Pith. sign in

REVIEW 5 major objections 6 minor 51 references

LLM-Assisted Iterative Evolution with Swarm Intelligence Toward SuperBrain

T0 review · 5 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that millions of persistent user–LLM pairs can co-evolve through genetic-algorithm loops and swarm coordination into a 'Superclass Brain' — an emergent collective intelligence surpassing any single model or user.

desk verdict A coherent architecture proposal whose central emergence claim is untested: the forward-evolution pilot is a real but small LLM+GA demonstration, not evidence of a Superclass Brain. read the letter →

arxiv 2509.00510 v1 pith:WBXDZPIS submitted 2025-08-30 cs.AI cs.CL

Li Weigang , Pedro Carvalho Brom , Lucas Ramson Siefert This is my paper
classification cs.AIcs.CL
keywords ArtificialIntelligenceExplainableAIGeneticAlgorithmsLargeLanguageModelsSubclassBrainSuperclassSwarmUrbanAirMobility
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out a roadmap by which the everyday interaction between one person and one large language model becomes a persistent cognitive unit, called a Subclass Brain. It argues that when many such units run genetic-algorithm-assisted forward and backward learning loops — refining prompts, fitness functions, and solutions — they can be aggregated through swarm intelligence into a Superclass Brain, an emergent meta-intelligence with abstraction, generalization, and self-improvement. The experimental core is a UAV take-off scheduling study in which LLM-assisted genetic algorithm variants improved worst-case wait time by up to about 60 percent relative to the baseline GA, approaching the theoretical Round-Robin benchmark. If the framework holds, it would shift AI from static monolithic models toward a continuously co-evolving collective cognition grounded in real human diversity.

What carries the argument

The load-bearing mechanism is the bidirectional GA-assisted evolution loop. Forward Iterative Evolution runs on the user side: an LLM diagnoses failures, prototypes multi-objective fitness functions fλ, and generates GA parameter sets, producing (prompt, fλ, solution) triplets stored in the Subclass Brain Registry. Backward Iterative Evolution runs on the LLM side: a Meta-LLM Controller mutates prompts under KU/KI (Key-Useful / Key-Irrelevant) keyword constraints and a diversity margin δ, while a Worker-LLM Evaluator scores candidates, and distilled patterns update the user-conditioned policy πθ|u. The swarm alignment operator A = Σu αu S(p, cu), with weights αu proportional to user reliabil

What would settle it

Score each of about 20 prompts from the UAV scheduling task ten times with the same Worker-LLM evaluator and have domain experts rank them independently. If the evaluator's within-prompt score variance approaches the between-prompt differences, or if evaluator and expert rankings disagree substantially, then the GA selection, KU/KI updates, and cognitive signatures have no stable signal on which to build a Superclass Brain.

Watch

Extended reading notes

Core claim

The central claim is that collective intelligence can be synthesized without merging model parameters, from the standardized residue of many human–LLM dyads. Each high-value user u and their LLM form a Subclass Brain SBu := (u, Hu, Mu, πθ|u) — interaction history, persistent memory, a user-conditioned response policy — exported as a cognitive signature cu. A swarm alignment operator A aggregates these signatures into a prompt distribution Q, and a Meta-LLM distills them into a pattern library Π; the update rule closes the loop, formalized as SuperBrain := ({SBu}, A, D, U). The experiments demonstrate the forward half: with identical prompts, expert-driven GA variants (v7, v8) beat a non-expe

Load-bearing premise

The whole pipeline assumes the Worker-LLM Evaluator's prompt and solution scores are stable and unbiased enough to rank candidates; if repeated scoring of the same prompt is noisy or self-preferential, the genetic selection, KU/KI updates, and cognitive signatures inherit that noise, and the Superclass Brain would simply optimize the model's own biases.

Editorial extensions

If this is right

  • If many Subclass Brains persist and co-evolve, LLM providers could feed distilled swarm patterns back through RLHF or RLAIF pipelines, letting accumulated user contributions recursively shape future model behavior.
  • The fitness-alignment finding suggests that iterative prompt refinement can bring LLM+GA scheduling near the Round-Robin bound — GA v5 and v8 came within about 90 seconds of it — without adding algorithmic complexity.
  • The KU/KI registry would make collective AI explainable: every aggregated strategy would trace back to specific prompt features and their measured performance links.
  • Cross-platform consistency (GPT-4.0, Gemini 2.5 Pro, Grok 3, Claude 4) indicates that LLM-assisted GA evolution transfers across models, supporting the framework's universality claim.
  • If realized, the Superclass Brain would exceed both isolated LLM instances and any single Subclass Brain in abstraction and cross-domain generalization, per the paper's formal definition.
  • The forward-loop workflow itself (LLM-assisted fitness function design) already yields measurable engineering gains, so the Subclass Brain stage is testable independently of the emergent meta-intelligence claim.

