Machine Learning · arXiv:2609.37105 v1 · submitted 29 Sep 2026 · 17 references
VACE: Validation-Gated Alternating Co-Evolution of Agent Models and Harnesses
The paper says
“Language model agents can be improved by updating their model weights or refining the harness that guides task execution.”
more of the abstract
These components are coupled: weight updates change how the model uses the harness, while harness updates change the trajectories used for training. We propose VACE, Validation-Gated Alternating CoEvolution, which alternates agentic reinforcement learning with trajectory-driven harness refinement. After each RL stage, VACE reuses the collected trajectories to propose a harness revision and evaluates the incumbent and candidate with the updated model held fixed. The candidate guides subsequent training only if it improves validation performance. With Qwen3.5-9B, VACE achieves 45.26% test accuracy on OfficeQA and a mean partial-credit score of 75.19% on AutomationBench, exceeding weight-only RL by 6.43 and 9.09 percentage points and ungated alternation by 4.59 and 6.95 points, respectively. Across 44 harness proposals, 17 reduce validation performance at the updated checkpoint and are rejected before subsequent RL training, highlighting the importance of validation gating.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 30 Sep 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 clean, honest study of a genuinely new validation-gated alternation schedule, but the headline gaps sit on noise-sized deltas from tiny reused validation splits, so the central claim is plausible rather than settled. This line is wrong
3 major objections · 4 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 17 references say.
The paper says, at §4.4, Eq. 5-6, Table 4
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.74):
“Validation reuse and evaluation noise.Repeated use of the validation set for harness acceptance can introduce adaptive selection effects [Dwork et al., 2015].”
Pith says: Objection 1 of 3 · major
The gate decides on single-evaluation point estimates from 53 (OfficeQA) and 58 (AutomationBench) validation tasks. For a 53-task binary split, a single accuracy estimate has a standard error of roughly 6 pp at p≈0.4; the paired same-checkpoint design reduces the variance of the difference, but the accepted OfficeQA delta of +3.77 pp and the headline VACE-vs-WHALE gaps (4.59 and 6.95 pp) do not convincingly clear the residual noise floor.
more of objection 1
The manuscript acknowledges evaluation noise and adaptive-selection risk (Sec. 5, b131) but neither quantifies whether the observed deltas exceed the residual noise nor provides multi-seed replication or significance testing. Without that, the 17/44 rejection count and the gating benefit cannot be distinguished from favorable draws on a reused split, and the central claim that the gate is load-bearing is not yet established.
The paper says, at Table 5 / Appendix A
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
The RL and harness-proposal compute caps BW and BH are left numerically unspecified. Since the abstract's first-cited result is the VACE-vs-RL-only gap (6.43/9.09 pp), and RL-only consumes BW while VACE consumes BW+BH, the gap over weight-only RL cannot currently be attributed to the gate rather than to unequal total training effort.
more of objection 2
The VACE-vs-WHALE comparison is better controlled (same caps, the gate is the only difference), which partially mitigates the concern, but the actual cap values are needed to interpret both comparisons, and the paper should state explicitly whether BW in the VACE condition equals BW in the RL-only condition.
The paper says, at §4.1
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.60):
“Validation reuse and evaluation noise.Repeated use of the validation set for harness acceptance can introduce adaptive selection effects [Dwork et al., 2015].”
Pith says: Objection 3 of 3 · major
The same validation split drives both the gate decisions and the best-checkpoint selection ('select the checkpoint and corresponding harness with the highest validation score'), and test scores are the mean of three evaluations of a single selected pair, with no per-seed error bars on the headline differences.
more of objection 3
The paired, same-checkpoint evaluation of incumbent vs. candidate is a genuine strength, but the reported 4.59/6.95 pp gating benefit needs either across-seed replication or resampling-based uncertainty on the test scores before the reader can assess whether the improvements exceed evaluation noise.
4 minor objections
The paper says, at §2 / §3.2
Pith says: minor objection 1 of 4
The harness proposer S is imported from HarnessEvolve [Jiang et al., 2026], a concurrent arXiv preprint with substantial author overlap with the present submission. The relationship should be stated explicitly, and the contribution boundary between the two works should be drawn in the text.
