Artificial Intelligence · arXiv:2610.18083 v1 · submitted 13 Oct 2026 · 39 references
AReaL-RSI: Adaptive Scaling and Coordination of Recursive Self-Improvement
The paper says
“Scaling recursive self-improvement requires additional computation to produce useful research progress in reaching persistent objectives.”
more of the abstract
Intuitively, more researcher agents can explore more hypotheses, but early evaluation scores might misrepresent their promise, shared findings could lose their applicability conditions, and useful procedures could remain confined to completed investigations. In this paper, we present AReaL-RSI, a framework for organizing a persistent research team of agents without updating backbone LLM weights. We formulate the scaling problem in terms of evaluated outcomes and whole-system resource expenditure. Concretely, a leader agent selects task-appropriate research orientations and guides ongoing investigations; conditional memory sharing distinguishes unresolved problems from supported findings and enables recipient adaptation. During recursive evolution, after each campaign, evidence-based review revises researcher orientation, knowledge, and skills for later campaigns. Experiments compare AReaL-RSI with Codex-Goal on downstream harness optimization, GPU kernel optimization, and AI research tasks using two backbones. Under the reported resource allocations, AReaL-RSI improves Terminal Bench performance by 2.2 percentage points for each backbone and increases mean KernelBench speedup by 32.2-36.8%; further scalability comparisons favor the design of AReaL-RSI over native scaling: combining experience sharing with specialized researcher roles yields 1.9-27.5% lower kernel latency than independent exploration and 1.3-33.9% lower latency than sharing alone across all seven comparable task-backbone pairs.
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 well-designed framework for coordinating RSI agents with a clear formulation and interesting mechanisms, but the evaluation lacks the statistical backbone to support the strong comparative claims. This line is wrong
3 major objections · 6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 39 references say.
The paper says, at §5.1, Table 1, Table 2
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.94):
“Comparisons are descriptive; repeated-run uncertainty is unavailable. 5.2 End-to-End Performance Compared with CODEX-GOAL Comparison.Table 1 addresses RQ1 within each research backbone.”
Pith says: Objection 1 of 3 · major
The evaluation lacks any repeated-run statistics. The paper explicitly states 'Comparisons are descriptive; repeated-run uncertainty is unavailable' (b145). All performance comparisons rest on single trials without error bars, confidence intervals, or hypothesis tests. Given the stochastic nature of LLM calls, leader decisions, and candidate evaluations, the central claim that 'AReaL-RSI outperforms Codex-Goal on all six metrics with both backbones' (b159) is not supported with sufficient statistical evidence.
more of objection 1
For example, the Astra jailbreak score difference of 0.119 (81.261 vs. 81.142) and several KernelBench latency differences below 2% (e.g., Astra L3P19: 0.479 vs. 0.473 ms) could easily be noise. Without quantifying uncertainty, the reader cannot assess whether the observed advantages are reliable or within measurement noise. This is a load-bearing limitation that must be addressed for the empirical claims to be substantiated.
The paper says, at §5.3, Table 2
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.90):
“This interpretation is consistent with task-conditioned harness selection and conditional memory sharing, but the comparisons do not isolate individual mechanism effects or measure research-direction diversity.”
Pith says: Objection 2 of 3 · major
The RQ2 comparisons of parallel configurations (independent, sharing, sharing+roles) are also based on single runs. The paper acknowledges that these comparisons 'do not isolate individual mechanism effects or measure research-direction diversity' (b174). This means the relative contributions of the proposed mechanisms (conditional memory sharing, specialization) are not disentangled.
more of objection 2
Combined with the lack of repeated runs, the evidence for the claimed ranking (sharing+roles best in all seven pairs) is weak. The authors should either provide repeated-run statistics or explicitly soften the conclusions to reflect the preliminary nature of the comparisons.
