Pith. sign in

REVIEW 2 major objections 4 minor 52 references

This paper proposes that every design choice in a multi-agent automated research system—who the agents are, what tools they hold, how they communicate, what state carries across runs, how actions are chosen, how a run starts, and how output

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

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

T0 review · deepseek-v4-flash

2026-08-02 07:01 UTC pith:XLI33PCT

load-bearing objection Useful framework paper with one real definitional bug: role/persona lives in both A and ι, so the claimed one-coordinate ablations are undefined for that axis; fix it and this is a solid contribution. the 2 major comments →

arxiv 2607.22682 v1 pith:XLI33PCT submitted 2026-07-13 cs.AI cs.LGcs.MA

A Vocabulary for Multi-Agent Automated Research Systems

Bardiya Akhbari This is my paper
classification cs.AI cs.LGcs.MA
keywords multi-agent systemsautomated researchLLM agentssystem design vocabularytrajectory distributionproxy-quality gapreward hackingevaluator integrity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that the usual way of talking about 'multi-agent' or 'autoresearch' systems is too coarse: a claim that one system is better than another is meaningless until we know which structural axis changed. To fix that, it proposes an eight-coordinate tuple—agents, operations, communication, capability assignment, shared state, control policy, initialization, and evaluator—and treats a run as a trajectory drawn from a distribution over behaviors. The payoff is that structural questions (when agents should talk, gain tools, or share memory across runs) become choices that can be varied one at a time, and that reported gains can be attributed to a coordinate instead of to the vague label 'multi-agent.' The paper also splits the complaint that these systems lack taste into two separable failures: generative taste (how often novel trajectories are proposed before scoring) and evaluative taste (how far the proxy score drifts from true quality). If the vocabulary is right, comparing and auditing automated research systems becomes an ablations-first activity rather than a leaderboard comparison.

Core claim

The central claim is that every meaningful structural design decision in a multi-agent automated research system can be captured by the tuple M = ⟨A, O, C, α, S, π, ι, e⟩, where A is the set of agents, O is the universe of callable operations, C = (Λ, σ) is the communication structure (who may message whom, and in what form), α assigns capabilities to agents, S splits shared state into within-run, world, and cross-run components, π decomposes into routing, stopping, meta-control, and exploration policies, ι sets how a run starts, and e is the evaluator the system optimizes. A trajectory records one run, and because agents and initialization are stochastic, a system induces a distribution ove

What carries the argument

The central object is the eight-tuple M = ⟨A,O,C,α,S,π,ι,e⟩, with C=(Λ,σ), S=(S_btw, S_world, S_cross), and π=(π_route, π_stop, π_meta, η). The corresponding trajectory τ records, per step, who acted, what operation they invoked, what they observed, the shared and private state, and the structural triple of agents, capabilities, and communication edges. These definitions do the work of decomposition: the tuple gives each design question a coordinate, the trajectory gives a distribution to measure, the disjoint action space (operation calls, messages, meta-control actions, halt) turns 'when should agents communicate?' into a policy question over edges, and the proxy-quality gap Δω = |ω(e(τ))

Load-bearing premise

The load-bearing premise is that every meaningful design choice can be assigned to exactly one of the eight tuple coordinates and that an ablation can vary one coordinate while holding the others fixed; if coordinates interact in practice—say, changing the communication topology forces a change in routing policy—then gains can no longer be attributed to a single axis.

What would settle it

Find a system whose behavior differs from another but that maps to identical tuples, or demonstrate a design change that cannot be expressed by varying one coordinate without forcing another to change (e.g., a communication topology that makes a previously valid routing policy impossible). Either observation would disprove the claimed isolation and sufficiency of the vocabulary.

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

If this is right

  • A claim that one autoresearch system beats another is incomplete until it names which coordinate changed; the tuple supplies the coordinates for that attribution.
  • Designers can run one-coordinate ablations—vary communication topology Λ, capability assignment α, initialization ι, or evaluator integrity e—while holding the other seven fixed.
  • Reward hacking and other evaluation failures become properties of the evaluator e, so the fix (variance control, sandbox isolation, blackout, contamination checks) targets the coordinate that failed.
  • The 'no taste' complaint splits into two fixable problems: generative taste is addressed by changing seeding, exploration, or backbones, while evaluative taste is addressed by strengthening evaluator integrity or mechanism.
  • The framework covers recursive self-improvement: meta-control actions can rewrite the policy, mutate capabilities, and even modify the operation universe or backbone mid-run.