The paper says, at Abstract / §4.4
Pith says: minor objection 2 of 4
The phrasing '17 reduce validation performance ... highlighting the importance of validation gating' should be qualified as 'on the validation split', since the paper itself acknowledges (Sec. 5) that the gate relies on noisy point estimates; as written it overstates what the rejections establish.
show 2 more
The paper says, at Throughout
Pith says: minor objection 3 of 4
The text contains rendering artifacts ('V ACE' with a stray space in many headings) and should be normalized; the captions of Table 3 ('A/T/R') could spell out accepted/tied/rejected for readability, and Figure 3 vs. Figure 5 should share a single statement of which validation score is reported (retained pair vs. best eligible).
The paper says, at §4.3
Pith says: minor objection 4 of 4
The WHALE-style and SIA-style conditions are reimplementations that 'adapt the scheduling ideas' within the authors' own optimizers rather than the published systems; this should be stated in the main text near Table 2 (it currently appears only in §4.2 and Appendix A) so the comparisons are not read as evaluations of the original methods.
What would settle it
Run VACE and WHALE-style with 5 independent seeds under the same stated numeric BW and BH budgets, and report the paired difference in mean test score per seed with the seed-to-seed standard deviation (plus a paired significance test); also bootstrap the per-round validation outcomes to quantify how often the Eq. 6 accept/reject decision would flip under evaluation noise. If the VACE-vs-WHALE test gap stays above roughly 2x the per-seed SD, the gating claim is supported; if it falls within one SD, the benefit is within measurement noise and the central claim is not established.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 30 Sep 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 4 minor
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
VACE is worth a read. The new thing is precise: each harness proposal is evaluated against the incumbent at the updated checkpoint, accepted only on strict validation improvement, and the rejected proposals never shape the next RL round. That distinguishes it from SIA, Co-Harness, and WHALE. The 17-of-44 rejection rate and the consistent gap over ungated alternation on both benchmarks are exactly what the construction predicts. The paper is cleanly written, Algorithm 1 is what the experiments actually run, and the round-level analysis in Table 4 is consistent with the method's behavior. The soft spots are real and the authors mostly name them. The stress-test arithmetic holds up: with 53 and 58 validation tasks, the standard error of a single accuracy estimate is near 6 percentage points, and the accepted deltas (e.g., +3.77 on OfficeQA) and the headline 4.59 and 6.95 point gaps over WHALE are the same order of magnitude. The same split drives both harness acceptance and checkpoint selection, and the reported means are three evaluations of one selected pair with no multi-seed replication or significance testing. The authors cite the adaptive-selection risk (b131, Dwork et al.) but do not quantify whether observed deltas clear the noise floor. That is the load-bearing gap: we cannot yet tell gating from favorable noise on a reused split. Two smaller issues. Table 5 leaves BW and BH numerically unspecified, so the RL-only gap could partly reflect unequal training effort rather than the gate. And the harness proposer S is imported from HarnessEvolve, an unpublished concurrent preprint with overlapping authors; the empirical claims are tested on external test splits so it is not circular, but it does make the revision mechanism hard to audit independently. Neither is fatal; both are fixable with code, data, numeric caps, and more seeds. Who should read it: anyone working on agent self-improvement or coupled weight-and-prompt optimization. It deserves a serious referee, not a desk reject, with the reasonable ask that the authors release artifacts and add multi-seed significance reporting before the gating claim is taken as established. I would not cite the headline numbers yet, but I would cite the method as related work.