Reading between the lines

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

  • A consequence the authors leave implicit: portable cognitive signatures would let a user's Subclass Brain migrate between LLM platforms, making personalization a transferable asset rather than a walled-garden feature.
  • A testable extension with the paper's own machinery: the Section 5 bilevel GA can be run without any LLM to establish a performance floor; only if the LLM-guided version beats that floor would the backward loop's contribution be demonstrated.
  • The paper's evidence covers only the forward (user-side) loop; backward evolution and Superclass Brain aggregation are specified but unvalidated, so the emergence claim currently rests on forward-loop results until the authors' stated next steps are run.
  • The aggregation operator could be stress-tested before scaling to millions of dyads: comparing the swarm-aligned prompt distribution Q against the best single Subclass Brain on a held-out task set would quantify how much the swarm layer actually adds.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes a conceptual framework, 'SuperBrain', in which persistent human–LLM dyads ('Subclass Brains') co-evolve through forward and backward iterative loops, are coordinated by swarm intelligence, and aggregate into an emergent 'Superclass Brain' capable of abstraction, generalization, and self-improvement. It formalizes these notions in Section 3.4 (Eqs. 12–16), describes a GA-assisted evolutionary architecture with KU/KI keyword filtering and a Subclass Brain Registry, and reports a pilot forward-evolution experiment on UAV vertiport take-off scheduling (Section 4, Tables 2–3). A bilevel, strongly convex formulation for the backward loop is presented in Section 5 but not implemented beyond a described validation protocol. The paper explicitly lists backward and swarm experiments as future work in Section 7.

Significance. If the central Superclass Brain claim were established, the framework would be a substantial contribution to human–LLM co-evolution and collective intelligence. The manuscript has genuine strengths: it gives explicit mathematical definitions for the proposed entities, provides a reproducible pilot of LLM-assisted GA forward evolution with real platforms and users, and contains a coherent theoretical construction of the backward optimization loop (Section 5.1–5.2). The paper is also honest in placing much of the full architecture in future work. However, the load-bearing claim—that swarm aggregation of Subclass Brains yields an emergent collective intelligence exceeding any single unit—is not tested end-to-end and is not derived from Eq. (15). At this stage the contribution is a conceptual roadmap plus a pilot study, not a validated collective-intelligence system.