Where Pith is reading between the lines

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

  • Going beyond the paper: the claimed independence of the eight coordinates is itself testable; an empirical research program could measure pairwise interactions, e.g., whether changing communication topology Λ constrains which routing policies π_route remain meaningful.
  • The proxy-quality gap could be measured on existing systems by scoring returned artifacts on a held-out true-quality metric and plotting how the gap grows with search effort N; if the overfitting tax is absent, the paper's central diagnostic would need revision.
  • The vocabulary suggests a reporting standard for the field: publish the tuple and trajectory distributions alongside final scores, which would let meta-analyses attribute gains across systems without re-running them.
  • Underexplored coordinates identified by the paper (dynamic capability assignment, run-time rewriting of the policy, changing the task distribution D) are natural candidates for the next generation of systems; each becomes a single-coordinate experiment under this framework.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes a vocabulary for describing multi-agent automated research systems as an eight-tuple M=⟨A,O,C,α,S,π,ι,e⟩. It separates problem specification P from system M, defines trajectories, formulates search as maximizing a proxy evaluator e under a budget, splits “taste” into generative and evaluative components, and applies the tuple to ten recent research systems. Its central claim is that the tuple decomposes design choices into coordinates fine enough to isolate each choice, so that ablations can vary one coordinate while holding others fixed, and that evaluator integrity is distinct from search quality. The paper is a formalization/position paper rather than an empirical study.

Significance. If the vocabulary were coherent, it would be a useful contribution: it makes explicit a set of axes — communication, capability, initialization, evaluator — that current survey labels blur, and it connects reward hacking to a component (e) rather than to agent intent, with a clear proxy-quality gap definition. The case-study coverage is broad (AIRA2, AlphaEvolve, Glia, AAR, AI Scientist-v2, MetaGPT, EvoX, ml-intern, SimpleTES, Engram), and the paper is transparent that its set/generic coding is descriptive, not a validated measurement. The formal definitions (Eqs. 1–9) are mostly clear and locally consistent, and the overfitting-tax discussion grounds itself in external reward-hacking literature. However, the main separability claim is currently undermined by an internal double-counting of role/persona and by unaddressed coordinate interactions, so the vocabulary does not yet deliver its advertised one-coordinate ablation semantics.

major comments (2)
  1. [§3, Eq. (4); §6 MetaGPT; §7] Role/persona is double-counted. Section 3 defines each agent as a_i=(θ_i, m_i, m0_i, ρ_i) with “role ρ_i specifies the agent’s prompt, persona, or capability tier,” while the initialization refinement (Eq. (4)) makes P0_{1:n}(x) “the role or system prompt.” The MetaGPT case study states “The personas are the initialization ι… so A sets how many distinct slots there are and ι sets what fills each,” and §7 says “the spread of roles ρ_i set by ι.” Thus a change in role/persona changes both A and ι, and the one-coordinate ablation promised in §1 and §7 is undefined on the axis the paper itself uses. Fix by assigning role/persona to exactly one coordinate and updating the case studies.
  2. [§1, §7; §3 Communication/Control] The claim that ablations can “hold all coordinates fixed except one” is not supported by the formal definitions, because coordinates are not independent. §3 requires π_route to return actions in α_t(a_t) and messages to use Λ-permitted edges; Table 2 lets π_meta mutate A_t, α_t, and Λ_t. So changing Λ may invalidate the current π_route, changing α may invalidate the action set, and changing A changes the domain of α and Λ. The paper should specify which derived consistency constraints are recomputed in a one-coordinate ablation, or state that separability is an idealization. Absent that, attribution of a performance gain to a single coordinate is confounded.
minor comments (4)
  1. [Abstract] “lacktasteinto” is missing a space; should read “lack taste into.”
  2. [§6, Figure 4] The set/generic coding is described as following each system’s own description, but the judgment is still subjective. Adding a short coding rubric or a second-coder agreement would improve reproducibility.
  3. [§5.2] The overfitting-tax claim — that ranking more trajectories increases the chance the top-scoring one overstates q — is stated informally. It is plausible and cites related literature, but since it motivates the “improve e before crediting the generator” recommendation, a formal or at least quantitative statement would strengthen it.
  4. [Throughout] Spelling of AIRA2 is inconsistent (“AIRA2”, “AIRA 2”). Minor typographic issue only.