The strongest objection
The central claim that validation gating is load-bearing rests on the VACE-vs-WHALE test gaps (4.59 pp OfficeQA, 6.95 pp AutomationBench) and on the 17/44 rejections being meaningful. Both hinge on the gate's repeated decisions under Eq. 6 (accept iff c_t - b_t > 0), where b_t and c_t are single-point estimates on |V| = 53 (OfficeQA) and 58 (AutomationBench) tasks (Eq. 5, Eq. 6). For a 53-task binary split, the standard error of a single accuracy estimate is roughly sqrt(p(1-p)/53), about 6 pp at p = 0.4; the paired difference has smaller but still material variability. The accepted delta in Table 4 is +3.77 pp (OfficeQA) and the first rejected is -11.31 pp; many of the 44 gate decisions operate on deltas that do not convincingly clear this noise floor. The same split drives both harness acceptance and best-checkpoint selection, and test scores are the mean of three evaluations of a single selected pair (Sec. 4.1) with no significance testing. The authors acknowledge the noise and adaptive-selection risk (b131, citing Dwork et al.) but do not quantify whether the observed deltas exceed the evaluation noise or provide multi-seed replication, so whether the gate selects genuinely better harnesses, or merely harnesses that drew favorable noise on a reused split, remains unresolved. Separately, Table 5 leaves BW and BH numerically unspecified, so the gap over RL-only cannot currently be attributed to the gate rather than to unequal training effort.
The reader's weakest_assumption — that small, repeatedly reused validation splits make the strict-improvement gate unreliable — is precisely the most load-bearing condition for the central claim, and the paper's own evidence (Table 4 deltas, 53/58-task splits, no significance testing) does not refute it. The authors do state the noise/adaptive-selection caveat (b131) and the compute-cap caveat (unspecified BW/BH in Table 5), but an acknowledged limitation still leaves the claim conditional: the headlined gaps (4.59/6.95 pp over ungated alternation) are on the same scale as plausible single-evaluation noise for these split sizes, so without multi-seed replication we cannot tell whether gating genuinely improves test performance or whether acceptance decisions are tracking noise on a reused split. I considered whether unequal compute is the sharper attack; it is real but chiefly threatens the VACE-vs-RL-only comparison (BH = 0 there), whereas the claim that gating itself is load-bearing is best isolated by VACE-vs-WHALE, where budgets are matched by construction if BW/BH are equal as titled. The decisive check is therefore multi-seed, matched-budget replication with explicit numeric budgets and a noise-floor analysis of the gate decisions. Since this matches the reader's identification and does not escalate the verdict, CONDITIONAL is retained.
The referee's report
If the gating benefit is real, this is a clean and practically actionable design contribution: separating harness proposal from acceptance, and evaluating the candidate with the updated checkpoint held fixed (paired, identical task IDs) before it shapes the next RL stage. Strengths include internally consistent numbers (Tables 2-4 reconcile with the reported deltas), a controlled comparison across scheduling variants built on the same weight and harness optimizers, explicit accounting of accept/tie/reject counts, and an honest acknowledgment of validation-reuse and evaluation-noise risk (Sec. 5, citing Dwork et al.). The submission is a preprint with no released code, data, or model artifacts, and the evidence base is one model and two benchmarks; the headline result, if it survives replication, would be of broad interest to the agent-training community.
Pith's reading of the paper (a machine's summary, not the authors')
The paper claims that when an agent's model weights and its "harness" — the instructions, skills, and execution rules wrapped around the model — are improved together by alternating reinforcement learning with trajectory-driven harness revisions, the acceptance step matters as much as the revision itself. VACE holds the freshly trained model fixed and compares the incumbent and candidate harness on the same validation tasks, adopting the candidate only if its score strictly improves. This validation gate is the paper's central ingredient: with Qwen3.5-9B it reaches 45.26% on OfficeQA and 75.19% on AutomationBench, beating weight-only RL by 6.43 and 9.09 points and ungated alternation by 4.59 and 6.95 points. Across 44 harness proposals, 17 lowered validation at the updated checkpoint and were rejected, evidence that proposals routinely regress and need an explicit decision before shaping the next training round.
What carries the argument. The validation gate is a checkpoint-specific comparison that decouples the proposal decision from the acceptance decision. Given the updated weights $W_{t+1}$, the incumbent and candidate harnesses are re-scored on the same validation tasks $V$ with the same evaluator, and the strict-improvement rule $\Delta_t^H = c_t - b_t > 0$ decides whether the candidate guides the next RL round, with ties retaining the incumbent. Its job is to keep the trajectory-diagnostic proposal stage (what to change, supplied by a HarnessEvolve-style proposer $S$) separate from the acceptance stage (whether the change helps with the model that will actually use it), so that regressive revisions never contaminate the next round's training trajectories.