major comments (5)
  1. [Section 3.4, Eq. (15)] The central emergence claim is unsupported. Eq. (15) defines Q(p) only as a weighted distribution over prompts computed from cognitive signatures; no theorem or experiment shows that this aggregation yields 'collective reasoning that exceeds any individual SBu' or produces abstraction, generalization, and self-improvement. A weighted reweighting over existing prompts does not, by itself, constitute a new cognitive entity. Section 7 explicitly defers swarm experiments. Please either provide a formal argument for why Eq. (15) + distillation yields emergent capabilities, or run a direct end-to-end test comparing the aggregate output against the best individual Subclass Brain on held-out tasks, or substantially weaken the claim to a hypothesis.
  2. [Section 5.3 and Section 7] The backward iterative evolution loop is not implemented. Section 5.3 says only that the 'expected outcome' is a consistent reduction in waiting time; no data, results, or repeated runs are reported. The closed-loop co-evolution of the dyad, which is essential to the SuperBrain concept, therefore rests on a theoretical construction rather than evidence. Since Section 7 also lists backward evolution implementation as future work, the manuscript should be reframed accordingly, and the forward-loop pilot should not be described as completing a closed-loop evolutionary cycle.
  3. [Tables 2 and 3, Section 4.3] The experimental claims about GA variants are not statistically grounded. Each GA variant appears to be a single run (n = 1 per condition), with no error bars, no repeated runs, and no significance tests. The 59.9% improvement in max wait time for GA v8 and the conclusions about expert versus student influence cannot be distinguished from noise under these conditions. The cross-platform and user-expertise comparisons additionally confound platform, user, and run variability. Please add repeated runs, variance measures, and an explicit account of how many independent trials underlie each number.
  4. [Section 3.2.2, Eq. (9), Figure 2] The stability of the Worker-LLM Evaluator is not examined. The GA selection, KU/KI updates, and cognitive signatures all depend on fλ scores produced by the same LLM that is being evolved. No calibration, repeated-evaluation agreement, or ground-truth check is reported. If the evaluator is noisy or self-preferential, the backward loop and the emergent claims inherit those biases. The paper itself acknowledges 'cognitive noise and self-bias' in Section 3.1 but does not mitigate this in the proposed loop. A stability/calibration experiment, or at least a sensitivity analysis over repeated scoring, is needed for the backward-evolution mechanism to be credible.
  5. [Section 3.3, Eq. (10)] The swarm fitness in Eq. (10) uses metrics such as 'Creativity', 'Generality', and 'Robustness' without operational definitions or measurement procedures, and the weights w_j are unspecified. This makes the swarm layer, and therefore the Superclass Brain formation pipeline, non-reproducible as stated. Please provide concrete operationalizations of these metrics or restrict the formal claims to the implemented forward-loop components.
minor comments (6)
  1. [Throughout] There are inconsistent spellings and spacing: 'UA V' vs. 'UAV', 'GA v1' vs. 'GAv1', and 'sucessfully' in the quoted prompt (the latter is authentic user text but could be marked as [sic]).
  2. [Eq. (7), Section 3.2.2] The notation Select(Mutate(Crossover(Pt); KU, KI)) is ambiguous about the order of filtering and how KU/KI constraints enter each operator. Please define the operators more precisely.
  3. [Eq. (15)] The reliability ρu used to define αu is not given an operational estimator. Clarify how ρu would be computed from interaction history or evaluation scores.
  4. [Section 5.1] The notation 'T95%' is introduced as a 95th percentile delay, but Section 5.1 says it is 'implemented through CVaR0.95 but kept in notation'. Please choose one definition consistently or explicitly state the relationship between the two.
  5. [Table 1 and Figure 3] The 'Source' column in Table 1 contains the unexplained entry 'Student EFO design', and Figure 3 is referenced in the text before its caption is introduced. Please clarify the figure/table integration and the meaning of 'EFO'.
  6. [Section 6.4] The Once-learning subsection, while interesting, is not well connected to the SuperBrain formal machinery. Consider integrating it into the architecture discussion or clearly stating its role as an inspirational analogy.

Circularity Check

1 steps flagged · score 6.0 of 10

Superclass Brain capability claim is stipulated in its definition; no derivation shows swarm aggregation yields superlinear collective cognition.

  1. self definitional [Section 3.4.2, Eq. (15)-(16)]
    "A Superclass Brain is the aggregation of multiple Subclass Brains via a swarm alignment operator A: Q(p) ∝ ∑ αu S(p, cu), αu ∝ ρu (15) ... The Meta-LLM consumes (SBR, Q) to produce a distilled pattern library Π, supporting collective reasoning that exceeds any individual SBu. ... the system self-organizes into a distributed, adaptive cognitive entity whose problem-solving capacity surpasses both isolated LLM instances and any single Subclass Brain."

    The load-bearing conclusion — aggregating Subclass Brains produces meta-intelligence beyond any individual — is not derived from the operators A, D, U; it is written into the definition of Superclass Brain/SuperBrain. Eq. (15) defines Q(p) as a weighted combination of existing cognitive signatures, and Eq. (16) labels the tuple as surpassing all individuals. No theorem or experiment shows that A and D applied to this reweighted prompt distribution create new capability; the asserted outcome is thus true by stipulation, reducing the 'prediction' to the definitional assertion.

full rationale

The paper's forward-evolution GA experiment (Tables 2-3) is a genuine, self-contained optimization exercise benchmarked against Round Robin, and the KU/KI loop, while self-referential, is not circular in the formal sense. The central Superclass Brain claim, however, is embedded in the definition itself: Eq. (15) defines the aggregation as a weighted distribution over prompts, and Eq. (16) simultaneously asserts that the resulting system surpasses all individuals, with no supporting derivation or end-to-end test. Section 7 explicitly defers backward and swarm experiments to future work, confirming that the emergence claim is stipulated rather than derived. Self-citations to Eixão-UAM and Once-learning supply methodology and historical priority but are not the basis of the circular step. Overall: partial circularity, since the framework's flagship prediction reduces to its own definition while the GA pilot remains independent.

