Pith. sign in

REVIEW 3 major objections 6 minor 57 references

FAIRTOPIA: Envisioning Multi-Agent Guardianship for Disrupting Unfair AI Pipelines

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper proposes FAIRTOPIA, a three-layered, agent-driven architecture that turns AI fairness from a stage-by-stage audit into continuous, knowledge-informed, human-centric guardianship.

desk verdict A serious position paper whose central algorithm doesn't do what the text says it does; worth discussing but not citing yet. read the letter →

arxiv 2506.09107 v1 pith:OSOFQ464 submitted 2025-06-10 cs.CY cs.AI

Athena Vakali , Ilias Dimitriadis This is my paper
classification cs.CYcs.AI
keywords AIfairnessmulti-agentsystemsbydesignknowledgegraphsagentichuman-in-the-loopalgorithmicbiassocio-technicalframework
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

FAIRTOPIA is a proposal paper rather than a deployed system. It argues that fairness in AI should be guarded continuously by three cooperating agents—a planner for data preparation, an action agent for model construction, and an optimizer for deployment—instead of being measured once at each pipeline stage. The authors claim this turns fairness from a reactive, post-hoc audit into a self-refining, human-centric design principle, with a knowledge base linking known human cognitive biases to computational biases. A generalized Algorithm 1 is offered so that any organization can customize the guardianship to its own fairness goals and risk constraints. If the proposal holds, high-stakes AI applications such as lending, healthcare, and chatbots could receive adaptive, realistic fairness guardrails rather than static fairness checks.

What carries the argument

The load-bearing mechanism is Algorithm 1, the Fairness Agentic Alignment algorithm. It pairs the AI pipeline with three agents: $F_{pla}$ consumes a task's fairness goals and knowledge graphs to emit a Guardrailset, $F_{act}$ executes candidate methods and recommends a Bestoutcome, and if that outcome fails the scenario-specific fairness goal $F_{GT}$, $F_{opt}$ self-critiques using the knowledge graphs and method set, then reflects and optimizes the task plans. The loop continues until $F_{opt}$ no longer fails on $F_{GT}$ or a trial budget $\max(\mathrm{trials})$ is reached. Architecturally, the scheme is organized as a bowtie: knowledge and planning form the causal side, the in-processing core sits in the middle, and optimization forms the consequence side.

What would settle it

Run Algorithm 1 on a public fair-ML benchmark with a concrete fairness target, such as an equalized-odds gap below 0.05 on a lending or recidivism prediction task, from multiple random seeds; if in a substantial share of runs the loop hits $\max(\mathrm{trials})$ without meeting the target, or does no better than a single one-shot mitigation, the convergence premise is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a three-layered socio-technical architecture can enact fairness watch across the whole AI pipeline. The AI layer keeps the standard data, model, and deployment stages, but these stages are no longer isolated: an agentic layer of three role-specialized agents—$F_{pla}$ for planning and pre-processing, $F_{act}$ for in-processing action, and $F_{opt}$ for post-processing optimization—iteratively produces guardrails, while a knowledge-and-reform layer supplies structured knowledge and refines task plans. Fairness goals $F_{GT}$ are defined per scenario, and Algorithm 1 loops until the optimizer agent no longer fails those goals or the trial budget is exhausted. The authors position this as fairness by design: human oversight appears at the points where guardrails indicate fairness leakage, and fairness is continuously negotiated rather than externally audited.

Load-bearing premise

The framework's success depends on the unproven assumption that the three agents will, through self-critique and re-optimization, actually converge on plans that satisfy the scenario's fairness goals within a bounded number of tries.

Editorial extensions

If this is right

  • Guardrails would be generated and refreshed at every pipeline stage, so fairness problems could be caught before deployment rather than after harm is reported.
  • The same generalized algorithm could be adapted across domains and regulatory regimes by substituting task-specific fairness goals, risk levels, and knowledge sources.
  • Human oversight would become targeted: people enter the loop only when the agents' guardrails signal fairness leakage, making human-in-the-loop involvement more efficient.
  • Cognitive and computational bias research could be consolidated into machine-readable knowledge graphs, allowing systematic reasoning about how human biases propagate into AI systems.
  • Fairness monitoring would evolve with the AI system instead of being a static, one-time audit.

