this as evidence that loop-control experience is portable, comparable, and reusable.

What carries the argument

The load-bearing object is the Bundle, a versioned asset pairing a Loop Policy with applicability conditions and parent evidence. The Loop Policy's eight components define the state-update and action-decision functions $f_\pi$ and $h_\pi$ in the loop dynamics $x_{t+1}=f_\pi(x_t,o_t,a_t,e_t)$. The mechanism that makes evolution verifiable is paired Champion-Challenger evaluation with the relative change score of Eq. (9), followed by the robust release gate of Eq. (10), whose four constraints require demonstrated benefit, valid evidence, acceptable tail risk, and resource compliance before any candidate can touch the execution path.

What would settle it

Rerun YC-Bench with the same seeds while matching the number of main-task calls across fixed and evolved policies, or compute paired confidence intervals across many seeded environments; if the fund differences shrink to within seed noise, the central claim fails.

Watch

Extended reading notes

Core claim

The discovery is that an agent's execution loop can be made into an external, governable asset: OLE formalizes the Loop Policy $\pi = \langle \pi_{\mathrm{obs}}, \pi_{\mathrm{plan}}, \pi_{\mathrm{mem}}, \pi_{\mathrm{act}}, \pi_{\mathrm{ver}}, \pi_{\mathrm{rec}}, \pi_{\mathrm{stop}}, \pi_{\mathrm{bud}}\rangle$ and packages it with applicability conditions and provenance into a Bundle $B_v = \langle \pi_v, \chi_v, \ell_v\rangle$ with stable version indices and a lineage. Traces from runs are distilled into evolution evidence $\mathcal{E} = \Phi(\mathcal{D})$, each item pointing at a component to change and carrying a measured direction. An LLM proposes candidate Bundles, and each candidate is compared with the current Champion in paired runs under shared conditions using the relative change $r_{i,j} = \Delta_{i,j}/\max(|q(\pi_c;g_i,z_{i,j})|,\epsilon)$; only candidates passing a four-part gate—benefit, evidence quality, tail risk, resource cost—are released. Online releases activate at the next task boundary and can be rolled back on degradation, while offline evolution runs a multi-generation population search. The YC-Bench results show both modes outperform a fixed initial Loop Policy on final funds, task success, annual survival, and drawdown.

Load-bearing premise

The empirical case rests on three official seeds on YC-Bench with no confidence intervals or significance tests, so the large measured gains could in principle come from seed randomness or from the evolved policies simply making more model calls, rather than from policy quality.

Editorial extensions

If this is right

If correct, Loop Policies become versioned assets that can be accumulated, diffed, reused, and rolled back across tasks and agent hosts.
The shared evidence-candidate-evaluation-release chain means online live feedback and offline archived traces can feed the same policy lineage.
On YC-Bench, evolved policies improve final funds, task success, and annual survival while reducing maximum drawdown, showing loop governance is a measurable performance lever.
The token accounting shows evolution has an explicit cost (29.82M and 24.02M tokens for online and offline modes) that buys longer survival and lower per-call usage at this scale.

Reading between the lines

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

A natural extension is to apply the same asset-and-gate protocol to other long-horizon interactive benchmarks, since the gate's constraints are not specific to YC-Bench.
If the release gate were made stricter or weaker, the framework predicts a direct trade-off between evolution speed and tail safety; this is testable without changing the Loop Policy representation.
The three-seed evaluation suggests that a confidence-interval-based gate, rather than point estimates, could be integrated into the release criterion itself.
Treating loop policies as versioned assets also creates an audit trail connecting each release to the trace evidence that motivated it, which could matter for deployment governance.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The headline: this is a coherent systems paper with a genuinely useful new object—the externalized, versioned Loop Policy—and a release-governance mechanism that ties self-improvement to evidence. The formal parts hold up. The empirical case does not yet support the title's claim to 'verifiable self-evolution,' and there is a possible leakage in the offline evaluation that the authors need to address.