Assumptions & free parameters 7 free parameters · 6 assumptions · 5 invented entities

The central claim rests on user-LLM dyad persistence, reliable LLM self-evaluation, and the existence of distillable collective patterns. None of these is supported by independent evidence in the paper. The main free parameters are GA fitness weights and diversity thresholds, tuned interactively or specified by hand; the formal framework contributes definitions and an algorithm sketch rather than a parameter-free derivation.

free parameters (7)
  • Fitness weights alpha_1, alpha_2, alpha_3 (Eq 5, 17) = not fixed in paper; interactively tuned per experiment
    Weights balancing average wait, std, tail and pad penalties are tuned by the user with LLM help across GA v1-v11; results depend on them.
  • Pad penalty weights w_k(t) (Eq 17)
    Time-varying per-pad penalties introduced in GA v3 and used in Eq 17 are task-specific and hand/LLM-set.
  • Diversity margin delta (Eq 8)
    Admissibility threshold for prompt diversity; chosen without sensitivity analysis.
  • Regularization coefficients lambda_tok, lambda_div, lambda_exp (Eq 11)
    Defined in fitness function but no values or fitting procedure given.
  • GA outer-loop hyperparameters (Npop, G, pc, pm, kappa, msamp, Nelite)
    Section 5.2 lists ranges and choices for the bilevel GA; results would vary with them.
  • Strong-convexity floor eta (Eq 21/23)
    Imposed alpha_{4} >= eta to guarantee strong convexity; value not specified nor tested for scheduling fidelity.
  • User registration threshold (10+ purposeful interactions) = 10+
    Section 1 declares high-value users are those with 10+ purposeful interactions; arbitrary cut-off affects who forms a Subclass Brain.
assumptions (6)
  • domain assumption Worker-LLM Evaluator returns reliable y and metrics M_k for any prompt p
    Eq 9 and Figure 2 treat the LLM as both generator and scorer; no validation of scoring stability.
  • ad hoc to paper Persistent interaction internalizes user patterns into a stable SBu with signature cu
    Section 3.1 and Eq 12-13 assert the dyad forms a persistent cognitive entity; this is the key postulate.
  • ad hoc to paper Aggregation operator A and meta-LLM distillation yield abstraction, generalization and self-improvement
    Section 3.4.2 defines Superclass Brain by this capability rather than deriving or testing it.
  • domain assumption Continuous convex relaxation of scheduling is faithful for real vertiport sequences
    Section 5.1 moves from discrete sequences to continuous z with capacity constraints; approximation error to Eq 5 not quantified.
  • domain assumption LLM platform differences and user expertise are the only explanatory variables in Tables 2 and 3
    Section 4.1 controls prompts but not iteration counts, temperature, or randomness; causal attribution in Section 4.3 assumes this.
  • standard math Log-sum-exp is convex and monotone; g4 strongly convex implies F strongly convex (Eq 19-21)
    This is standard math used to justify the bilevel inner problem; accepted as a lemma without proof in Section 5.1.
invented entities (5)
  • Subclass Brain (SBu)
    purpose: Persistent user-LLM dyad that accumulates memory and cognitive signature
    Defined only inside the paper; no external observable beyond ordinary chat logs, no falsifiable prediction.
  • Superclass Brain
    purpose: Emergent collective meta-intelligence from many Subclass Brains
    Claimed to generalize and self-improve, but no operational test or external handle.
  • Cognitive signature cu
    purpose: Vector encoding user history, KU/KI, behavior tags and reliability rho_u
    No defined measurement protocol or benchmark; Eq 13 is a specification.
  • Subclass Brain Registry (SBR)
    purpose: Store and retrieve cognitive signatures for swarm alignment
    Not implemented or evaluated.
  • KU/KI keyword sets
    purpose: Filter prompts correlated with high or low fitness
    Needs LLM scoring and quartile thresholds, both internal to the loop.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLM-Assisted Iterative Evolution with Swarm Intelligence Toward SuperBrain." pith.science (2026). https://pith.science/paper/WBXDZPIS

@misc{pith2026250900510,
  author       = {Pith},
  title        = {Pith review of: LLM-Assisted Iterative Evolution with Swarm Intelligence Toward SuperBrain},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WBXDZPIS}},
  note         = {Machine review of arXiv:2509.00510}
}
read the original abstract

We propose a novel SuperBrain framework for collective intelligence, grounded in the co-evolution of large language models (LLMs) and human users. Unlike static prompt engineering or isolated agent simulations, our approach emphasizes a dynamic pathway from Subclass Brain to Superclass Brain: (1) A Subclass Brain arises from persistent, personalized interaction between a user and an LLM, forming a cognitive dyad with adaptive learning memory. (2) Through GA-assisted forward-backward evolution, these dyads iteratively refine prompts and task performance. (3) Multiple Subclass Brains coordinate via Swarm Intelligence, optimizing across multi-objective fitness landscapes and exchanging distilled heuristics. (4) Their standardized behaviors and cognitive signatures integrate into a Superclass Brain, an emergent meta-intelligence capable of abstraction, generalization and self-improvement. We outline the theoretical constructs, present initial implementations (e.g., UAV scheduling, KU/KI keyword filtering) and propose a registry for cross-dyad knowledge consolidation. This work provides both a conceptual foundation and an architectural roadmap toward scalable, explainable and ethically aligned collective AI.

Figures

Figures reproduced from arXiv: 2509.00510 by the authors.

Figure 1
Figure 1. Logical architecture of the proposed SuperBrain framework. Subclass Brains evolve through swarm-based [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Closed-loop Forward/Backward Iterative Evolution for Subclass Brain formation. The Subclass Brain Registry [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Swarm intelligence framework [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Bilevel GA with Dirichlet–Multinomial weight sampling and LSE-based convex interactions for strongly [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: Timeline of Once-learning and related concepts, highlighting its precedence and conceptual continuity in AI development. Looking forward, Once-learning could inspire global tokenization and encoding mechanisms in future LLM/Transformer architectures, enabling models to…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 38 canonical work pages

  1. [1]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30, pages 5998--6008, 2017

    work page 2017

  2. [2]

    Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors

    Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chi-Min Chan, Heyang Yu, Yaxi Lu, Yi-Hsin Hung, Chen Qian, et al. Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors. In The International Conference on Learning Representations (ICLR), 2024

    work page 2024

  3. [3]

    Multi-agent collaboration mechanisms: A survey of llms

    Khanh-Tung Tran, Dung Dao, Minh-Duong Nguyen, Quoc-Viet Pham, Barry O'Sullivan, and Hoang D Nguyen. Multi-agent collaboration mechanisms: A survey of llms. arXiv preprint arXiv:2501.06322, 2025

    arXiv 2025

  4. [4]

    Multi-agent systems powered by large language models: applications in swarm intelligence

    Cristian Jimenez-Romero, Alper Yegenoglu, and Christian Blum. Multi-agent systems powered by large language models: applications in swarm intelligence. Frontiers in Artificial Intelligence, 8: 0 1593017, 2025

    work page 2025

  5. [5]

    Dynamic llm-agent network: An llm-agent collaboration framework with agent team optimization

    Zijun Liu, Yanzhe Zhang, Peng Li, Yang Liu, and Diyi Yang. Dynamic llm-agent network: An llm-agent collaboration framework with agent team optimization. arXiv preprint arXiv:2310.02170, 2023

    arXiv 2023

  6. [6]

    Emergent analogical reasoning in large language models

    Taylor Webb, Keith J Holyoak, and Hongjing Lu. Emergent analogical reasoning in large language models. Nature Human Behaviour, 7 0 (9): 0 1526--1541, 2023

    work page 2023

  7. [7]

    Llm4pm: A case study on using large language models for process modeling in enterprise organizations

    Clara Ziche and Giovanni Apruzzese. Llm4pm: A case study on using large language models for process modeling in enterprise organizations. In International Conference on Business Process Management, pages 472--483. Springer, 2024

    work page 2024

  8. [8]

    Large language models for manufacturing

    Yiwei Li, Huaqin Zhao, Hanqi Jiang, Yi Pan, Zhengliang Liu, Zihao Wu, Peng Shu, Jie Tian, Tianze Yang, Shaochen Xu, et al. Large language models for manufacturing. arXiv preprint arXiv:2410.21418, 2024 a

    arXiv 2024

Show all 51 references
  1. [9]