The paper says, at §5.2, Table 1
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
The Astra jailbreak score improvement from Codex-Goal (81.142) to AReaL-RSI (81.261) is only 0.119 points, which is likely within the noise of a single trial. Including such a small difference in the list of '12 favorable comparisons' across all metrics gives a misleading impression of uniform improvement. The authors should either provide error bars to show this difference is meaningful or acknowledge that not all improvements are practically significant.
6 minor objections
The paper says, at §5.1
Pith says: minor objection 1 of 6
The term 'Terminal Bench' in the abstract and later 'Terminus harness optimization' in §5.1 should be harmonized for clarity.
The paper says, at §5.1, Metrics
Pith says: minor objection 2 of 6
The notation 'PPL↓' for data curation perplexity is clear, but the arrow notation could be explained in the caption of Table 1 for completeness (though it is conventional).
show 4 more
The paper says, at §2.1, Eq. (1)
Pith says: minor objection 3 of 6
The per-campaign objective is described as defining a comparison rather than a runtime optimizer. This could be clarified earlier to avoid the impression that the system solves Eq. (1) directly.
The paper says, at §5.1, Evaluation Protocol
Pith says: minor objection 4 of 6
It is stated that for KernelBench, seed 137 was used for independent final reevaluation. It would be helpful to mention whether the same seed was used across all methods to ensure fair comparison.
The paper says, at §4.2, Evidence-Based Revision
Pith says: minor objection 5 of 6
The description states that review instructions 'exclude complete task answers and hidden-test information from reusable assets' as a semantic requirement (b115). The practical enforcement of this requirement is not discussed; if there is a risk of leakage, the transfer assessment (§4.3) could be compromised.
The paper says, at §2.3, Quality vs. Promise
Pith says: minor objection 6 of 6
The distinction between artifact quality and investigative promise is conceptually useful, but the paper does not evaluate how well the leader agent actually distinguishes them in practice. A brief qualitative example or analysis would strengthen the section.
What would settle it
Run each condition in Tables 1 and 2 for five independent replicates with different random seeds, preserving the same resource allocations. For each metric, report the mean and 95% bootstrap confidence interval. If any metric shows overlapping intervals between AReaL-RSI and Codex-Goal (Table 1) or between any pair of configurations (Table 2), the claim of consistent superiority is not supported at the 95% level. If all intervals are non-overlapping in the claimed direction, the concern is resolved.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 14 Oct 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 6 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
Here's my take on the AReaL-RSI paper. The thing to know is that this is a genuinely thoughtful architecture for coordinating multiple research agents in a recursive self-improvement setting. The framing—leader-guided organization with conditional memory sharing and cross-campaign harness evolution—is clean and addresses real problems like premature convergence and loss of applicability conditions. I give them credit for thinking carefully about what experience should contain (sources, conditions, uncertainty) and for introducing a simple routing rule to limit spread of unresolved problems. The formulation in §2 connects scaling to resources and evaluation outcomes, which is honest.
The soft spot is the one you'd expect: the evaluation rests on single runs with no error bars. The paper says so explicitly ($5.1: 'Comparisons are descriptive; repeated-run uncertainty is unavailable.'). That's a problem because the systems are stochastic (LLM calls, leader decisions, candidate evaluations), and some reported margins are tiny—the Astra jailbreak score is 81.261 vs. 81.142. Without variance estimates, we can't tell if the claimed universal superiority (all six metrics, both backbones) reflects real advantage or noise. The authors acknowledge this and also note that matched-compute comparisons, mechanism ablation, and cross-campaign evolution gains remain unmeasured. Those aren't omissions they hide; they're stated as future work. But the central empirical case for the framework's benefits is weaker than the confident prose suggests.
Who is this for? Researchers working on multi-agent coordination and RSI architecture. The paper offers a plausible blueprint with novel components. The evaluation, while incomplete, shows consistent directional trends that are worth investigating. I'd say it deserves a serious referee—the conceptual contribution is real—but the revision should require repeated runs with confidence intervals and a matched-compute or cost-controlled comparison. Without that, the magnitude claims are unsubstantiated.