Circularity Check

0 steps flagged

No significant circularity: the paper is a definitional vocabulary with no fitted predictions or load-bearing self-citations.

full rationale

The paper's central contribution is a formal vocabulary, M=⟨A,O,C,α,S,π,ι,e⟩, and a descriptive mapping of independently published systems; it does not derive an empirical prediction, fit any parameter, or invoke a self-authored uniqueness/ansatz result. The overfitting-tax discussion explicitly rests on external reward-hacking literature (MLR-Bench, METR, Skalse et al., Gao et al.), so that support is independent. The case-study codings are presented as 'a descriptive map, not a validated measurement,' so no coordinate-effectiveness claim reduces to the vocabulary's own definitions. The one definitional overlap—role ρ_i in the agent identity A and P0 in initialization ι—could complicate the paper's one-coordinate-ablation language, but it is an internal-consistency concern, not a circular derivation: no output in the paper is equivalent to an input by construction, and the vocabulary is the proposed object rather than a derived result. Hence the honest finding is no significant circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 3 invented entities

The paper's central claim rests on domain assumptions about the adequacy and separability of the tuple, the usefulness of trajectory distributions, and the accuracy of the case-study coding. No free parameters are fit. The invented entities (generative taste, evaluative taste, overfitting tax) are conceptual tools proposed without independent empirical validation in this paper.

axioms (4)
  • domain assumption The trajectory distribution PM(·|x) is the correct object of study for LLM-agent systems.
    Assumed in Section 4 and borrowed from Dec-POMDP framing; justifies treating systems as stochastic generators rather than deterministic programs.
  • domain assumption The eight-tuple M = ⟨A,O,C,α,S,π,ι,e⟩ is a sufficient decomposition of any multi-agent autoresearch system.
    Underpins Section 1 and the Section 6 case-study mapping; if a design choice maps to multiple coordinates or to none, the attribution goal fails.
  • domain assumption True quality q is generally uncomputable, so a proxy-quality gap necessarily exists.
    Stated in Section 2; motivates the evaluator coordinate and the overfitting-tax argument, but not proven.
  • domain assumption Case-study descriptions in Section 6 faithfully represent the primary systems.
    The paper explicitly says it follows each system's own description rather than re-evaluating choices (Section 6), so the coding is a descriptive claim, not a verified measurement.
invented entities (3)
  • generative taste no independent evidence
    purpose: To name the rate at which a system proposes novel trajectories before any score is observed.
    Defined in Section 7 as a property of the trajectory distribution; deliberately left qualitative with no novelty measure, so no independent falsifiable handle yet.
  • evaluative taste no independent evidence
    purpose: To name the gap between proxy score and true quality.
    Defined as Δω in Section 5.2; measurable in principle but not measured in this paper.
  • overfitting tax no independent evidence
    purpose: To describe the penalty of optimizing a miscalibrated proxy: harder search finds trajectories where the proxy overstates true quality.
    Argued qualitatively in Section 5.2; supported by cited reward-hacking evidence (MLR-Bench, METR) but not formally derived or directly measured here.

pith-pipeline@v1.3.0-alltime-deepseek · 20981 in / 7516 out tokens · 68444 ms · 2026-08-02T07:01:48.371317+00:00 · methodology

0 comments
read the original abstract

We introduce a vocabulary for automated research systems built from one or more agents to make their design choices easier to describe and compare. The vocabulary specifies 1) who the agents are, 2) what operations are available in the system, 3) who may invoke them, 4) how agents communicate, 5) what information is visible within and across runs, 6) how the next action is chosen, 7) how a run begins, and 8) how outputs are evaluated. A trajectory records one run from the input task to the returned artifact. Because agents, operations, and initialization may be stochastic, repeated runs on the same task induce a distribution over trajectories rather than a single behavior. Our vocabulary turns structural design questions, such as when agents should communicate, gain or lose a capability, or carry information across runs, into testable choices. It also makes the evaluator a component of the system, since reported gains depend on how closely the proxy score matches true quality. That separation also splits the vague complaint that these systems lack taste into two failures with different solutions. Generative taste is the rate at which a system proposes novel trajectories before any score is observed, and evaluative taste is the gap between the proxy score and the quality it should match. We instantiate the vocabulary on recent autoresearch systems to illustrate that it covers designs that differ widely in structure.

Figures

Figures reproduced from arXiv: 2607.22682 by Bardiya Akhbari.

Figure 1
Figure 1. Figure 1: A multi-agent system is a set of components. A research task enters on the left. The [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The communication space Λ has the shape axis which fixes which edges exist (top row), [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Each panel shows the structural shape of each system. AIRA [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Each system sets only a few coordinates and leaves the rest generic. A filled cell marks [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 23 linked inside Pith

  1. [1]

    AIRA 2: Overcoming bottlenecks in AI research agents, 2026

    Karen Hambardzumyan, Nicolas Baldwin, Edan Toledo, Rishi Hazra, Michael Kuchnik, Bassel Al Omari, Thomas Simon Foster, Anton Protopopov, Jean-Christophe Gagnon-Audet, Ishita Mediratta, Kelvin Niu, Michael Shvartsman, Alisia Lupidi, Alexis Audran-Reiss, Parth Pathak, Tatiana Shavrina, Despoina Magka, Hela Momand, Derek Dunfield, Nicola Cancedda, Pontus Ste...

    Pith/arXiv arXiv 2026

  2. [2]

    Glia: A human-inspired AI for auto- mated systems design and optimization, 2025

    Pouya Hamadanian, Pantea Karimi, Arash Nasr-Esfahany, Kimia Noorbakhsh, Joseph Chandler, Ali ParandehGheibi, Mohammad Alizadeh, and Hari Balakrishnan. Glia: A human-inspired AI for auto- mated systems design and optimization, 2025. arXiv:2510.27176 [cs.AI]

    Pith/arXiv arXiv 2025

  3. [3]

    Automated weak-to-strong researcher

    Jiaxin Wen, Liang Qiu, Joe Benton, Jan Hendrik Kirchner, and Jan Leike. Automated weak-to-strong researcher. Alignment Science Blog, 2026. Accessed June 2026

    2026

  4. [4]

    Harness engineering for self-improvement.lilianweng.github.io, July 2026

    Lilian Weng. Harness engineering for self-improvement.lilianweng.github.io, July 2026. Accessed July 2026

    2026

  5. [5]

    Airs-bench: a suite of tasks for frontier ai research science agents, 2026

    Alisia Lupidi, Bhavul Gauri, Thomas Simon Foster, Bassel Al Omari, Despoina Magka, Alberto Pepe, Alexis Audran-Reiss, Muna Aghamelu, Nicolas Baldwin, Lucia Cipolina-Kun, Jean-Christophe Gagnon- Audet, Chee Hau Leow, Sandra Lefdal, Hossam Mossalam, Abhinav Moudgil, Saba Nazir, Emanuel 17 Tewolde, Isabel Urrego, Jordi Armengol Estape, Amar Budhiraja, Gaurav...

    arXiv 2026

  6. [6]

    When AI builds itself: Our progress toward recursive self-improvement, and its implications

    Marina Favaro and Jack Clark. When AI builds itself: Our progress toward recursive self-improvement, and its implications. The Anthropic Institute, 2026. Accessed June 2026

    2026

  7. [7]

    First steps toward automated AI research

    Recursive. First steps toward automated AI research. Recursive (Recursive Superintelligence, Inc.) Blog, jun 2026. Accessed June 2026

    2026

  8. [8]

    Darwin G¨ odel Machine: Open- ended evolution of self-improving agents, 2025

    Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, and Jeff Clune. Darwin G¨ odel Machine: Open- ended evolution of self-improving agents, 2025. arXiv:2505.22954 [cs.AI]