Reading between the lines

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

  • The self-critique loop is effectively a test-time optimization scheme, which implies fairness could be improved by extra inference-time computation rather than retraining; the paper leaves this economic consequence implicit.
  • A concrete next step would be to realize the three agents with current tool-using language models and measure loop convergence on existing fairness benchmarks; that experiment would either support or undermine the proposal's feasibility.
  • The bowtie layout invites a causal reading: the knowledge base could be used to model how cognitive biases propagate into specific pipeline-stage harms, and to predict where fairness leakage will emerge in a new deployment.
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

3 major / 6 minor

Summary. The manuscript is a position statement proposing FAIRTOPIA, a three-layer architecture (AI, Agentic, Knowledge and Reform) in which three AI agents—Fpla, Fact, and Fopt—act as fairness guardians across pre-, in-, and post-processing pipeline stages. It introduces a Fairness Knowledge Warehouse built from KGs, LLMs, and RAGs, and presents Algorithm 1 as a 'generalized AI fairness alignment algorithm' that iteratively plans, recommends, and optimizes an AI task against fairness goals F_GT. The paper argues that fairness should shift from reactive, stage-specific bias metrics to continuous, human-centric, socio-technical guardrailing. No experiments or formal proofs are provided; Section 5 offers a structured critique-and-advocacy discussion.

Significance. The paper's interdisciplinary framing and explicit critique-and-advocacy structure are strengths; it identifies a real gap—agentic AI fairness is underexplored—and its call for shared fairness knowledge graphs and evaluation benchmarks is constructive. If the FAIRTOPIA architecture were realized, it could provide a useful template for embedding human oversight and multi-agent reflection into AI pipelines. However, the contribution is currently at the vision level: the central algorithm is the only formal artifact, and it has a control-flow defect (Algorithm 1, lines 8–15) that prevents it from delivering the claimed fairness guarantee. Because the claims are framed as achievements ('will resolve current AI fairness bottlenecks', 'safeguards fairness leakage end-to-end') rather than as research proposals, the paper overstates its current support. The conceptual direction is defensible, but the formal specification and the claim-evidence calibration need revision.