    New directions for artificial intelligence: human, machine, biological, and quantum intelligence

    Li Weigang, Liriam Michi Enamoto, Denise Leyi Li, and Geraldo Pereira Rocha Filho. New directions for artificial intelligence: human, machine, biological, and quantum intelligence. Frontiers of Information Technology & Electronic Engineering, 23 0 (6): 0 984--990, 2022

    2022

  2. [10]

    Daniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, and Romeo Dean. Ai 2027. https://ai-2027.com/, 2025. Accessed August 2025

    2027

  3. [11]

    Artificial general intelligence for medical imaging analysis

    Xiang Li, Lin Zhao, Lu Zhang, Zihao Wu, Zhengliang Liu, Hanqi Jiang, Chao Cao, Shaochen Xu, Yiwei Li, Haixing Dai, et al. Artificial general intelligence for medical imaging analysis. IEEE Reviews in Biomedical Engineering, 2024 b

    2024

  4. [12]

    A survey on llm-based multi-agent systems: workflow, infrastructure, and challenges

    Xinyi Li, Sai Wang, Siqi Zeng, Yu Wu, and Yi Yang. A survey on llm-based multi-agent systems: workflow, infrastructure, and challenges. Vicinagearth, 1 0 (1): 0 9, 2024 c

    2024

  5. [13]

    Collaborative problem-solving with llm: a multi-agent system approach to solve complex tasks using autogen

    Ricardo Barbosa, Ricardo Santos, and Paulo Novais. Collaborative problem-solving with llm: a multi-agent system approach to solve complex tasks using autogen. In International Conference on Practical Applications of Agents and Multi-Agent Systems, pages 203--214. Springer, 2024

    2024

  6. [14]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 0 2...

    2022

  7. [15]

    Weigang, A

    L. Weigang, A. Juliano. Maia, Emilia Stensel, and R. Lucas Siefert. Eixao-uam: Llm-assisted iterative development of a low-altitude urban air mobility corridor in brasilia. Frontiers of Information Technology and Electronic Engineering, submitted, 2025

    2025

  8. [16]

    Artificial Brain: No Longer Fantasy

    Andrezej Buller. Artificial Brain: No Longer Fantasy. Hunan Science & Technology Press, in Chinese, thanslated by Liu Juan and Li Weigang, 2001, from Proszyrisky i S-ka,, 1998

    2001

  9. [17]

    Two paths for a.i

    Joshua Rothman. Two paths for a.i. https://www.newyorker.com/culture/open-questions/two-paths-for-ai?, 2025. Accessed August 2025

    2025

  10. [18]

    Generative agents: Interactive simulacra of human behavior

    Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, pages 1--22, 2023

    2023

  11. [19]

    Using large language models to simulate multiple humans and replicate human subject studies

    Gati V Aher, Rosa I Arriaga, and Adam Tauman Kalai. Using large language models to simulate multiple humans and replicate human subject studies. In International conference on machine learning, pages 337--371. PMLR, 2023

    2023

  12. [20]

    Can ai replace human subjects? a large-scale replication of psychological experiments with llms

    Ziyan Cui, Ning Li, and Huaikang Zhou. Can ai replace human subjects? a large-scale replication of psychological experiments with llms. A Large-Scale Replication of Psychological Experiments with LLMs (August 25, 2024), 2024

    2024

  13. [21]

    How large language models can reshape collective intelligence

    Jason W Burton, Ezequiel Lopez-Lopez, Shahar Hechtlinger, Zoe Rahwan, Samuel Aeschbach, Michiel A Bakker, Joshua A Becker, Aleks Berditchevskaia, Julian Berger, Levin Brinkmann, et al. How large language models can reshape collective intelligence. Nature human behaviour, 8 0 (...

    2024

  14. [22]

    Wisdom of the machines: Exploring collective intelligence in llm crowds

    Yashar Talebirad, Ali Parsaee, Vishwajeet Ohal, Amirhossein Nadiri, Csongor Szepesvari, Yash Mouje, and Eden Redman. Wisdom of the machines: Exploring collective intelligence in llm crowds. In First Workshop on Social Simulation with LLMs, 2025

    2025

  15. [23]

    Collective reasoning among llms a framework for answer validation without ground truth

    Seyed Pouyan Mousavi Davoudi, Alireza Shafiee Fard, and Alireza Amiri-Margavi. Collective reasoning among llms a framework for answer validation without ground truth. arXiv preprint arXiv:2502.20758, 2025

    2025 arXiv

  16. [24]

    Agent-based human behavior modeling for crowd simulation

    Linbo Luo, Suiping Zhou, Wentong Cai, Malcolm Yoke Hean Low, Feng Tian, Yongwei Wang, Xian Xiao, and Dan Chen. Agent-based human behavior modeling for crowd simulation. Computer Animation and Virtual Worlds, 19 0 (3-4): 0 271--281, 2008

    2008

  17. [25]

    We are raising a tiger: Waic 2025 keynote

    Geoffrey Hinton. We are raising a tiger: Waic 2025 keynote. https://pandaily.com/ai-godfather-geoffrey-hinton-urges-global-ai-cooperation-at-waic-2025-in-shanghai, 2025. Accessed August 2025

    2025

  18. [26]

    China proposes global framework for benevolent ai at waic 2025

    State Council of China . China proposes global framework for benevolent ai at waic 2025. https://example.com/china-ai-framework-waic2025, 2025. Accessed August 2025

    2025

  19. [27]

    Genetic algorithms and game theory for airport departure decision making: Gedman and codman

    Deborah Mendes Ferreira, Lucas Pessoa Rosa, Vitor Filincowsky Ribeiro, Fl \'a vio de Barros Vidal, and Li Weigang. Genetic algorithms and game theory for airport departure decision making: Gedman and codman. In International Conference on Knowledge Management in Organizations,...

    2014

  20. [28]

    Dynamic super round-based distributed task scheduling for uav networks

    Subir Halder, Amrita Ghosal, and Mauro Conti. Dynamic super round-based distributed task scheduling for uav networks. IEEE Transactions on Wireless Communications, 22 0 (2): 0 1014--1028, 2022

    2022

  21. [29]

    Multi-objective routing optimization in electric and flying vehicles: a genetic algorithm perspective

    Muhammad Alolaiwy, Tarik Hawsawi, Mohamed Zohdy, Amanpreet Kaur, and Steven Louis. Multi-objective routing optimization in electric and flying vehicles: a genetic algorithm perspective. Applied Sciences, 13 0 (18): 0 10427, 2023

    2023

  22. [30]

    Swarm intelligence

    James Kennedy. Swarm intelligence. In Handbook of nature-inspired and innovative computing: integrating classical models with emerging technologies, pages 187--219. Springer, 2006

    2006

  23. [31]

    Swarm intelligent systems, volume 26

    Nadia Nedjah and Luiza de Macedo Mourelle. Swarm intelligent systems, volume 26. Springer, 2006

    2006

  24. [32]

    Using ant algorithm to arrange taxiway sequencing in airport

    Kamila B Nogueira, Paulo HC Aguiar, Li Weigang, et al. Using ant algorithm to arrange taxiway sequencing in airport. International Journal of Computer Theory & Engineering, 6 0 (4): 0 857--361, 2014

    2014

  25. [33]

    A review: Swarm robotics: Cooperative control in multi-agent systems

    Kskn Venkata Ramana Devi, BS Smitha, Sorabh Lakhanpal, Ravi Kalra, Vandana Arora Sethi, and Sadiq Khader Thajil. A review: Swarm robotics: Cooperative control in multi-agent systems. In E3S Web of Conferences, volume 505, page 03013. EDP Sciences, 2024

    2024

  26. [34]

    Deep reinforcement learning from human preferences

    Paul F Christiano, Jan Leike, Tom B Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. In Advances in Neural Information Processing Systems (NeurIPS), pages 4299--4307, 2017

    2017

  27. [35]

    Reflexion: Language agents with verbal reinforcement learning

    Noah Shinn, Federico Cassano, Arjun Gopinath, Shunyu Zhang, Chelsea Finn, and Jiajun Wu. Reflexion: Language agents with verbal reinforcement learning. arXiv preprint arXiv:2303.11366, 2023

    2023 arXiv

  28. [36]

    Self-refine: Iterative refinement with self-feedback

    Aman Madaan, Yuchen Zhou, Uri Alon, Maor Ivgi, Shuyan Zhang, Boaz Barak, Graham Neubig, Tianyi Yang, Mohit Bansal, Jonathan Clark, et al. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  29. [37]