    Pith/arXiv arXiv 2025

  9. [9]

    Self-taught optimizer (STOP): Recursively self-improving code generation, 2024

    Eric Zelikman, Eliana Lorch, Lester Mackey, and Adam Tauman Kalai. Self-taught optimizer (STOP): Recursively self-improving code generation, 2024. arXiv:2310.02304 [cs.CL]

    Pith/arXiv arXiv 2024

  10. [10]

    MLR-Bench: Evaluating AI agents on open-ended machine learning research

    Hui Chen, Miao Xiong, Yujie Lu, Wei Han, Ailin Deng, Yufei He, Jiaying Wu, Yibo Li, Yue Liu, and Bryan Hooi. MLR-Bench: Evaluating AI agents on open-ended machine learning research. InAdvances in Neural Information Processing Systems 38 (NeurIPS 2025), Datasets and Benchmarks Track, 2025

    2025

  11. [11]

    MALT: A dataset of natural and prompted behaviors that threaten evaluation integrity

    Neev Parikh and Hjalmar Wijk. MALT: A dataset of natural and prompted behaviors that threaten evaluation integrity. METR, 2025

    2025

  12. [12]

    Alexander Novikov, Ngˆ an V˜ u, Marvin Eisenberger, Emilien Dupont, Po-Sen Huang, Adam Zsolt Wag- ner, Sergey Shirobokov, Borislav Kozlovskii, Francisco J. R. Ruiz, Abbas Mehrabian, M. Pawan Kumar, Abigail See, Swarat Chaudhuri, George Holland, Alex Davies, Sebastian Nowozin, Pushmeet Kohli, and Matej Balog. AlphaEvolve: A coding agent for scientific and ...

    Pith/arXiv arXiv 2025

  13. [13]

    The AI scientist-v2: Workshop-level automated scientific discovery via agentic tree search, 2025

    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

    2025

  14. [14]

    MetaGPT: Meta programming for a multi-agent collaborative framework

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Ceyao Zhang, Jinlin Wang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and Juergen Schmidhuber. MetaGPT: Meta programming for a multi-agent collaborative framework. InInternational Conference on Learning Representations, 2024

    2024

  15. [15]

    Pan, Alexander Du, Kurt Keutzer, Alvin Cheung, Alexandros G

    Shu Liu, Shubham Agarwal, Monishwaran Maheswaran, Mert Cemri, Zhifei Li, Qiuyang Mang, Ashwin Naren, Ethan Boneh, Audrey Cheng, Melissa Z. Pan, Alexander Du, Kurt Keutzer, Alvin Cheung, Alexandros G. Dimakis, Koushik Sen, Matei Zaharia, and Ion Stoica. EvoX: Meta-evolution for auto- mated discovery, 2026. arXiv:2602.23413 [cs.LG]

    arXiv 2026

  16. [16]

    ml-intern: A single-agent ReAct loop for ML engineering on the Hugging Face ecosystem

    Hugging Face. ml-intern: A single-agent ReAct loop for ML engineering on the Hugging Face ecosystem. GitHub repository,https://github.com/huggingface/ml-intern, 2025. Accessed June 2026

    2025

  17. [17]

    Evaluation-driven scaling for scientific discovery, 2026

    Haotian Ye, Haowei Lin, Jingyi Tang, Yizhen Luo, Caiyin Yang, Chang Su, Rahul Thapa, Rui Yang, Ruihua Liu, Zeyu Li, Chong Gao, Dachao Ding, Guangrong He, Miaolei Zhang, Lina Sun, Wenyang Wang, Yuchen Zhong, Zhuohao Shen, Di He, Jianzhu Ma, Stefano Ermon, Tongyang Li, Xiaowen Chu, James Zou, and Yuzhi Xu. Evaluation-driven scaling for scientific discovery,...

    Pith/arXiv arXiv 2026

  18. [18]

    Improving coherence and persistence in agentic AI for system optimization, 2026

    Pantea Karimi, Kimia Noorbakhsh, Mohammad Alizadeh, and Hari Balakrishnan. Improving coherence and persistence in agentic AI for system optimization, 2026. arXiv:2603.21321 [cs.AI]

    arXiv 2026

  19. [19]

    Pawan Kumar, Emilien Dupont, Francisco J

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M. Pawan Kumar, Emilien Dupont, Francisco J. R. Ruiz, Jordan S. Ellenberg, Pengming Wang, Omar Fawzi, Pushmeet Kohli, and Alhussein Fawzi. Mathematical discoveries from program search with large lan- guage models.Nature, 625(7995):468–475, 2024. 18

    2024

  20. [20]

    An empirical study of multi-agent collaboration for automated research, 2026

    Yang Shen, Zhenyi Yi, Ziyi Zhao, Lijun Sun, Dongyang Li, Chin-Teng Lin, and Yuhui Shi. An empirical study of multi-agent collaboration for automated research, 2026. arXiv:2603.29632 [cs.MA]

    Pith/arXiv arXiv 2026

  21. [21]

    Exploring design of multi-agent LLM dialogues for research ideation

    Keisuke Ueda, Wataru Hirota, Takuto Asakura, Takahiro Omi, Kosuke Takahashi, Kosuke Arima, and Tatsuya Ishigaki. Exploring design of multi-agent LLM dialogues for research ideation. InProceedings of the 26th Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL), 2025

    2025

  22. [22]

    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 science of scaling agent systems, 2026. arXiv:2512.08296 [cs.AI]

    Pith/arXiv arXiv 2026

  23. [23]

    Automated design of agentic systems, 2025

    Shengran Hu, Cong Lu, and Jeff Clune. Automated design of agentic systems, 2025. arXiv:2408.08435 [cs.AI]

    Pith/arXiv arXiv 2025

  24. [24]

    AFlow: Automating agentic workflow generation, 2025

    Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Jiaqi Chen, Mingchen Zhuge, Sirui Hong, Jinlin Wang, Bang Liu, Yuyu Luo, and Chenglin Wu. AFlow: Automating agentic workflow generation, 2025. arXiv:2410.10762 [cs.AI]

    Pith/arXiv arXiv 2025

  25. [25]

    Learning to replicate expert judgment in financial tasks.Thinking Machines Lab: News, 2026

    Sarah Su, Kevin Zhu, Emily Xiao, Rohan Alur, and Daniel Kang. Learning to replicate expert judgment in financial tasks.Thinking Machines Lab: News, 2026. https://thinkingmachines.ai/news/learning-to- replicate-expert-judgment-in-financial-tasks/

    2026

  26. [26]

    Rui Wang, Joel Lehman, Jeff Clune, and Kenneth O. Stanley. Paired open-ended trailblazer (POET): Endlessly generating increasingly complex and diverse learning environments and their solutions.arXiv preprint arXiv:1901.01753, 2019

    Pith/arXiv arXiv 1901

  27. [27]

    CORAL: Towards autonomous multi-agent evolution for open-ended discovery, 2026

    Ao Qu, Han Zheng, Zijian Zhou, Yihao Yan, Yihong Tang, Shao Yong Ong, Fenglu Hong, Kaichen Zhou, Chonghe Jiang, Minwei Kong, Jiacheng Zhu, Xuan Jiang, Sirui Li, Cathy Wu, Bryan Kian Hsiang Low, Jinhua Zhao, and Paul Pu Liang. CORAL: Towards autonomous multi-agent evolution for open-ended discovery, 2026. arXiv:2604.01658 [cs.AI]

    Pith/arXiv arXiv 2026

  28. [28]

    Autodata: An agentic data scientist to create high quality synthetic data, 2026

    Ilia Kulikov, Chenxi Whitehouse, Tianhao Wu, Yixin Nie, Swarnadeep Saha, Eryk Helenowski, Weizhe Yuan, Olga Golovneva, Jack Lanchantin, Yoram Bachrach, Jakob Foerster, Xian Li, Han Fang, Sainba- yar Sukhbaatar, and Jason Weston. Autodata: An agentic data scientist to create high quality synthetic data, 2026. arXiv:2606.25996 [cs.AI]

    Pith/arXiv arXiv 2026

  29. [29]

    SIA: Self improving AI with harness & weight updates, 2026

    Prannay Hebbar, Yogendra Manawat, Samuel Verboomen, Alesia Ivanova, Selvam Palanimalai, Ku- nal Bhatia, and Vignesh Baskaran. SIA: Self improving AI with harness & weight updates, 2026. arXiv:2605.27276 [cs.AI]

    Pith/arXiv arXiv 2026

  30. [30]

    Rao and Michael P

    Anand S. Rao and Michael P. Georgeff. BDI agents: From theory to practice. InProceedings of the First International Conference on Multi-Agent Systems, pages 312–319, 1995

    1995

  31. [31]

    Wooldridge.An introduction to multiagent systems

    Michael J. Wooldridge.An introduction to multiagent systems. Wiley, 2. ed., repr edition, 2012

    2012

  32. [32]

    Cambridge University Press, 2008

    Yoav Shoham and Kevin Leyton-Brown.Multiagent Systems: Algorithmic, Game-Theoretic, and Logical Foundations. Cambridge University Press, 2008

    2008

  33. [33]

    Bernstein, Shlomo Zilberstein, and Neil Immerman

    Daniel S. Bernstein, Shlomo Zilberstein, and Neil Immerman. The Complexity of Decentralized Control of Markov Decision Processes, January 2013. arXiv:1301.3836 [cs.AI]

    Pith/arXiv arXiv 2013

  34. [34]

    Michael L. Littman. Markov games as a framework for multi-agent reinforcement learning. InMachine Learning Proceedings 1994, pages 157–163. Elsevier, 1994

    1994

  35. [35]

    White, Doug Burger, and Chi Wang

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W. White, Doug Burger, and Chi Wang. AutoGen: Enabling next-gen LLM applications via multi-agent conversation, 2023. arXiv:2308.08155 [cs.AI]

    Pith/arXiv arXiv 2023

  36. [36]

    CAMEL: Communicative agents for “mind” exploration of large language model society, 2023

    Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. CAMEL: Communicative agents for “mind” exploration of large language model society, 2023. arXiv:2303.17760 [cs.AI]. 19

    Pith/arXiv arXiv 2023

  37. [37]

    ChatDev: Communica- tive agents for software development, 2023

    Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, Juyuan Xu, Dahai Li, Zhiyuan Liu, and Maosong Sun. ChatDev: Communica- tive agents for software development, 2023. arXiv:2307.07924 [cs.SE]

    Pith/arXiv arXiv 2023

  38. [38]

    AgentVerse: Facilitating multi-agent collaboration and exploring emergent behaviors, 2023

    Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chi-Min Chan, Heyang Yu, Yaxi Lu, Yi-Hsin Hung, Chen Qian, Yujia Qin, Xin Cong, Ruobing Xie, Zhiyuan Liu, Maosong Sun, and Jie Zhou. AgentVerse: Facilitating multi-agent collaboration and exploring emergent behaviors, 2023. arXiv:2308.10848 [cs.CL]

    Pith/arXiv arXiv 2023

  39. [39]

    Voyager: An open-ended embodied agent with large language models, 2023

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models, 2023. arXiv:2305.16291 [cs.AI]

    Pith/arXiv arXiv 2023

  40. [40]

    Agentic context engineering: Evolving contexts for self-improving language models

    Qizheng Zhang, Changran Hu, Shubhangi Upasani, Boyuan Ma, Fenglu Hong, Vamsidhar Kamanuru, Jay Rainton, Chen Wu, Mengmeng Ji, Hanchen Li, Urmish Thakker, James Zou, and Kunle Olukotun. Agentic context engineering: Evolving contexts for self-improving language models. InInternational Conference on Learning Representations, 2026

    2026

  41. [41]

    Meta context engineering via agentic skill evolution, 2026

    Haoran Ye, Xuning He, Vincent Arak, Haonan Dong, and Guojie Song. Meta context engineering via agentic skill evolution, 2026. arXiv:2601.21557 [cs.AI]

    arXiv 2026

  42. [42]

    A survey of self-evolving agents: What, when, how, and where to evolve on the path to artificial super intelligence.Transactions on Machine Learning Research, 2026

    Huan-ang Gao, Jiayi Geng, Wenyue Hua, Mengkang Hu, Xinzhe Juan, Hongzhang Liu, Shilong Liu, Jiahao Qiu, Xuan Qi, Yiran Wu, Hongru Wang, Han Xiao, Yuhang Zhou, Shaokun Zhang, Jiayi Zhang, Jinyu Xiang, Yixiong Fang, Qiwen Zhao, Dongrui Liu, Qihan Ren, Cheng Qian, Zhenhailong Wang, Minda Hu, Huazheng Wang, Qingyun Wu, Heng Ji, and Mengdi Wang. A survey of se...

    2026

  43. [43]

    Yu, and Jianfeng Gao

    Guiyao Tie, Jiawen Shi, Dingjie Song, Yixiao Huang, Ziji Sheng, Xueyang Zhou, Daizong Liu, Pan Zhou, Yongchao Chen, Ran Xu, Lifang He, Qingsong Wen, Manling Li, Cong Lu, Shuai Li, Pengtao Xie, Yixuan Yuan, Rui Meng, Lei Xing, Lichao Sun, Caiming Xiong, Philip S. Yu, and Jianfeng Gao. Au- toresearch ai: Towards ai-powered research automation for scientific...

    Pith/arXiv arXiv 2026

  44. [44]

    Towards end-to-end automation of AI research.Nature, 651(8107):914–919, 2026

    Chris Lu, Cong Lu, Robert Tjarko Lange, Yutaro Yamada, Shengran Hu, Jakob Foerster, David Ha, and Jeff Clune. Towards end-to-end automation of AI research.Nature, 651(8107):914–919, 2026

    2026

  45. [45]

    SR-Scientist: Scientific equation discovery with agentic AI

    Shijie Xia, Yuhan Sun, and Pengfei Liu. SR-Scientist: Scientific equation discovery with agentic AI. In International Conference on Learning Representations, 2026

    2026

  46. [46]

    Concrete problems in AI safety.arXiv preprint arXiv:1606.06565, 2016

    Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mane. Concrete problems in AI safety.arXiv preprint arXiv:1606.06565, 2016

    Pith/arXiv arXiv 2016

  47. [47]

    Specification gaming: The flip side of AI ingenuity

    Victoria Krakovna, Jonathan Uesato, Vladimir Mikulik, Matthew Rahtz, Tom Everitt, Ramana Kumar, Zac Kenton, Jan Leike, and Shane Legg. Specification gaming: The flip side of AI ingenuity. DeepMind Blog, 2020

    2020

  48. [48]

    Categorizing variants of goodhart’s law.arXiv preprint arXiv:1803.04585, 2018

    David Manheim and Scott Garrabrant. Categorizing variants of goodhart’s law.arXiv preprint arXiv:1803.04585, 2018

    Pith/arXiv arXiv 2018

  49. [49]

    Joar Skalse, Nikolaus H. R. Howe, Dmitrii Krasheninnikov, and David Krueger. Defining and charac- terizing reward hacking. InAdvances in Neural Information Processing Systems 35 (NeurIPS 2022), 2022

    2022

  50. [50]

    Scaling laws for reward model overoptimization

    Leo Gao, John Schulman, and Jacob Hilton. Scaling laws for reward model overoptimization. In Proceedings of the 40th International Conference on Machine Learning, pages 10835–10866, 2023

    2023

  51. [51]

    Reward gaming in conditional text generation

    Richard Yuanzhe Pang, Vishakh Padmakumar, Thibault Sellam, Ankur Parikh, and He He. Reward gaming in conditional text generation. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL), Volume 1: Long Papers, pages 4746–4763, 2023

    2023

  52. [52]

    playbook

    Jiacheng Wang and Jinbin Huang. Reward hacking as equilibrium under finite evaluation, 2026. arXiv:2603.28063 [cs.AI]. 20 A Related work We inherit our formulation from several literatures, generalize some of it, and depart from the rest. Classical multi-agent systems.The structural core of agents, communication, and shared state has a history in the mult...

    arXiv 2026