What's new: OLE separates the full agent loop (observation, planning, memory, action, verification, recovery, stopping, budget) into a portable Bundle with version lineage, and applies a Champion–Challenger paired evaluation and a four-criterion gate (benefit, evidence quality, tail risk, resource cost) before any policy is released. The online mode adds activation-at-task-boundary and rollback. That is a real contribution, and the definitions in Section III are clear enough to implement. The paper gets credit for reporting that Fixed-π0 is 29% worse than the native baseline, for tracking token usage, and for linking to code. The evolved policies do survive longer and use fewer tokens per call.

The soft spots. The evidence is thin: one benchmark, three seeds, no confidence intervals, no significance tests, no ablation separating policy quality from the additional compute. The bigger issue is offline score leakage. Section V.A says all settings use official seeds 1/2/3 and start from the same initial state, and Algorithm 2 selects the offline successor from archived traces Doff. The paper never states whether those archived traces include the exact seeds used in the final evaluation. If they do, the offline policy was selected on outcome information from the same runs being scored, so the 974k final funds and 3/3 survival would not demonstrate reusable policy improvement. This is a possible flaw rather than a proven one, but the absence of a data-split description or limitation statement makes the offline result incomplete.

The online mode is less vulnerable because it starts from π0 and only learns from feedback during the evaluated run, but the 12 update attempts mean it also gets substantially more main-task calls, so even the online comparison is confounded.

Bottom line: the framework is worth engaging, and the formalization is a solid base. But the paper needs a clear statement of where Doff comes from, a train/evaluation split, and at minimum per-seed results plus an ablation that controls for compute. The citation pattern is fine; the issue is data provenance, not self-citation. I'd send it to review—the contribution is original and the authors can fix the empirical holes.

Referee Report

5 major / 7 minor

Summary. The paper introduces OpenLoopEvolve (OLE), a framework for externalizing and evolving 'Loop Policies' for long-horizon LLM-agent tasks. It defines a Loop Policy as an eight-component control specification (observation, planning, memory, action, verification, recovery, stopping, budget), packages it with applicability conditions and provenance in a versioned 'Bundle', and proposes online and offline evolution modes that share a Champion–Challenger paired evaluation and a robust release gate. Both modes propose candidate policies via an LLM, evaluate them in paired runs against the current Champion, and release a successor only if it passes the gate; the online mode also includes degradation monitoring and rollback. The experimental section evaluates the framework on YC-Bench with DeepSeek-V4-flash on three official seeds, reporting that OLE-online and OLE-offline improve mean final funds, task success rate, annual survival, and drawdown relative to a fixed initial Loop Policy.

Significance. If the reported results are robust, OLE would offer a useful contribution to the growing area of agent self-evolution by providing a concrete asset representation and release-governance mechanism for loop-level control experience. The paper's strengths include a clear separation between candidate generation, evaluation, release, and activation; formal definitions of traces, evidence, and policy assets; and the release of code on GitHub. However, the empirical support for the headline improvements is currently incomplete, as detailed in the major comments. With appropriate additional experiments and reporting, this line of work could be valuable to the community.

major comments (5)