major comments (3)
  1. [3.2, Algorithm 1 (lines 8–15)] The loop condition 'while Fopt fails on F_GT or t < max(trials)' is a disjunction, and the inner increment t ← t+1 is inside the 'if Bestoutcome fails' branch. Consequently, if Bestoutcome succeeds, t never changes and the second disjunct keeps the loop running forever; if Fopt never succeeds, the first disjunct keeps the loop running forever after t reaches max(trials). The loop therefore has no correctness-based termination at all: it cannot exit because fairness is achieved, and it returns no fairness certificate. In addition, 'Fopt fails on F_GT' is evaluated before Fopt has produced any output, and the failure predicate for F_GT is never formally defined. As written, Algorithm 1 cannot provide the 'fairness alignment' or 'fairness watch' guarantee that Section 3.2 and Section 4 attribute to it. This is load-bearing because Algorithm 1 is the only formal specification of the framework's core mechanism. I recommend rewriting the loop so that termination is governed by an explicit convergence predicate on Bestoutcome, with a bounded iteration count as a safety stop, and defining the failure test on F_GT and the return value (Guardrailset, Bestoutcome) precisely in the pseudocode.
  2. [Abstract and Section 4] The manuscript uses achievement-level language—'we enact fairness watch in all of the AI pipeline stages' (Abstract) and 'safeguards fairness leakage end-to-end' (Section 4)—but Section 5 itself lists major open research questions and critiques, and no empirical or formal evidence is supplied. For a position paper this would be acceptable if the claims were phrased as proposals or hypotheses; as written, the claims promise capabilities that are neither demonstrated nor specified. In particular, Section 4 states that Fpla 'will follow the proposed algorithmic steps' to 'flag the guardrails generated', but no implementation, instantiation, or concrete test of the algorithm is provided. I recommend reframing the contributions as a research agenda with explicit falsifiable predictions and an evaluation roadmap, and softening the claimed guarantees throughout.
  3. [5, Critique 2] The metric-space formulation used to rebut Critique 2 is not defined enough to support the argument: CS=(B, db) and AS=(Bhat, dhat) require specifying what the distance functions db and dhat measure and how the bias reflection function r: B → Bhat is constructed from the proposed knowledge graphs. Without these definitions, the claim that 'human and AI bias types similarities detection' and their 'reflections, hierarchies, patterns, and dynamics' can be harvested is an assertion rather than a mechanism. If this counterargument is retained, the authors should either give a concrete instantiation or explicitly mark it as an open research direction.
minor comments (6)
  1. [Abstract] The sentence 'AI unfairness in heavily criticized' is ungrammatical; it should be 'AI unfairness is heavily criticized'. Similarly, 'so called FAIRTOPIA framework' should be 'so-called FAIRTOPIA framework'.
  2. [Title and Section 2] There are several typos: the title uses spacing artifacts in 'E NVISIONING' and 'P IPELINES'; Section 1 has 'dymamic'; Section 2 has 'refedined' and 'out-of-the-self' (twice, in Section 3.1); Section 4 uses 'quardrails' and 'quadrails' inconsistently; Section 5 has 'Advovacy' twice. A careful proofreading pass is needed.
  3. [3.2, Algorithm 1] The pseudocode notation mixes assignment and function application, e.g., 'Fpla ← [F_GT; Tplans] and produce-fairness(Guardrailset)' and 'Fact ← [Guardrailset; Tplans] and recommend(Bestoutcome)'; it is unclear whether these are assignments, function calls, or parallel updates. The pseudocode should use a consistent convention for inputs, outputs, and function calls.
  4. [4, Figure 4] The bowtie-like design is described as 'particularly useful in risk analysis' but the analogy is not used to derive any design constraint or component of the architecture. Either make the bowtie analogy operational (e.g., specify the central event, threat paths, and controls) or remove it to avoid giving the false impression of a formal risk-analysis method.
  5. [References] Several references are malformed: [22] duplicates 'Haofen Wang' as author; [35] begins 'Dwivedi S Dwivedi S, Ghosh S.'; and [36] lacks a publication venue. These should be corrected to meet the journal's reference style.
  6. [5, Critique 1] The statement that 'by 2028, 33% of enterprise software applications will incorporate it' is attributed to reference [56], which is a Gartner prediction; the citation should specify the original source and date, and the wording should indicate that this is an analyst forecast rather than an established fact.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: FAIRTOPIA is a conceptual architecture paper with no fitted quantities or derivations that reduce to their inputs.

full rationale

FAIRTOPIA is a position/vision paper. It introduces a three-layer agentic guardianship architecture and a generalized fairness-alignment algorithm, but it makes no quantitative predictions and fits no parameters. The central proposals—the knowledge warehouse, the Fpla/Fact/Fopt agents, and the layered framework—are presented as designs to be implemented, not as results derived from data or from the paper's own outputs. The inputs (F_GT, KGs, Mset, max(trials)) are stipulated, and Bestoutcome/Guardrailset are proposed outputs of the loop; no equation or definition identifies an output with an input. The only mathematical objects, the metric spaces CS=(B, db), AS=(Bhat, dhat), and the reflection function r: B→Bhat, are explicitly introduced as suggestions for future research within Critique 2, not as proven derivations. The two self-citations ([3] and [57]) are used for motivating examples and existing health-domain fairness guidelines; they are not load-bearing for the central claim and do not constitute a self-citation chain or imported uniqueness theorem. The known issue in Algorithm 1—its while-loop condition being always satisfied until max(trials) is reached—is a specification/correctness concern about termination and fairness certification, not a circularity between input and output. The paper's own Section 5 acknowledges that substantial research remains, which is consistent with a non-circular, open proposal. Therefore no circular step can be exhibited, and the circularity score is 0.