The strongest objection
The central claim of consistent superiority over Codex-Goal on all six metrics and on all seven KernelBench task-backbone pairs depends on point estimates from what appear to be single trials. The paper explicitly states that 'comparisons are descriptive; repeated-run uncertainty is unavailable' (§5.1), meaning no variance, confidence intervals, or hypothesis tests are reported. This is critical because the system involves stochastic LLM calls, leader decisions, and candidate evaluations; a single run cannot rule out that the observed gains, especially marginal ones (e.g., Astra jailbreak score: 81.261 vs. 81.142), are within the noise envelope. The headline statement 'outperforms Codex-Goal on all six metrics with both backbones' is therefore not falsifiable from the reported data. The same concern applies to the RQ2 latency comparisons: e.g., Astra L3P19 experience sharing latency 0.479 ms vs. sharing+z specialization 0.473 ms (1.3% difference) could easily flip under repetition. Without error bars, the reader cannot assess the reliability of the claimed ranking, making the evaluation unsubstantiated in magnitude and direction.
The reader's weakest assumption focused on the leader agent's judgment quality, which is a plausible internal threat but is not directly tested by the current experiments. I find the more immediate load-bearing gap to be the absence of any measure of variability or significance. This concern is mentioned in the reader's rationale ('lacks repeated-run uncertainty, statistical tests'), so there is partial agreement. However, the reader did not elevate it as the weakest assumption, whereas I argue it is the most foundational: if single-run results are noisy, the entire empirical case collapses. The paper itself acknowledges this limitation (§5.1), but it does not mitigate it. A concrete repeated-run check would directly address whether the observed advantages are reproducible. The verdict should remain CONDITIONAL, as the authors have an accepted path to strengthen the evidence (provide repeated-run statistics). I recommend no change to the reader's proposed condition.
The referee's report
If the reported results are reliable, the framework addresses a practical challenge in scaling recursive self-improvement by proposing explicit mechanisms for coordination, conditional sharing, and reuse of research procedures. The design is clearly motivated and the problem—turning additional compute into useful research progress—is timely. However, the significance is tempered by the limited empirical support, as discussed below.
Pith's reading of the paper (a machine's summary, not the authors')
The paper introduces AReaL-RSI, a framework that organizes a persistent team of researcher agents—guided by a leader agent—to scale recursive self-improvement without updating the underlying language model weights. Three mechanisms carry the argument: a leader agent selects task-appropriate research orientations and evaluates progress by distinguishing current artifact quality from investigative promise; conditional memory sharing limits the spread of unresolved problems while allowing supported findings to reach relevant peers with their applicability conditions; and after each campaign, an evidence-based review revises reusable researcher harnesses (orientations, knowledge, and skills) for future tasks. Under fixed resource budgets, AReaL-RSI outperforms Codex-Goal on all six reported metrics across two LLM backbones, and combining experience sharing with specialized researcher roles yields lower kernel latencies than independent exploration or sharing alone in every comparable task–backbone pair.
What carries the argument. The core mechanism is the three-part coupling of (1) a leader agent that uses evidence-driven coordination (π_coord) to distinguish artifact quality from investigative promise and to assign researcher harnesses, (2) conditional memory sharing that labels experiences as 'problem' or 'positive', limits proactive problem routing to at most one additional recipient, and requires new evidence for relabeling, and (3) an evidence-based review process (π_review) that converts completed campaign evidence into revisions of researcher harnesses (orientation, knowledge, and skills) that are versioned and published for future campaigns.
Core claim. AReaL-RSI establishes that coordinating parallel researcher agents through a leader-driven policy of evidence-based guidance, conditional memory sharing, and cross-campaign harness evolution can consistently improve research outcomes over the baseline Codex-Goal under comparable resource allowances. The central discovery is that the coupling of selective information propagation (which preserves source and condition metadata and restricts problem routing) with differentiated investigative roles (selected by the leader from an evolving harness library) produces better final artifacts than either independent parallel exploration or undifferentiated sharing alone, across all seven task–backbone pairs tested.