[V.A, IV.C, Eq. (9)] The most serious concern is that OLE-offline may be evaluated on the very runs used to construct its archived traces. Section V.A states that all settings use official seeds 1/2/3 and start from the same initial state, but Section IV.C does not specify whether Doff—the archived traces used for offline evolution—includes runs from those same seeds and that initial state. If the archived traces overlap with the evaluation runs, then the paired evaluation in Eq. (9) and the selection in Algorithm 2 use outcome information from the exact evaluation trajectories, and the Table II gains for OLE-offline would reflect in-sample fitting rather than a reusable policy. The release gate in Eq. (10) does not rule this out because it checks task-contract compliance (benefit, evidence quality, tail risk, resource cost), not out-of-sample generalization. Please specify the composition of Doff and, if necessary, evaluate on held-out seeds or runs.
[Table II] The statistical support for the main empirical claim is thin: all aggregate results are means over three seeds, with no standard deviations, confidence intervals, or significance tests. Annual survival is a coarse 1/3 vs 2/3 vs 3/3, and the shaded regions in Fig. 3 show cross-seed ranges that are not sufficient to establish that the mean differences (e.g., 878,601.79 vs 974,627.52 vs 365,976.30) are reliable. Please report per-seed results, confidence intervals, and paired statistical tests (e.g., bootstrap or permutation tests) for the key metrics.
[Table III(a)] The comparison of OLE against Fixed-π0 in Tables II and III(a) confounds policy quality with the amount of compute: OLE-online and OLE-offline make 1575 and 1907 main-task calls against 799 for Fixed-π0, and their mean survival durations are 294 and 365 days versus 207.78 days. The higher final funds and success rates could therefore result from the larger number of task calls (i.e., longer survival) rather than from the evolution mechanism. To support the claim that evolution itself improves the policy, please add an ablation that holds the number of calls or the resource budget fixed across the compared settings—for example, a Fixed-π0 variant with a matched or increased budget, or a budget-capped OLE variant.
[IV.A, Eq. (10)] The 'verifiable' claim is not fully substantiated because the release gate is left abstract: the forms of the four constraint functions Γ_b, Γ_e, Γ_t, Γ_c and the threshold configuration Θ are never specified, and the experiments do not report how many candidates passed or failed each constraint. Since the abstract and title emphasize verifiability, please provide the concrete gate implementation and a summary of gate outcomes (e.g., for the 12 online candidate attempts and the K offline generations), including for each candidate which constraints were satisfied or violated.
[IV.B, Algorithm 1] The online evolution mechanism depends on the update condition Trig and the degradation monitor CANARYMONITOR, but neither is specified and no rollback events are reported. Without describing what Trig checks (feedback amount, period, or task stage) and the exact degradation condition used for rollback, it is difficult to assess whether the online mode's improvement comes from the proposed release/rollback governance or simply from repeatedly prompting the LLM with recent feedback. Please provide these details and a log of the online update decisions (candidate, gate result, release, rollback).

minor comments (7)

[Throughout] There are several typos and LaTeX artifacts, e.g., 'diﬀicult' (Abstract), 'oﬀicial' (Section V.A), and 'Y ao' in references [1] and [5]; these should be corrected.
[Eq. (10) and Eq. (6)] The symbol Π is used both for the Loop Policy space in Eq. (6) and as the product operator in Eq. (10), which is confusing; please use a different symbol (e.g., ⋀) for the conjunction over constraints.
[Fig. 3 caption] The statement that after early termination terminal funds are carried forward 'only to retain a fixed aggregation denominator' is unusual; please clarify the justification and note that the shaded region is a cross-seed range, not a confidence band, so readers are not misled.
[III.C, Fig. 2] The eight Loop Policy components are given only as high-level YAML descriptions; for reproducibility, please provide formal domains or a more concrete example instantiated for YC-Bench.
[V.A] The term 'oﬀicial medium configuration' for the deepseek-v4-flash model is not defined; please specify the exact inference settings (temperature, top-p, max tokens, etc.) used in the experiments.
[II.C] The claim of being 'the first to apply Champion–Challenger release mechanism to both offline and online self-evolution of externalized Loop Policies' is a strong novelty claim; please soften it or provide a more thorough comparison with related work to make the claim verifiable.
[IV–V] The gate thresholds Θ and offline search hyperparameters (number of generations, candidate count, mutation rates) are not listed; please include them in the experimental setup or appendix for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the OLE derivation is definitional and empirical, and the reported gains are measured outcomes rather than quantities forced by the method's own definitions or by self-citation.