Core claim. The central claim is that validation gating is the load-bearing component of model–harness co-evolution, not just a cosmetic check. The model and harness are coupled through the trajectory distribution $p_{W,H}$: a weight update changes how the model uses the harness, and a harness update changes the trajectories available for future training. VACE formalizes the loop as alternating updates to the pair $(W,H)$: stage I runs a block of agentic RL with the harness fixed, stage II reuses the collected trajectories to propose a revision $H'_t = S(H_t, T_t)$, and stage III compares $b_t = \widehat{R}_V(W_{t+1}, H_t)$ with $c_t = \widehat{R}_V(W_{t+1}, H'_t)$ on the same validation tasks with the same evaluator, adopting the revision only when $c_t > b_t$ (strict improvement, Eq. 6) and keeping the incumbent on ties. The empirical assertion is that this gate, not raw alternation, accounts for the gains: gated alternation exceeds ungated alternation by 4.59 and 6.95 points, and rejects 17 of 44 proposals that would otherwise have guided the next RL stage. The paper also reports that joint adaptation beats either single-component route, with VACE ahead of harness-only optimization by 9.98 and 13.58 points.
The premise it rests on. The result rests on the validation gate being a trustworthy signal on small, repeatedly-reused splits (53 OfficeQA and 58 AutomationBench tasks), where the strict-improvement rule compares noisy point estimates and the same split drives both harness acceptance and checkpoint selection; if gate noise dominates, accepted harnesses may not genuinely improve, and the unspecified RL compute caps mean part of the gap could reflect training effort rather than the gate.
If the paper is right (Pith's inferences, not the authors')
If validation gating is the load-bearing ingredient, co-evolution frameworks should budget for acceptance decisions, not just proposals, and checkpoint-specific evaluation becomes a design requirement rather than a nicety.
Joint adaptation dominates both single-component routes — VACE beats harness-only by 9.98/13.58 points and RL-only by 6.43/9.09 points — implying that the coupling the paper formalizes produces real gains, not incidental ones.
The 17-of-44 rejection rate implies that trajectory-driven harness proposals routinely regress at the checkpoint they target, so harness optimization without an outer decision layer is likely to degrade training.
Since harness quality is checkpoint-dependent, evaluation must be performed at the updated model to be meaningful; the paper's own suggestion of cross-evaluating checkpoints and harness versions would test how compatibility changes during training.
A strict-improvement gate with ties kept in favor of the incumbent is a conservative rule: it can only slow harness evolution relative to a tolerance-aware rule, so the reported gains are a floor for what acceptance control can buy.
The gate operates partly as a variance filter on the harness search: proposals that overfit the training trajectories get rejected on held-out validation, so part of the benefit may be search stabilization rather than raw quality control; a testable extension is to compare gated alternation against strong regularization of the proposer itself.
Because the same validation split drives both harness acceptance and checkpoint selection, the reported gap is vulnerable to adaptive selection; a clean test would reserve a fresh holdout to certify that accepted harnesses transfer beyond the reused split.
The framework suggests a general principle for coupled discrete/continuous optimization: whenever a discrete proposal changes the input distribution of a later gradient step, gate the proposal against the held-fixed intermediate — a pattern that could transfer to prompt-pool selection, tool-interface redesign, or curriculum scheduling in agentic RL.
The numerically unspecified RL compute caps imply the comparison with RL-only may mix the gate's effect with unequal training effort; equating budgets across conditions would separate the gate's contribution from raw compute.
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
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.