The premise it rests on. The leader agent can reliably judge which research directions are truly promising and which artifacts are merely superficially good, without any special training, using only the base LLM's reasoning on the available evidence.
If the paper is right (Pith's inferences, not the authors')
Under the reported resource budgets, AReaL-RSI achieves 2.2 percentage-point higher resolution on Terminal Bench 2.1 and 32.2–36.8% higher mean speedup on KernelBench over Codex-Goal, demonstrating that coordinated parallel investigation can produce better artifacts than the baseline approach.
Combining experience sharing with specialized researcher roles reduces kernel latency by 1.9–27.5% relative to independent exploration and by 1.3–33.9% relative to sharing alone across all seven comparable task–backbone pairs, indicating that both information routing and task-coditioned role assignment matter for scalability.
The cross-campaign harness evolution allows the system to retain reusable research procedures (orientations, knowledge, skills) without retraining the backbone LLM, which could compound improvements over successive tasks.
The conditional memory sharing scheme prevents premature convergence by restricting the broadcastu of unresolved problems while allowing supported findings to propagate, which helps maintain exploration diversity.
The leader agent's reliance on reasoning-based π_coord without additional training is a critical bottleneck: a systematic failure to distinguish promising from unpromising directions would erode all benefits of coordination. A head-to-heard comparison of leader-guided vs. randomly-assigned directions on a suite of tasks would directly test this assumption.
The framework's effectiveness likely depends on the quality of the backbone LLM's reasoning (here GPT-5.6-Luna and GPT-6-Astra). Weaker backbones may not be able to execute the required evidence interpretation and harness revision steps, limiting the transferability of the method to less capable models.
The paper does not measure the benefit of cross-campaign harness evolution separately from within-campaign coordination. An ablation that disables harness publication while retaining within-campaign sharing and specialization would reveal how much of the gain comes from reusable procedures vs. real-time coordination alone.
If the leader agent's guidance and the conditional sharing rules require close to a fixed overhead per campaign (as opposed to scaling gracefully with the number of researcher agents or tasks), the framework may have limited beneit for very small or very large teams and tasks.
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
Scaling recursive self-improvement requires additional computation to produce useful research progress in reaching persistent objectives. Intuitively, more researcher agents can explore more hypotheses, but early evaluation scores might misrepresent their promise, shared findings could lose their applicability conditions, and useful procedures could remain confined to completed investigations. In this paper, we present AReaL-RSI, a framework for organizing a persistent research team of agents without updating backbone LLM weights. We formulate the scaling problem in terms of evaluated outcomes and whole-system resource expenditure. Concretely, a leader agent selects task-appropriate research orientations and guides ongoing investigations; conditional memory sharing distinguishes unresolved problems from supported findings and enables recipient adaptation. During recursive evolution, after each campaign, evidence-based review revises researcher orientation, knowledge, and skills for later campaigns. Experiments compare AReaL-RSI with Codex-Goal on downstream harness optimization, GPU kernel optimization, and AI research tasks using two backbones. Under the reported resource allocations, AReaL-RSI improves Terminal Bench performance by 2.2 percentage points for each backbone and increases mean KernelBench speedup by 32.2-36.8%; further scalability comparisons favor the design of AReaL-RSI over native scaling: combining experience sharing with specialized researcher roles yields 1.9-27.5% lower kernel latency than independent exploration and 1.3-33.9% lower latency than sharing alone across all seven comparable task-backbone pairs.
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 1.00
“Under the reported resource allocations, AReaL-RSI improves Terminal Bench performance by 2.2 percentage points for each backbone and increases mean KernelBench speedup by 32.2-36.8%; further scalability comparisons favor the design of AReaL-RSI over native scaling: combining experience sharing with specialized researcher roles yields 1.9-27.5% lower kernel latency than independent exploration and 1.3-33.9% lower latency than sharing alone across all seven comparable task-backbone pairs.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.38 · human participants 0.03 · controlled comparison 0.96 · 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 |
|---|---|---|---|---|
| Scaling recursive self-improvement requires additional computation to produce useful research progress in reaching persistent objectives. | background 0.98 | 0.11 | 0.05 | asserted 1.00 |
| Intuitively, more researcher agents can explore more hypotheses, but early evaluation scores might misrepresent their promise, shared findings could lose their applicability conditions, and useful procedures could remain confined to completed investigations. | background 0.74 | 0.07 | 0.02 | speculative 0.70 |
| In this paper, we present AReaL-RSI, a framework for organizing a persistent research team of agents without updating backbone LLM weights. | contribution 0.99 | 0.83 | 0.05 | asserted 1.00 |
| We formulate the scaling problem in terms of evaluated outcomes and whole-system resource expenditure. | method 0.75 | 0.79 | 0.03 | asserted 0.95 |
| Concretely, a leader agent selects task-appropriate research orientations and guides ongoing investigations; conditional memory sharing distinguishes unresolved problems from supported findings and enables recipient adaptation. | method 0.99 | 0.74 | 0.05 | asserted 1.00 |
| During recursive evolution, after each campaign, evidence-based review revises researcher orientation, knowledge, and skills for later campaigns. | method 0.99 | 0.66 | 0.05 | asserted 0.99 |
| Experiments compare AReaL-RSI with Codex-Goal on downstream harness optimization, GPU kernel optimization, and AI research tasks using two backbones. | method 0.88 | 0.96 | 0.05 | asserted 0.89 |
| Under the reported resource allocations, AReaL-RSI improves Terminal Bench performance by 2.2 percentage points for each backbone and increases mean KernelBench speedup by 32.2-36.8%; further scalability comparisons favor the design of AReaL-RSI over native scaling: combining experience sharing with specialized researcher roles yields 1.9-27.5% lower kernel latency than independent exploration and 1.3-33.9% lower latency than sharing alone across all seven comparable task-backbone pairs. | result 1.00 | 0.96 | 0.80 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.97
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, 6,396 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 (1)
Figures are the authors', reproduced from the paper.
References (39)
- [1] Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, and Jeff Clune. Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents, 2025. URLhttps://arxiv.org/abs/2505.22954v3. arXiv:2505.22954v3 on PithPith's reviewarXiv
- [2] Jenny Zhang, Bingchen Zhao, Wannan Yang, Jakob Foerster, Jeff Clune, Minqi Jiang, Sam Devlin, and Tatiana Shavrina. Hyperagents, 2026. URLhttps://arxiv.org/abs/2603.19461v1. arXiv:2603.19461v1 on PithPith's reviewarXiv
- [3] Tong Zheng, Xidong Wu, Zheng Zhang, Zhankui He, Chaoyi Zhang, Benjamin Coleman, Ruoqiao Wei, Di Bai, Haolin Liu, Rui Liu, Xue Wang, Yue Zhuan, Wang-Cheng Kang, Renkai Xiang, Heng Huang, Xinwu Cheng, and Yunsong Guo. Dream-RSI: Recursive Self-Improvement through Evolving Worlds... on PithPith's reviewarXiv
- [4] Anne Ouyang, Simon Guo, Simran Arora, Alex L. Zhang, William Hu, Christopher Ré, and Azalia Mirhoseini. KernelBench: Can LLMs Write Efficient GPU Kernels?, 2025. URL https://arxiv.org/abs/2502.10517v1 . arXiv:2502.10517v1 on PithPith's reviewarXiv
- [5] Hjalmar Wijk, Tao Lin, Joel Becker, Sami Jawhar, Neev Parikh, Thomas Broadley , Lawrence Chan, Michael Chen, Josh Clymer, Jai Dhyani, Elena Ericheva, Katharyn Garcia, Brian Goodrich, Nikola Jurkovic, Holden Karnofsky, Megan Kinniment, Aron Lajko, Seraphina Nix, Lucas Sato, Wil... on PithPith's reviewarXiv
- [6] Jun Shern Chan, Neil Chowdhury, Oliver Jaffe, James Aung, Dane Sherburn, Evan Mays, Giulio Starace, Kevin Liu, Leon Maksin, Tejal Patwardhan, Lilian Weng, and Aleksander M ˛ adry. MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering, 2024. URL https://... on PithPith's reviewarXiv
- [7] Summer Eunhyung Ann, Haokun Liu, and Chenhao Tan. The Interaction Tax: When Communication Erases Diversity in Multi-Agent Teams, 2026. URLhttps://arxiv.org/abs/2608.23541v1. arXiv:2608.23541v1 on PithPith's reviewarXiv
- [8] Juraj Gottweis, Wei-Hung Weng, Alexander Daryin, Tao Tu, Anil Palepu, Petar Sirkovic, Artiom Myaskovsky, Felix Weissenberger, Keran Rong, Ryutaro Tanno, Khaled Saab, Dan Popovici, Jacob Blum, Fan Zhang, Katherine Chou, Avinatan Hassidim, Burak Gokturk, Amin Vahdat, Pushmeet Ko... on Pith
- [9] Jiajie Jin, Yuyang Hu, Kai Qiu, Qi Dai, Chong Luo, Guanting Dong, Xiaoxi Li, Tong Zhao, Xiaolong Ma, Gongrui Zhang, Zhirong Wu, Bei Liu, Zhengyuan Yang, Linjie Li, Lijuan Wang, Hongjin Qian, Yutao Zhu, and Zhicheng Dou. Toward Generalist Autonomous Research via Hypothesis-Tree... on PithPith's reviewarXiv
- [10] Chengxin Yu, Zhaoxin Fan, Faguo Wu, Hongwei Zheng, Yun Zhou, and Zhiyu Li. CoMem: Collective-Individual Memory Synergy for Evolutionary Multi-Agent Systems, 2026. URL https://arxiv.org/abs/2609.15009v1 . arXiv:2609.15009v1 on PithPith's reviewarXiv
- [11] Max Jaderberg, Valentin Dalibard, Simon Osindero, Wojciech M. Czarnecki, Jeff Donahue, Ali Razavi, Oriol Vinyals, Tim Green, Iain Dunning, Karen Simonyan, Chrisantha Fernando, and Koray Kavukcuoglu. Population Based Training of Neural Networks, 2017. URLhttps://arxiv.org/abs/1... on PithPith's reviewarXiv
- [12] Haiqian Yang and Yuan Cao. Efficiency Matters in Autonomous Research, 2026. URL https://arxiv.org/abs/26 07.24647v1. arXiv:2607.24647v1 on PithPith's reviewarXiv
- [13] Alexander Novikov, Ngân V˜u, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wagner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushm... on PithPith's reviewarXiv
- [14] OpenAI. Using goals in codex. https://developers.openai.com/cookbook/examples/codex/using_goals_in _codex, 2026. Accessed Sep 25, 2026 on Pith
- [15] Shengran Hu, Cong Lu, and Jeff Clune. Automated Design of Agentic Systems, 2024. URL https://arxiv.org/ab s/2408.08435v2. arXiv:2408.08435v2 on PithPith's reviewarXiv
- [16] Lakshya A Agrawal, Shangyin Tan, Dilara Soylu, Noah Ziems, Rishi Khare, Krista Opsahl-Ong, Arnav Singhvi, Herumb Shandilya, Michael J Ryan, Meng Jiang, Christopher Potts, Koushik Sen, Alexandros G. Dimakis, Ion Stoica, Dan Klein, Matei Zaharia, and Omar Khattab. GEPA: Reflecti... on Pith
- [17] Quan Shi, Keshav Dhandhania, Karthik Narasimhan, and Victor Barres. τ τ-Bench: An Environment for End-To-End, Realistic Agent Construction, 2026. URLhttps://arxiv.org/abs/2609.04611v1. arXiv:2609.04611v1 on PithPith's reviewarXiv
- [18] URLhttps://arxiv.org/abs/2507.19457v2. arXiv:2507.19457v2 on PithPith's reviewarXiv
- [19] Yutaro Yamada, Robert Tjarko Lange, Cong Lu, Shengran Hu, Chris Lu, Jakob Foerster, Jeff Clune, and David Ha. The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search, 2025. URL https://arxiv.org/abs/2504.08066v1. arXiv:2504.08066v1 on PithPith's reviewarXiv
- [20] Zhengyao Jiang, Dominik Schmidt, Dhruv Srikanth, Dixing Xu, Ian Kaplan, Deniss Jacenko, and Yuxiang Wu. AIDE: AI- Driven Exploration in the Space of Code, 2025. URL https://arxiv.org/abs/2502.13138v1. arXiv:2502.13138v1 on PithPith's reviewarXiv
show 19 more
- [21] Haoyang Fang and Bernie Wang. Exploit More, Explore Smarter for Budget-Constrained Agentic Search, 2026. URL https://arxiv.org/abs/2608.23848v1. arXiv:2608.23848v1 on PithPith's reviewarXiv
- [22] Mohan Chen. Loreley: Repository-Scale Program Evolution with Quality-Diversity Search, 2026. URL https: //arxiv.org/abs/2608.19703v1. arXiv:2608.19703v1 on PithPith's reviewarXiv
- [23] Anas Mahmoud, MohammadHossein Rezaei, Mingchen Zhuge, Chenguang Wang, Yuandong Tian, and Yunzhong He. Announcing RSI Bench, August 2026. URL https://labs.scale.com/rsi-benchmark/blog/announcing-rsi-b ench. Official announcement, August 7, 2026 on Pith
- [24] Thomas Simon Foster, Bassel Al Omari, Tingchen Fu, Thomas Mann, Carl Domond, Lucia Cipolina-Kun, Bhavul Gauri, Muna Aghamelu, Alexander D. Goldie, Eryk Helenowski, Jean-Christophe Gagnon-Audet, Alberto Pepe, Saba Nazir, Daniel Izcovich, Noam Levi, Rishi Hazra, Karen Hambardzum... on PithPith's reviewarXiv
- [25] Amy Xin, Jiening Siow, Junjie Wang, Zijun Yao, Fanjin Zhang, Jian Song, Lei Hou, and Juanzi Li. EurekAgent: Agent Environment Engineering is All You Need For Autonomous Scientific Discovery, 2026. URL https://arxiv.org/ab s/2606.13662v2. arXiv:2606.13662v2 on PithPith's reviewarXiv
- [26] Marjan Moodi, Xuankang Zhu, Fernando De Mesentier Silva, Harold Chaput, and Mohammad Reza Tae- siri. AutoWorldModel-Bench: A State-Centric Benchmark for Automated World-Model Research, 2026. URL https://arxiv.org/abs/2608.11216v2. arXiv:2608.11216v2 on PithPith's reviewarXiv