    Distilling the knowledge in a neural network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015

    2015 arXiv

  30. [38]

    A survey on knowledge distillation: Recent advancements

    Amir Moslemi, Anna Briskina, Zubeka Dang, and Jason Li. A survey on knowledge distillation: Recent advancements. Machine Learning with Applications, 18: 0 100605, 2024

    2024

  31. [39]

    Raft: Adapting language model to domain specific rag

    Tianjun Zhang, Shishir G Patil, Naman Jain, Sheng Shen, Matei Zaharia, Ion Stoica, and Joseph E Gonzalez. Raft: Adapting language model to domain specific rag. In First Conference on Language Modeling, 2024

    2024

  32. [40]

    Slim-raft: A novel fine-tuning approach to improve cross-linguistic performance for mercosur common nomenclature

    Vin \' cius Di Oliveira, Yuri Fa c anha Bezerra, Li Weigang, Pedro Carvalho Brom, and Victor Rafael R Celestino. Slim-raft: A novel fine-tuning approach to improve cross-linguistic performance for mercosur common nomenclature. In Proceedings of the 20th International Conferenc...

    2024

  33. [41]

    Llmquoter: enhancing rag capabilities through efficient quote extraction from large contexts

    Yuri Fa c anha Bezerra and Li Weigang. Llmquoter: enhancing rag capabilities through efficient quote extraction from large contexts. In Proceedings of the 17th International Conference on Agents and Artificial Intelligence, volume 3, pages 1335--1342. Scitepress, 2025

    2025

  34. [42]

    The paradox of poetic intent in back-translation: Evaluating the quality of large language models in chinese translation

    Li Weigang and Pedro Carvalho Brom. The paradox of poetic intent in back-translation: Evaluating the quality of large language models in chinese translation. arXiv preprint arXiv:2504.16286, 2025

    2025 arXiv

  35. [43]

    A study of parallel self-organizing map

    Li Weigang. A study of parallel self-organizing map. arXiv preprint quant-ph/9808025, 1998

    1998 arXiv

  36. [44]

    A study of parallel neural networks

    Li Weigang and Nilton Correia da Silva. A study of parallel neural networks. In IJCNN'99. International Joint Conference on Neural Networks. Proceedings (Cat. No. 99CH36339), volume 2, pages 1113--1116. IEEE, 1999

    1999

  37. [45]

    A bayesian approach to unsupervised one-shot learning of object categories

    Li Fe-Fei et al. A bayesian approach to unsupervised one-shot learning of object categories. In proceedings ninth IEEE international conference on computer vision, pages 1134--1141. IEEE, 2003

    2003

  38. [46]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33: 0 1877--1901, 2020

    1901

  39. [47]

    You only look once: Unified, real-time object detection

    Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779--788, 2016

    2016

  40. [48]

    Once learning for looking and identifying based on yolo-v5 object detection

    Lucas Althoff, Myl \`e ne CQ Farias, and Li Weigang. Once learning for looking and identifying based on yolo-v5 object detection. In Proceedings of the Brazilian Symposium on Multimedia and the Web, pages 298--304, 2022

    2022

  41. [49]

    Benchmarking human--ai collaboration for common evidence appraisal tools

    Tim Woelfle, Julian Hirt, Perrine Janiaud, Ludwig Kappos, John PA Ioannidis, and Lars G Hemkens. Benchmarking human--ai collaboration for common evidence appraisal tools. Journal of Clinical Epidemiology, 175: 0 111533, 2024

    2024

  42. [50]

    Improving language models by retrieving from trillions of tokens

    Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean-Baptiste Lespiau, Jonathan Godwin, et al. Improving language models by retrieving from trillions of tokens. arXiv preprint arXiv:2112.04426, 2022

    2022 arXiv

  43. [51]

    Taskmatrix.ai: Completing tasks by connecting foundation models with millions of apis

    Weizhe Liang, Yizhong Zhang, Yuchen Zhang, Zhiyuan Shao, Huan Zhang, Hongyu Zhou, Yue Yu, Jiayi Li, Kunpeng Zhang, et al. Taskmatrix.ai: Completing tasks by connecting foundation models with millions of apis. arXiv preprint arXiv:2303.16434, 2023

    2023 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.