full rationale

The paper's derivation chain is self-contained. Section III defines the long-horizon task contract g and the eight-component Loop Policy π, and Eq. (3) states the optimization objective as maximizing the value score q(π; g, z) under verification and resource constraints. Section IV constructs evolution evidence E = Φ(D) from attributable traces, generates candidate Bundles with an LLM, compares them with the Champion via paired differences Δ_i,j and relative changes r_i,j (Eq. 9), and gates release through separate benefit, evidence-quality, tail-risk, and resource-cost constraints (Eq. 10). The final YC-Bench funds are not defined by any of these equations: they are measured outcomes of the executed policies. No fitted parameter is renamed as a prediction; the update trigger and gate thresholds are hand-set protocol choices, and they do not algebraically determine the mean final funds in Table II. The cited prior work on loop engineering and Champion–Challenger management ([16]–[19]) is external to the present authors, and no load-bearing premise is justified by a self-citation or by an imported uniqueness theorem. The only substantive concern is experimental rather than circular: Section V.A states that all settings use official seeds 1/2/3 and the same initial state, while the provenance of the archived traces D_off used by OLE-offline is not specified. If those traces included the evaluation seeds, the offline comparison could suffer from in-sample leakage. That would be an external-validity or data-split flaw, not a reduction of the claimed improvement to the method's inputs by construction, and the text provides no exhibited identity between the archived traces and the evaluation runs. Therefore, no circular step is established, and the appropriate circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The framework's claims rest on definitions and empirical choices rather than a derivation. Most assumptions are domain assumptions about trace attribution, paired evaluation control, and LLM candidate quality. No physical or mathematical constants are introduced, and no parameters are fitted to data in the traditional sense; the hand-chosen gate thresholds and hyperparameters are the main free choices.

free parameters (4)

Release gate thresholds Theta = {Theta_b, Theta_e, Theta_t, Theta_c}
Hand-configured thresholds in Eq. (10) determine whether a Challenger is released; no values or sensitivity analysis are reported.
Update trigger Trig
Application-defined condition in Section IV.B decides when online evolution starts; no concrete rule or sensitivity is given.
Offline search hyperparameters = 3 candidates per round, 4 candidate rounds
Chosen for the experiment in Section V.A; no ablation of these choices is reported.
Initial Loop Policy pi0
Hand-authored initial policy used by Fixed-pi0 and both OLE modes; its content is not shown, so part of the measured gain could depend on this starting point.

assumptions (4)

domain assumption Evolution evidence Phi(D) can be extracted from traces and reliably identifies which Loop Policy component to modify.
Section III-B and Eq. (5) assume the evidence operator produces testable, attributable adjustment directions with metric changes.
domain assumption Paired evaluation under shared task g_i and controlled run condition z_{i,j} isolates the effect of the Loop Policy change.
Section IV.A and Eq. (9) assume no unmeasured confounders such as stochastic simulator noise or order effects.
domain assumption An LLM proposer can generate valid candidate policies that are improvements or informative failures.
Online and offline algorithms call G_LLM and V_LLM and assume valid Bundles are formed; no coverage guarantee is provided.
domain assumption The task contract's verification V_g and evidence sources S_g make outcomes verifiable and attributable.
Section III-A and III-B assume that verification protocols and evidence-source restrictions are sufficient to establish run outcomes.

invented entities (1)

Loop Policy and Bundle asset
purpose: Unit of loop evolution: externalized eight-component control rules with applicability conditions and version lineage.
This is a design object proposed by the paper. Its usefulness is asserted through the YC-Bench results, which are not independent of the authors' implementation and evaluation choices.

how reviews work

0 comments

Cite this review

Pith. "Pith review of OpenLoopEvolve: A Verifiable Self-Evolution Framework for Loop Policies in Long-Horizon Complex Tasks." pith.science (2026). https://pith.science/paper/CSSABSVV

@misc{pith2026260809380,
  author       = {Pith},
  title        = {Pith review of: OpenLoopEvolve: A Verifiable Self-Evolution Framework for Loop Policies in Long-Horizon Complex Tasks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CSSABSVV}},
  note         = {Machine review of arXiv:2608.09380}
}

read the original abstract

Long-horizon complex tasks require agents to repeatedly observe states, formulate plans, invoke tools, verify results, and recover from failures in continuously changing environments. However, such control experience often remains confined to a single context or a fixed prompt, and is difficult to accumulate and reuse across historical traces. This paper presents OpenLoopEvolve (OLE), a self-evolution framework centered on the Loop Policy. OLE represents an agent's observation, planning, memory, action, verification, recovery, stopping, and budget control as portable policy assets with versions and lineages, and provides online and offline evolution modes that can be selected according to practical needs: the online mode triggers candidate generation based on feedback from continuous operation, whereas the offline mode searches for candidate policies from archived traces and failure evidence. Both modes share an evolution mechanism consisting of autonomous proposals by a large language model, Champion--Challenger paired evaluation, and robust release. Policies released online are activated at a subsequent task boundary, monitored using subsequent feedback, and rolled back to their parent versions when degradation conditions are met. On the simulated business benchmark YC-Bench, both modes improve aggregate task performance, task success rate, and risk metrics relative to a fixed initial Loop Policy. The results indicate that treating the Loop Policy as a governable asset can support the accumulation, comparison, release, and reuse of control experience and improve agent performance on long-horizon complex tasks.