- [27] Faramarz Jabbarvaziri. Remember, Don’t Re-read: Stateful ReAct Agents for Token-Efficient Autonomous Experimen- tation, 2026. URLhttps://arxiv.org/abs/2606.14945v1. arXiv:2606.14945v1 on PitharXiv
- [28] Yanwen Peng, Delvin Ce Zhang, Xi Wang, and Nikolaos Aletras. BusMA: A Bus Communication Substrate for Multi-Agent Systems, 2026. URLhttps://arxiv.org/abs/2609.15054v1. arXiv:2609.15054v1 on PithPith's reviewarXiv
- [29] Eric Hanchen Jiang, Zhi Zhang, Yuchen Wu, Levina Li, Dong Liu, Xiao Liang, Rui Sun, Yubei Li, Edward Sun, Haozheng Luo, Zhaolu Kang, Aylin Caliskan, Kai-Wei Chang, and Ying Nian Wu. Memory as a Controlled Process: 11 Learned Adaptive Memory Management for LLM Agents, 2026. URL... on PithPith's reviewarXiv
- [30] Guibin Zhang, Muxin Fu, Guancheng Wan, Miao Yu, Kun Wang, and Shuicheng Yan. G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems, 2025. URLhttps://arxiv.org/abs/2506.07398v1. arXiv:2506.07398v1 on PithPith's reviewarXiv
- [31] Jianlyu Chen, Yuyang Hu, Hongjin Qian, Jiawei Liu, Wenqing Wei, Xiaolong Chen, Defu Lian, Zhicheng Dou, Chaozhuo Li, Qiwei Ye, and Zheng Liu. Repo-To-Skill: Distilling GitHub Repositories Into AI4AI Skills, 2026. URL https://arxiv.org/abs/2609.02749v1. arXiv:2609.02749v1 on PithPith's reviewarXiv
- [32] Siyu Liu, Bo Hu, Beilin Ye, He Cao, David J. Srolovitz, and Tongqi Wen. Harnessing agent memory to build lifelong AI partners for materials scientists, 2026. URLhttps://arxiv.org/abs/2608.11224v1. arXiv:2608.11224v1 on PithPith's reviewarXiv
- [33] Yiwei Li, Wanli Yang, Hexiang Tan, Xiangzhou Huang, Zhengyu Chen, Ziran Li, Borun Chen, Shanglin Lei, Huaisheng Zhu, Hao Tian, Fei Sun, Xunliang Cai, and Jingang Wang. Beyond Final Scores: A Systematic Evaluation of Agents for Long-Horizon AI Research and Development, 2026. UR... on PithPith's reviewarXiv
- [34] Deyao Zhu, Xin Zhou, Shengling Qin, Xuekai Zhu, Hangliang Ding, Shu Zhong, Zixin Wen, Zhonglin Xie, Chenhui Gou, Linxuan Ren, Yueyang Wang, Junfeng Zhong, Rui Liu, Tian Gao, Yangguang Lin, Jingyuan Zhang, Maojia Song, Xuan Qi, Jinhong Wu, Chenyang Zhang, Yinzhu Piao, Ziru Niu,... on Pith
- [35] Yubin Kim, Ken Gu, Chanwoo Park, Chunjong Park, Samuel Schmidgall, A. Ali Heydari, Yao Yan, Zhihan Zhang, Yuchen Zhuang, Yun Liu, Mark Malhotra, Paul Pu Liang, Hae Won Park, Yuzhe Yang, Xuhai Xu, Yilun Du, Shwetak Patel, Tim Althoff, Daniel McDuff, and Xin Liu. Towards a Scien... on PithPith's reviewarXiv
- [36] Davide Paglieri, Logan Cross, Tim Genewein, Joel Z. Leibo, Nenad Tomasev, and Alexander Sasha Vezhnevets. A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms, 2026. URL https: //arxiv.org/abs/2609.04170v1. arXiv:2609.04170v1. 12 on PithPith's reviewarXiv
- [37] Junyou Li, Qin Zhang, Yangbin Yu, Qiang Fu, and Deheng Ye. More Agents Is All You Need, 2024. URL https: //arxiv.org/abs/2402.05120v2. arXiv:2402.05120v2 on PithPith's reviewarXiv
- [2025] URLhttps://arxiv.org/abs/2502.18864v1. arXiv:2502.18864v1 on PithPith's reviewarXiv
- [2026] URLhttps://arxiv.org/abs/2607.05155v1. arXiv:2607.05155v1 on PithPith's reviewarXiv
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