Assumptions & free parameters 1 free parameters · 4 assumptions · 2 invented entities

The FAIRTOPIA framework rests on untested assumptions about agentic AI reliability, the representability of biases, and the ability to build a comprehensive knowledge warehouse. These are domain assumptions stated or implied in the paper, not established facts.

free parameters (1)
  • max(trials)
    Unspecified termination bound in Algorithm 1; the behavior of the loop depends on it.
assumptions (4)
  • domain assumption Agentic AI systems can reliably plan, act, and self-critique to achieve fairness goals.
    Algorithm 1's success depends on agents improving outcomes, but no evidence is provided.
  • domain assumption Cognitive and AI biases can be represented as metric spaces and knowledge graphs with meaningful structure.
    Section 5 Critique 2 assumes CS and AS metric spaces and a bias reflection function without proof.
  • domain assumption A comprehensive fairness knowledge warehouse can be built from fragmented existing research.
    Section 3.1 assumes feasibility of harvesting and unifying diverse sources.
  • domain assumption Human-in-the-loop involvement will not reduce automation and can be selectively invoked by agents.
    Section 4 states this as a design goal without operational validation.
invented entities (2)
  • Fpla, Fact, and Fopt fairness guardian agents
    purpose: To generate and enforce fairness guardrails across pre-, in-, and post-processing stages
    Proposed as the core mechanism of FAIRTOPIA, but no implementation or evaluation exists.
  • Fairness knowledge warehouse
    purpose: To aggregate cognitive and AI bias knowledge into knowledge graphs, embeddings, and vectorized data for the agents
    Described as a planned repository; no proof of construction or coverage.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FAIRTOPIA: Envisioning Multi-Agent Guardianship for Disrupting Unfair AI Pipelines." pith.science (2026). https://pith.science/paper/OSOFQ464

@misc{pith2026250609107,
  author       = {Pith},
  title        = {Pith review of: FAIRTOPIA: Envisioning Multi-Agent Guardianship for Disrupting Unfair AI Pipelines},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OSOFQ464}},
  note         = {Machine review of arXiv:2506.09107}
}
read the original abstract

AI models have become active decision makers, often acting without human supervision. The rapid advancement of AI technology has already caused harmful incidents that have hurt individuals and societies and AI unfairness in heavily criticized. It is urgent to disrupt AI pipelines which largely neglect human principles and focus on computational biases exploration at the data (pre), model(in), and deployment (post) processing stages. We claim that by exploiting the advances of agents technology, we will introduce cautious, prompt, and ongoing fairness watch schemes, under realistic, systematic, and human-centric fairness expectations. We envision agents as fairness guardians, since agents learn from their environment, adapt to new information, and solve complex problems by interacting with external tools and other systems. To set the proper fairness guardrails in the overall AI pipeline, we introduce a fairness-by-design approach which embeds multi-role agents in an end-to-end (human to AI) synergetic scheme. Our position is that we may design adaptive and realistic AI fairness frameworks, and we introduce a generalized algorithm which can be customized to the requirements and goals of each AI decision making scenario. Our proposed, so called FAIRTOPIA framework, is structured over a three-layered architecture, which encapsulates the AI pipeline inside an agentic guardian and a knowledge-based, self-refining layered scheme. Based on our proposition, we enact fairness watch in all of the AI pipeline stages, under robust multi-agent workflows, which will inspire new fairness research hypothesis, heuristics, and methods grounded in human-centric, systematic, interdisciplinary, socio-technical principles.

Figures

Figures reproduced from arXiv: 2506.09107 by the authors.

Figure 1
Figure 1. Indicative biases detected at the AI pipeline stages. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Generation of multiple Knowledge Graphs based on heterogeneous sources [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Agents for fair AI planning, deployment and optimization. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Agentic AI fairness framework : architecture overview [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 46 canonical work pages

  1. [1]

    S. F. Brosnan and F. B. De Waal. Evolution of responses to (un) fairness. Science, 346(6207), 2014

    work page 2014

  2. [2]

    Digitized trust in human-in-the-loop health research

    Andrew Sutton, Reza Samavi, Thomas E Doyle, and David Koff. Digitized trust in human-in-the-loop health research. In 2018 16th Annual Conference on Privacy, Security and Trust (PST), pages 1–10. IEEE, 2018

    work page 2018

  3. [3]

    Uncovering bias in personal informatics

    Sofia Yfantidou, Pavlos Sermpezis, Athena Vakali, and Ricardo Baeza-Yates. Uncovering bias in personal informatics. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 7(3):1–30, 2023

    work page 2023

  4. [4]

    Trust in ai-assisted decision making: Perspectives from those behind the system and those for whom the decision is made