Figures

Figures reproduced from arXiv: 2608.09380 by the authors.

**Figure 1.** Overall OpenLoopEvolve framework. A released Loop Policy is activated at a subsequent task boundary, enters the host environment, and controls interactions in long-horizon tasks, producing attributable traces. The traces are converted into evolution evidence under a shared evidence contract and enter online or offline Loop Policy evolution. Candidates are evaluated against the Champion under identical conditions and… view at source ↗

**Figure 2.** presents a simplified Loop Policy example in a YAML-style format. This representation organizes the eight categories of loop rules into a readable and modifiable external structure, allowing Loop Policy content to be stored, compared, and reused independently of any single-run context. To support the reuse of loop-control experience across tasks, we design a Loop Policy as an external asset independent of a specifi… view at source ↗

**Figure 3.** Annual task performance of the four settings on YC-Bench. Curves and shaded regions show mean month-end funds and cross-seed ranges over three seeds, respectively; after early termination, terminal funds are carried forward only to retain a fixed aggregation denominator, not to indicate continued execution. TABLE II OVERALL RESULTS ON YC-BENCH. Method Mean final funds ($) ∆Native ∆π0 Task success rate Annual surviva… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 10 canonical work pages

[1]

ReAct: Synergizing reasoning and acting in language models,

S. Y ao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y . Cao, “ReAct: Synergizing reasoning and acting in language models,” in International Conference on Learning Representations , 2023. [Online]. Available: https://arxiv.org/abs/2210.03629 12

arXiv 2023
[2]

AgentBench: Evaluating LLMs as agents,

X. Liu, H. Yu, H. Zhang, Y . Xu, X. Lei, H. Lai, Y . Gu, H. Ding, K. Men, K. Y ang, 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,” in International Conference on Learning Representations , 2024. [Online]. Available: https://arxiv.org/abs/2308.03688

arXiv 2024
[3]

AgentBoard: An analytical evaluation board of multi-turn LLM agents,

C. Ma, J. Zhang, Z. Zhu, C. Y ang, Y . Y ang, Y . Jin, Z. Lan, L. Kong, and J. He, “AgentBoard: An analytical evaluation board of multi-turn LLM agents,” in Advances in Neural Information Processing Systems , vol. 37, 2024. [Online]. Available: https://arxiv.org/abs/2401.13178

arXiv 2024
[4]

WebArena: A realistic web environment for building autonomous agents,

S. Zhou, F. F. Xu, H. Zhu, X. Zhou, R. Lo, A. Sridhar, X. Cheng, T. Ou, Y . Bisk, D. Fried, U. Alon, and G. Neubig, “WebArena: A realistic web environment for building autonomous agents,” in International Conference on Learning Representations , 2024. [Online]. Available: https://arxiv.org/abs/2307.13854

arXiv 2024
[5]

τ -bench: A benchmark for tool-agent-user interaction in real-world domains,

S. Y ao, N. Shinn, P . Razavi, and K. Narasimhan, “ τ -bench: A benchmark for tool-agent-user interaction in real-world domains,” in International Conference on Learning Representations , 2025. [Online]. Available: https://arxiv.org/abs/2406.12045

arXiv 2025
[6]

YC-Bench: Benchmarking AI agents for long-term planning and consistent execution,

M. He, A. Jain, A. Kumar, V . Tu, S. Bakshi, S. Patro, and N. Rajani, “YC-Bench: Benchmarking AI agents for long-term planning and consistent execution,” arXiv preprint arXiv:2604.01212 , 2026. [Online]. Available: https://arxiv.org/abs/2604.01212

arXiv 2026
[7]

Generative agents: Interactive simulacra of human behavior,

J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P . Liang, and M. S. Bernstein, “Generative agents: Interactive simulacra of human behavior,” in Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology , 2023. [Online]. Available: https://doi.org/10.1145/3586183.3606763

arXiv 2023
[8]

Reflexion: Language agents with verbal reinforcement learning,

N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, and S. Y ao, “Reflexion: Language agents with verbal reinforcement learning,” in Advances in Neural Information Processing Systems , vol. 36, 2023. [On- line]. Available: https://proceedings.neurips.cc/paper_files/paper/2023/ hash/1b44b878bb782e6954cd888628510e90-Abstract-Conference.html

work page 2023

Show all 20 references

[9]

ExpeL: LLM agents are experiential learners,

A. Zhao, D. Huang, Q. Xu, M. Lin, Y .-J. Liu, and G. Huang, “ExpeL: LLM agents are experiential learners,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 17, 2024, pp. 19 632–19 642. [Online]. Available: https://ojs.aaai.org/index.php/ AAAI/art...

2024
[10]

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

G. Wang, Y . Xie, Y . Jiang, A. Mandlekar, C. Xiao, Y . Zhu, L. Fan, and A. Anandkumar, “Voyager: An open-ended embodied agent with large language models,” Transactions on Machine Learning Research , 2024. [Online]. Available: https://openreview.net/forum?id=ehfRiF0R3a

2024
[11]

DSPy: Compiling declarative language model calls into self-improving pipelines,

O. Khattab, A. Singhvi, P . Maheshwari, Z. Zhang, K. Santhanam, S. V . A, S. Haq, A. Sharma, T. Joshi, H. Moazam, H. Miller, M. Zaharia, and C. Potts, “DSPy: Compiling declarative language model calls into self-improving pipelines,” in International Conference on Learning Repr...

2024
[12]

Automated design of agentic systems,

S. Hu, C. Lu, and J. Clune, “Automated design of agentic systems,” in International Conference on Learning Representations , 2025. [Online]. Available: https://arxiv.org/abs/2408.08435

2025 arXiv
[13]

AFlow: Automating agentic workflow generation,

J. Zhang, J. Xiang, Z. Yu, F. Teng, X. Chen, J. Chen, M. Zhuge, X. Cheng, S. Hong, J. Wang, B. Zheng, B. Liu, Y . Luo, and C. Wu, “AFlow: Automating agentic workflow generation,” in International Conference on Learning Representations , 2025. [Online]. Available: https://proce...

2025
[14]

AgentSquare: Automatic LLM agent search in modular design space,

Y . Shang, Y . Li, K. Zhao, L. Ma, J. Liu, F. Xu, and Y . Li, “AgentSquare: Automatic LLM agent search in modular design space,” in International Conference on Learning Representations , 2025. [Online]. Available: https://proceedings.iclr.cc/paper_files/paper/2025/ hash/0ae940...

2025
[15]

SE-Agent: Self-evolution trajectory optimization in multi-step reasoning with LLM-based agents,

Y . Guo, J. Lin, H. Wang, Y . Han, S. Hu, Z. Ni, L. Wang, and M. Chen, “SE-Agent: Self-evolution trajectory optimization in multi-step reasoning with LLM-based agents,” in Advances in Neural Information Processing Systems , vol. 38, 2025. [Online]. Available: https://proceedin...

2025
[17]

Proof-or-stop: Don’t trust the agent, trust the evidence— loop engineering for verifiable evidence-gated lifecycle control,

J. Huang, J. Hsia, J. Sun, F. Shi, W. Huang, and I. H. White, “Proof-or-stop: Don’t trust the agent, trust the evidence— loop engineering for verifiable evidence-gated lifecycle control,” arXiv preprint arXiv:2607.14890 , 2026. [Online]. Available: https: //arxiv.org/abs/2607.14890

2026 arXiv
[18]

Champion-challenger based predictive model selection,

S. V . Nath, “Champion-challenger based predictive model selection,” in Proceedings 2007 IEEE SoutheastCon , 2007, pp. 254–254. [Online]. Available: https://doi.org/10.1109/SECON.2007.342897

2007
[19]

ChaCha for online AutoML,

Q. Wu, C. Wang, J. Langford, P . Mineiro, and M. Rossi, “ChaCha for online AutoML,” in Proceedings of the 38th International Conference on Machine Learning , ser. Proceedings of Machine Learning Research, vol. 139, 2021, pp. 11 263–11 273. [Online]. Available: https://proceedi...