The paper
Abstract
Language model agents can be improved by updating their model weights or refining the harness that guides task execution. These components are coupled: weight updates change how the model uses the harness, while harness updates change the trajectories used for training. We propose VACE, Validation-Gated Alternating CoEvolution, which alternates agentic reinforcement learning with trajectory-driven harness refinement. After each RL stage, VACE reuses the collected trajectories to propose a harness revision and evaluates the incumbent and candidate with the updated model held fixed. The candidate guides subsequent training only if it improves validation performance. With Qwen3.5-9B, VACE achieves 45.26% test accuracy on OfficeQA and a mean partial-credit score of 75.19% on AutomationBench, exceeding weight-only RL by 6.43 and 9.09 percentage points and ungated alternation by 4.59 and 6.95 points, respectively. Across 44 harness proposals, 17 reduce validation performance at the updated checkpoint and are rejected before subsequent RL training, highlighting the importance of validation gating.
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 · 2026-09-30
The authors' main claim, as the classifier picked it out 0.94
“With Qwen3.5-9B, VACE achieves 45.26% test accuracy on OfficeQA and a mean partial-credit score of 75.19% on AutomationBench, exceeding weight-only RL by 6.43 and 9.09 percentage points and ungated alternation by 4.59 and 6.95 points, respectively.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.31 · human participants 0.03 · controlled comparison 0.95 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · 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.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Language model agents can be improved by updating their model weights or refining the harness that guides task execution. | background 1.00 | 0.05 | 0.03 | asserted 0.71 |
| These components are coupled: weight updates change how the model uses the harness, while harness updates change the trajectories used for training. | background 1.00 | 0.11 | 0.03 | asserted 1.00 |
| We propose VACE, Validation-Gated Alternating CoEvolution, which alternates agentic reinforcement learning with trajectory-driven harness refinement. | contribution 0.85 | 0.82 | 0.09 | asserted 0.98 |
| After each RL stage, VACE reuses the collected trajectories to propose a harness revision and evaluates the incumbent and candidate with the updated model held fixed. | method 1.00 | 0.92 | 0.06 | asserted 1.00 |
| The candidate guides subsequent training only if it improves validation performance. | method 0.99 | 0.83 | 0.12 | asserted 1.00 |
| With Qwen3.5-9B, VACE achieves 45.26% test accuracy on OfficeQA and a mean partial-credit score of 75.19% on AutomationBench, exceeding weight-only RL by 6.43 and 9.09 percentage points and ungated alternation by 4.59 and 6.95 points, respectively. | result 1.00 | 0.96 | 0.54 | asserted 1.00 |
| Across 44 harness proposals, 17 reduce validation performance at the updated checkpoint and are rejected before subsequent RL training, highlighting the importance of validation gating. | result 0.82 | 0.96 | 0.14 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.95
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, 2026-09-30: 1 request, 5,663 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 (5)
Figures are the authors', reproduced from the paper.
References (17)
- [5] URLhttps://arxiv.org/abs/2605.27276. Wen Jiang, Mingmin Chu, Yimeng Tian, Qianxin Zhang, Haofei Yang, Rui Yang, Yang Liu, Tao Lv, and Fangming Li. HarnessEvolve: Learning from reference trajectories for reliable agent self-evolution, on PithPith's reviewarXiv
- [6] URLhttps://arxiv.org/abs/2609.00829. Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. Search-R1: Training LLMs to reason and leverage search engines with reinforcement learning.arXiv preprint arXiv:2503.09516, on PithPith's reviewarXiv