    Oleksandra Vereschak, Fatemeh Alizadeh, Gilles Bailly, and Baptiste Caramiaux. Trust in ai-assisted decision making: Perspectives from those behind the system and those for whom the decision is made. In Proceedings of the CHI Conference on Human Factors in Computing Systems, pages 1–14, 2024

    work page 2024

  5. [5]

    Fairness and bias in multimodal ai: A survey

    Tosin Adewumi, Lama Alkhaled, Namrata Gurung, Goya van Boven, and Irene Pagliai. Fairness and bias in multimodal ai: A survey. arXiv preprint arXiv:2406.19097, 2024

    arXiv 2024

  6. [6]

    Caton and C

    S. Caton and C. Haas. Fairness in machine learning: A survey. ACM Computing Surveys, 2024

    work page 2024

  7. [7]

    Fairness and bias in artificial intelligence: A brief survey of sources, impacts, and mitigation strategies

    Emilio Ferrara. Fairness and bias in artificial intelligence: A brief survey of sources, impacts, and mitigation strategies. Sci, 6(1):3, 2023

    work page 2023

  8. [8]

    Bias mitigation for machine learning classifiers: A comprehensive survey

    Max Hort, Zhenpeng Chen, Jie M Zhang, Mark Harman, and Federica Sarro. Bias mitigation for machine learning classifiers: A comprehensive survey. ACM Journal on Responsible Computing, 1(2):1–52, 2024

    work page 2024

Show all 57 references
  1. [9]

    A survey on bias and fairness in machine learning

    Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM computing surveys (CSUR), 54(6):1–35, 2021

    2021

  2. [10]

    Fair enough: Searching for sufficient measures of fairness

    Suvodeep Majumder, Joymallya Chakraborty, Gina R Bai, Kathryn T Stolee, and Tim Menzies. Fair enough: Searching for sufficient measures of fairness. ACM Transactions on Software Engineering and Methodology, 32(6):1–22, 2023

    2023

  3. [11]

    The possibility of fairness: Revisiting the impossibility theorem in practice

    Andrew Bell, Lucius Bynum, Nazarii Drushchak, Tetiana Zakharchenko, Lucas Rosenblatt, and Julia Stoyanovich. The possibility of fairness: Revisiting the impossibility theorem in practice. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, ...

    2023

  4. [12]

    Fair enough? a map of the current limitations of the requirements to have fair algorithms

    Daniele Regoli, Alessandro Castelnovo, Nicole Inverardi, Gabriele Nanino, and Ilaria Penco. Fair enough? a map of the current limitations of the requirements to have fair algorithms. arXiv preprint arXiv:2311.12435, 2023

    2023 arXiv

  5. [13]

    The (im) possibility of fairness: Different value systems require different mechanisms for fair decision making.Communications of the ACM, 64(4):136–143, 2021

    Sorelle A Friedler, Carlos Scheidegger, and Suresh Venkatasubramanian. The (im) possibility of fairness: Different value systems require different mechanisms for fair decision making.Communications of the ACM, 64(4):136–143, 2021

    2021

  6. [14]

    Humans’ bias blind spot and its societal significance

    Emily Pronin and Lori Hazel. Humans’ bias blind spot and its societal significance. Current Directions in Psychological Science, 32(5):402–409, 2023

    2023

  7. [15]

    Cognitive biases that are warping your perception of reality, 2021

    World Economic Forum. Cognitive biases that are warping your perception of reality, 2021. World Economic Forum report

    2021

  8. [16]

    Judgment under uncertainty: Heuristics and biases

    Amos Tversky and Daniel Kahneman. Judgment under uncertainty: Heuristics and biases. 1990

    1990

  9. [17]

    Responsible artificial intelligence—from principles to practice: A keynote at thewebconf 2022

    Virginia Dignum. Responsible artificial intelligence—from principles to practice: A keynote at thewebconf 2022. In ACM SIGIR Forum, volume 56, pages 1–6. ACM New York, NY , USA, 2023

    2022

  10. [18]

    Overcoming the pitfalls and perils of algorithms: A classification of machine learning biases and mitigation methods

    Benjamin Van Giffen, Dennis Herhausen, and Tobias Fahse. Overcoming the pitfalls and perils of algorithms: A classification of machine learning biases and mitigation methods. Journal of Business Research, 144(6):93–106, 2022

    2022

  11. [19]

    Towards a standard for identifying and managing bias in artificial intelligence, volume 3

    Reva Schwartz, Reva Schwartz, Apostol Vassilev, Kristen Greene, Lori Perine, Andrew Burt, and Patrick Hall. Towards a standard for identifying and managing bias in artificial intelligence, volume 3. US Department of Commerce, National Institute of Standards and Technology . . . , 2022

    2022

  12. [20]

    A collaborative, human-centred taxonomy of ai, algorithmic, and automation harms

    Gavin Abercrombie, Djalel Benbouzid, Paolo Giudici, Delaram Golpayegani, Julio Hernandez, Pierre Noro, Harshvardhan Pandit, Eva Paraschou, Charlie Pownall, Jyoti Prajapati, et al. A collaborative, human-centred taxonomy of ai, algorithmic, and automation harms. arXiv preprint ...

    2024 arXiv

  13. [21]

    Erc science journalism initiative, 2024

    European Research Council. Erc science journalism initiative, 2024. https://erc.europa.eu/sites/ default/files/content/pages/pdf/ERC-Science-Journalism-Initiative.pdf

    2024

  14. [22]

    Retrieval-augmented generation for large language models: A survey

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yixin Dai, Jiawei Sun, Haofen Wang, and Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2:1, 2023

    2023 arXiv

  15. [23]

    Generative ai service implementation using llm application architecture: based on rag model and langchain framework

    Cheonsu Jeong. Generative ai service implementation using llm application architecture: based on rag model and langchain framework. Journal of Intelligence and Information Systems, 29(4):129–164, 2023

    2023

  16. [24]

    A fine-tuning enhanced rag system with quantized influence measure as ai judge

    Keshav Rangan and Yiqiao Yin. A fine-tuning enhanced rag system with quantized influence measure as ai judge. Scientific Reports, 14(1):27446, 2024

    2024

  17. [25]

    Fairerml: An extensible platform for analysing, visualising, and mitigating biases in machine learning [application notes]

    Bo Yuan, Shenhao Gui, Qingquan Zhang, Ziqi Wang, Junyi Wen, Bifei Mao, Jialin Liu, and Xin Yao. Fairerml: An extensible platform for analysing, visualising, and mitigating biases in machine learning [application notes]. IEEE Computational Intelligence Magazine, 19(2):129–141, 2024

    2024

  18. [26]

    Knowledge graphs: Opportunities and challenges

    Ciyuan Peng, Feng Xia, Mehdi Naseriparsa, and Francesco Osborne. Knowledge graphs: Opportunities and challenges. Artificial Intelligence Review, 56(11):13071–13102, 2023

    2023

  19. [27]

    From local to global: A graph rag approach to query-focused summarization

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. From local to global: A graph rag approach to query-focused summarization. arXiv preprint arXiv:2404.16130, 2024

    2024 arXiv

  20. [28]

    Comprehensible artificial intelligence on knowledge graphs: A survey

    Simon Schramm, Christoph Wehner, and Ute Schmid. Comprehensible artificial intelligence on knowledge graphs: A survey. Journal of Web Semantics, 79:100806, 2023

    2023

  21. [29]

    Practices for governing agentic ai systems

    Yonadav Shavit, Sandhini Agarwal, Miles Brundage, Steven Adler, Cullen O’Keefe, Rosie Campbell, Teddy Lee, Pamela Mishkin, Tyna Eloundou, Alan Hickey, et al. Practices for governing agentic ai systems. Research Paper, OpenAI, 2023

    2023

  22. [30]

    Recursive introspection: Teaching language model agents how to self-improve, 2024

    Yuxiao Qu, Tianjun Zhang, Naman Garg, and Aviral Kumar. Recursive introspection: Teaching language model agents how to self-improve, 2024

    2024

  23. [31]

    Harms from increasingly agentic algorithmic systems

    Alan Chan, Rebecca Salganik, Alva Markelius, Chris Pang, Nitarshan Rajkumar, Dmitrii Krasheninnikov, Lauro Langosco, Zhonghao He, Yawen Duan, Micah Carroll, Michelle Lin, Alex Mayhew, Katherine Collins, Maryam Molamohammadi, John Burden, Wanru Zhao, Shalaleh Rismani, Konstanti...

    2023

  24. [32]

    Fundamental limitations of generative llms

    Andrei Kucharavy. Fundamental limitations of generative llms. In Large Language Models in Cybersecurity: Threats, Exposure and Mitigation, pages 55–64. Springer Nature Switzerland Cham, 2024

    2024

  25. [33]

    Artificial intelligence, trust, and perceptions of agency

    Bart S Vanneste and Phanish Puranam. Artificial intelligence, trust, and perceptions of agency. Academy of Management Review, (ja):amr–2022, 2024

    2022

  26. [34]

    Fft: Towards harmlessness evaluation and analysis for llms with factuality, fairness, toxicity, 2024

    Shiyao Cui, Zhenyu Zhang, Yilong Chen, Wenyuan Zhang, Tianyun Liu, Siqi Wang, and Tingwen Liu. Fft: Towards harmlessness evaluation and analysis for llms with factuality, fairness, toxicity, 2024

    2024

  27. [35]

    Gender fairness in llms using prompt engineering and in-context learning

    Dwivedi S Dwivedi S, Ghosh S. Gender fairness in llms using prompt engineering and in-context learning. Rupkatha Journal on Interdisciplinary Studies in Humanities, 2023

    2023

  28. [36]

    From ’showgirls’ to ’performers’: Fine-tuning with gender-inclusive language for bias reduction in llms, 2024

    Marion Bartl and Susan Leavy. From ’showgirls’ to ’performers’: Fine-tuning with gender-inclusive language for bias reduction in llms, 2024

    2024

  29. [37]

    Air: Adaptive incremental embedding updating for dynamic knowledge graphs

    Zhifeng Jia, Haoyang Li, and Lei Chen. Air: Adaptive incremental embedding updating for dynamic knowledge graphs. In International Conference on Database Systems for Advanced Applications, pages 606–621. Springer, 2023

    2023

  30. [38]

    Knowledge graph embedding: A survey from the perspective of representation spaces

    Jiahang Cao, Jinyuan Fang, Zaiqiao Meng, and Shangsong Liang. Knowledge graph embedding: A survey from the perspective of representation spaces. ACM Computing Surveys, 56(6):1–42, 2024

    2024

  31. [39]

    Making large language models perform better in knowledge graph completion

    Yichi Zhang, Zhuo Chen, Lingbing Guo, Yajing Xu, Wen Zhang, and Huajun Chen. Making large language models perform better in knowledge graph completion. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 233–242, 2024

    2024

  32. [40]

    Glam: Fine- tuning large language models for domain knowledge graph alignment via neighborhood partitioning and generative subgraph encoding

    Stefan Dernbach, Khushbu Agarwal, Alejandro Zuniga, Michael Henry, and Sutanay Choudhury. Glam: Fine- tuning large language models for domain knowledge graph alignment via neighborhood partitioning and generative subgraph encoding. In Proceedings of the AAAI Symposium Series, ...

    2024

  33. [41]

    A comprehensive survey on automatic knowledge graph construction

    Lingfeng Zhong, Jia Wu, Qian Li, Hao Peng, and Xindong Wu. A comprehensive survey on automatic knowledge graph construction. ACM Computing Surveys, 56(4):1–62, 2023. 10

    2023

  34. [42]

    Fairrag: Fair human generation via fair retrieval augmentation

    Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li, Yusheng Xie, and Siqi Deng. Fairrag: Fair human generation via fair retrieval augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11996–12005, 2024

    2024

  35. [43]

    The eu artificial intelligence act “up-to-date developments and analyses of the eu ai act, 2024

    European Commission. The eu artificial intelligence act “up-to-date developments and analyses of the eu ai act, 2024

    2024

  36. [44]

    Reflexion: Language agents with verbal reinforcement learning

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36:8634–8652, 2023

    2023

  37. [45]

    Fairness in multi-agent ai: A unified framework for ethical and equitable autonomous systems

    Rajesh Ranjan, Shailja Gupta, and Surya Narayan Singh. Fairness in multi-agent ai: A unified framework for ethical and equitable autonomous systems. arXiv preprint arXiv:2502.07254, 2025

    2025 arXiv

  38. [46]

    W. Zhang. Ai fairness in practice: Paradigm, challenges, and prospects. Ai Magazine, 45(3):386–395, 2024

    2024

  39. [47]

    Igniting language intelligence: The hitchhiker’s guide from chain-of-thought reasoning to language agents

    Zhuosheng Zhang, Yao Yao, Aston Zhang, Xiangru Tang, Xinbei Ma, Zhiwei He, Yiming Wang, Mark Gerstein, Rui Wang, Gongshen Liu, et al. Igniting language intelligence: The hitchhiker’s guide from chain-of-thought reasoning to language agents. ACM Computing Surveys, 57(8):1–39, 2025

    2025

  40. [48]

    Agent-pro: Learning to evolve via policy-level reflection and optimization

    Wenqi Zhang, Ke Tang, Hai Wu, Mengna Wang, Yongliang Shen, Guiyang Hou, Zeqi Tan, Peng Li, Yueting Zhuang, and Weiming Lu. Agent-pro: Learning to evolve via policy-level reflection and optimization. arXiv preprint arXiv:2402.17574, 2024

    2024 arXiv

  41. [49]

    Learning to communicate using a communication critic and counterfactual reasoning

    Simon Vanneste, Astrid Vanneste, Kevin Mets, Tom De Schepper, Ali Anwar, Siegfried Mercelis, and Peter Hellinckx. Learning to communicate using a communication critic and counterfactual reasoning. Neural Computing and Applications, pages 1–18, 2025

    2025

  42. [50]

    Ai, meet human: Learning paradigms for hybrid decision making systems

    Clara Punzi, Roberto Pellungrini, Mattia Setzu, Fosca Giannotti, and Dino Pedreschi. Ai, meet human: Learning paradigms for hybrid decision making systems. arXiv preprint arXiv:2402.06287, 2024

    2024 arXiv

  43. [51]

    Akbar, A

    O. Akbar, A. & Conlan. Towards integrating human-in-the-loop control in proactive intelligent personalised agents. In Adjunct Proceedings of the 32nd ACM Conference on User Modeling, Adaptation and Personalization., pages 394–398, 2024

    2024

  44. [52]

    Human-in- the-loop machine learning: Reconceptualizing the role of the user in interactive approaches

    Oihane Gómez-Carmona, Diego Casado-Mansilla, Diego López-de Ipiña, and Javier García-Zubia. Human-in- the-loop machine learning: Reconceptualizing the role of the user in interactive approaches. Internet of Things, 25:101048, 2024

    2024

  45. [53]

    Symphony: Composing interactive interfaces for machine learning

    Alex Bäuerle, Ángel Alexander Cabrera, Fred Hohman, Megan Maher, David Koski, Xavier Suau, Titus Barik, and Dominik Moritz. Symphony: Composing interactive interfaces for machine learning. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems, pages 1...

    2022

  46. [54]

    Human-in-the-loop machine learning: a state of the art.Artificial Intelligence Review, 56(4):3005– 3054, 2023

    Eduardo Mosqueira-Rey, Elena Hernández-Pereira, David Alonso-Ríos, José Bobes-Bascarán, and Ángel Fernández-Leal. Human-in-the-loop machine learning: a state of the art.Artificial Intelligence Review, 56(4):3005– 3054, 2023

    2023

  47. [55]

    Superintelligent agents pose catastrophic risks: Can scientist ai offer a safer path? arXiv preprint arXiv:2502.15657, 2025

    Yoshua Bengio, Michael Cohen, Damiano Fornasiere, Joumana Ghosn, Pietro Greiner, Matt MacDermott, Sören Mindermann, Adam Oberman, Jesse Richardson, Oliver Richardson, et al. Superintelligent agents pose catastrophic risks: Can scientist ai offer a safer path? arXiv preprint ar...

    2025 arXiv

  48. [56]

    Intelligent agents in ai really can work alone

    Tom Coshow. Intelligent agents in ai really can work alone. here’s how, 2024. Gartner

    2024

  49. [57]

    The state of algorithmic fairness in mobile human-computer interaction

    Sofia Yfantidou, Marios Constantinides, Dimitris Spathis, Athena Vakali, Daniele Quercia, and Fahim Kawsar. The state of algorithmic fairness in mobile human-computer interaction. In Proceedings of the 25th International Conference on Mobile Human-Computer Interaction, pages 1...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.