2021
[20]

DeepSeek-V4: Towards highly eﬀicient million-token context intelligence,

DeepSeek-AI, “DeepSeek-V4: Towards highly eﬀicient million-token context intelligence,” arXiv preprint arXiv:2606.19348 , 2026. [Online]. Available: https://arxiv.org/abs/2606.19348

2026
[2026]

Available: https://arxiv.org/abs/2607.00038

[Online]. Available: https://arxiv.org/abs/2607.00038

arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

ReAct: Synergizing reasoning and acting in language models,

S. Y ao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y . Cao, “ReAct: Synergizing reasoning and acting in language models,” in International Conference on Learning Representations , 2023. [Online]. Available: https://arxiv.org/abs/2210.03629 12

arXiv 2023

[2] [2]

AgentBench: Evaluating LLMs as agents,

X. Liu, H. Yu, H. Zhang, Y . Xu, X. Lei, H. Lai, Y . Gu, H. Ding, K. Men, K. Y ang, 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,” in International Conference on Learning Representations , 2024. [Online]. Available: https://arxiv.org/abs/2308.03688

arXiv 2024

[3] [3]

AgentBoard: An analytical evaluation board of multi-turn LLM agents,

C. Ma, J. Zhang, Z. Zhu, C. Y ang, Y . Y ang, Y . Jin, Z. Lan, L. Kong, and J. He, “AgentBoard: An analytical evaluation board of multi-turn LLM agents,” in Advances in Neural Information Processing Systems , vol. 37, 2024. [Online]. Available: https://arxiv.org/abs/2401.13178

arXiv 2024

[4] [4]

WebArena: A realistic web environment for building autonomous agents,

S. Zhou, F. F. Xu, H. Zhu, X. Zhou, R. Lo, A. Sridhar, X. Cheng, T. Ou, Y . Bisk, D. Fried, U. Alon, and G. Neubig, “WebArena: A realistic web environment for building autonomous agents,” in International Conference on Learning Representations , 2024. [Online]. Available: https://arxiv.org/abs/2307.13854

arXiv 2024

[5] [5]

τ -bench: A benchmark for tool-agent-user interaction in real-world domains,

S. Y ao, N. Shinn, P . Razavi, and K. Narasimhan, “ τ -bench: A benchmark for tool-agent-user interaction in real-world domains,” in International Conference on Learning Representations , 2025. [Online]. Available: https://arxiv.org/abs/2406.12045

arXiv 2025

[6] [6]

YC-Bench: Benchmarking AI agents for long-term planning and consistent execution,

M. He, A. Jain, A. Kumar, V . Tu, S. Bakshi, S. Patro, and N. Rajani, “YC-Bench: Benchmarking AI agents for long-term planning and consistent execution,” arXiv preprint arXiv:2604.01212 , 2026. [Online]. Available: https://arxiv.org/abs/2604.01212

arXiv 2026

[7] [7]

Generative agents: Interactive simulacra of human behavior,

J. S. Park, J. C. O’Brien, C. J. Cai, M. R. Morris, P . Liang, and M. S. Bernstein, “Generative agents: Interactive simulacra of human behavior,” in Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology , 2023. [Online]. Available: https://doi.org/10.1145/3586183.3606763

arXiv 2023

[8] [8]

Reflexion: Language agents with verbal reinforcement learning,

N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, and S. Y ao, “Reflexion: Language agents with verbal reinforcement learning,” in Advances in Neural Information Processing Systems , vol. 36, 2023. [On- line]. Available: https://proceedings.neurips.cc/paper_files/paper/2023/ hash/1b44b878bb782e6954cd888628510e90-Abstract-Conference.html

work page 2023