- [7] URLhttps://arxiv.org/abs/2503.09516. Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. DSPy: Compiling declarative... on PithPith's reviewarXiv
- [9] URLhttps://arxiv.org/abs/2609.00196. Xufang Luo, Yuge Zhang, Zhiyuan He, Zilong Wang, Siyun Zhao, Dongsheng Li, Luna K. Qiu, and Yuqing Yang. Agent lightning: Train ANY AI agents with reinforcement learning.arXiv preprint arXiv:2508.03680, on PithPith's reviewarXiv
- [10] URLhttps://arxiv.org/abs/2508.03680. Cheng Qian, Emre Can Acikgoz, Qi He, Hongru Wang, Xiusi Chen, Dilek Hakkani-Tür, Gokhan Tur, and Heng Ji. ToolRL: Reward is all tool learning needs.arXiv preprint arXiv:2504.13958, on PithPith's reviewarXiv
- [11] URL https://arxiv.org/abs/2504.13958. Qwen Team. Qwen3.5: Towards native multimodal agents, February on PithPith's reviewarXiv
- [14] URL https://arxiv.org/abs/2604.18934. Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning.arXiv preprint arXiv:2303.11366, on PithPith's reviewarXiv
- [15] URLhttps://arxiv.org/abs/2303.11366. 9 Arnav Singhvi, Krista Opsahl-Ong, Jasmine Collins, Ivan Zhou, Cindy Wang, Ashutosh Baheti, Jacob Portes, Sam Havens, Erich Elsen, Michael Bendersky, Matei Zaharia, and Xing Chen. Introducing OfficeQA: A benchmark for end-to-end grounded r... on PithPith's reviewarXiv
- [17] URL https://arxiv.org/abs/2504.20073. Zhepei Wei, Wenlin Yao, Yao Liu, Weizhi Zhang, Qin Lu, Liang Qiu, Changlong Yu, Puyang Xu, Chao Zhang, Bing Yin, Hyokun Yun, and Lihong Li. WebAgent-R1: Training web agents via end-to-end multi-turn reinforcement learning.arXiv preprint ar... on PithPith's reviewarXiv
- [18] URL https://arxiv.org/abs/2505.16421. Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, Hang Zhu,... on PitharXiv
- [19] URLhttps://arxiv.org/abs/2503.14476. Mert Yuksekgonul, Federico Bianchi, Joseph Boen, Sheng Liu, Zhi Huang, Carlos Guestrin, and James Zou. TextGrad: Automatic “differentiation” via text.arXiv preprint arXiv:2406.07496, on PithPith's reviewarXiv
- [20] URL https://arxiv.org/abs/2406.07496. Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, and Jeff Clune. Darwin Gödel machine: Open-ended evolution of self-improving agents.arXiv preprint arXiv:2505.22954, 2025a. URL https://arxiv.org/abs/2505.22954. Qizheng Zhang, Changran Hu, ... on PithPith's reviewarXiv
- [21] URL https://arxiv.org/abs/2308.10144. 10 A Optimization settings Table 5 summarizes the optimization settings for the compared methods. Table 5: Optimization settings.B W andBH denote the planned RL and harness-proposal caps. Method RL cap H cap Harness gate Static agent 0 0 —... on PithPith's reviewarXiv
- [2015] URL https://arxiv.org/abs/1506.02629. Prannay Hebbar, Yogendra Manawat, Samuel Verboomen, Alesia Ivanova, Selvam Palanimalai, Kunal Bhatia, and Vignesh Baskaran. SIA: Self improving AI with harness & weight updates.arXiv preprint arXiv:2605.27276, on PithPith's reviewarXiv
- [2023] URLhttps://arxiv.org/abs/2310.03714. Haechan Kim, Yoonho Lee, Gisang Lee, Chelsea Finn, and Kangwook Lee. WHALE: A simple recipe for joint harness-weight optimization, on PithPith's reviewarXiv
- [2024] URL https://arxiv.org/abs/2402.03300. Daniel Shepard and Robin Salimans. AutomationBench.arXiv preprint arXiv:2604.18934, on PithPith's reviewarXiv
- [2025] URL https://arxiv.org/abs/2507.19457. Haotian Chen, Xin Cong, Shengda Fan, Yuyang Fu, Ziqin Gong, Yaxi Lu, Yishan Li, Boye Niu, Chengjun Pan, Zijun Song, Huadong Wang, Yesai Wu, Yueying Wu, Zihao Xie, Yukun Yan, Zhong Zhang, Yankai Lin, Zhiyuan Liu, and Maosong Sun. AgentCPM-E... on PithPith's reviewarXiv
Cited by 3 Pith papers
Reviewed papers on Pith that cite this one. The full citation record
- ModHarness-RL: Recasting Harnesses as Learnable and Evolving Programs for Agentic RL cs.AI · Nov 2026
- CoEvo-RSI: When Harnesses Evolve, Which Models Should Follow? cs.LG · Oct 2026
- A Survey on RLOps for Large Language Models cs.SE · Nov 2026
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.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment