Adaptive Arena-based Contestable Argumentative Network-of-Experts for Open-Ended Care Plan Coordination

Hoang-Loc Cao; Hung Cao; Phuc Ho; Phuc Truong Loc Nguyen; Ren\'e Richard; Truong Thanh Hung Nguyen

REVIEW 5 major objections 5 minor 29 references

Adaptive Arena-based Contestable Argumentative Network-of-Experts for Open-Ended Care Plan Coordination

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

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

Pith's one-line read CANOE builds care plans as contestable argument graphs that clinicians can edit, with deterministic recomputation after each change.

desk verdict Well-built framework, but the headline claim about contestability and faithfulness is not tested by the experiments. read the letter →

arxiv 2608.05391 v1 pith:CTRCRHJ3 submitted 2026-08-05 cs.AI cs.MA

Truong Thanh Hung Nguyen , Hoang-Loc Cao , Phuc Ho , Phuc Truong Loc Nguyen , René Richard , Hung Cao This is my paper · ORCID

classification cs.AIcs.MA

keywords contestableAImulti-agentsystemsargumentativecomputationopen-endeddecisioncareplancoordinationquantitativebipolarargumentationLLM-as-a-judgeclinicalsupport

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

CANOE is a framework for generating care plans that replaces a single LLM generating text in one pass with a team of role-specialized agents who argue for and against each candidate intervention. The arguments are assembled into a graph, near-tied conflicts are resolved by an arena-style adjudication step, and a deterministic propagation rule turns the graph into a final acceptability score for each intervention. Care planners can accept, reject, edit, or add arguments, and the framework recomputes the plan after every edit. On two care-planning benchmarks, the paper finds that a medically fine-tuned model gives the strongest factual consistency and safety, while a larger general-purpose model gives more coherent but less grounded text. If the framework works as claimed, every recommendation becomes traceable to surviving arguments, evidence links, and scores that a human can inspect and contest.

What carries the argument

The central machinery is the Arena-based Quantitative Bipolar Argumentation Framework (A-QBAF). For each candidate intervention, arguments from role-specialized agents carry intrinsic strengths $\tau(\alpha)$; opposing pairs whose strengths differ by less than a threshold $\delta$ are sent to an arena adjudicator that computes a win rate $w(\alpha)$ and recalibrates strengths as $\tilde{\tau}(\alpha) = \mathrm{clip}(\tau(\alpha) + \beta(2w(\alpha)-1),0,1)$; then a gradual semantics $\sigma(x) = \tilde{\tau}(x) + (1-\tilde{\tau}(x))h(E(x)) - \tilde{\tau}(x)h(-E(x))$ with $h(z) = \max\{z,0\}^2/(1+\max\{z,0\}^2)$ propagates net support-minus-attack energy $E(x)$, and the intervention's converged score becomes its acceptability $F(o_i)$. Because the propagation is deterministic, any human edit to an argument or relation yields a reproducible recomputation of the plan.

What would settle it

A concrete test: take a set of discharge cases, have clinicians independently rank the candidate interventions, run CANOE, and check whether its acceptability scores $F(o_i)$ match the clinicians' ranking; if the correlation is weak, or if removing a safety-critical argument fails to change the plan's ranking, the central claim that the scores encode clinically meaningful acceptability fails.

Watch

Extended reading notes

Core claim

CANOE's central claim is that open-ended care plan coordination can be restructured as a contestable argumentative computation. Role-based LLM agents propose supporting and attacking arguments for candidate interventions, an Arena-based Quantitative Bipolar Argumentation Framework resolves near-tied conflicts through pairwise adjudication and propagates acceptability scores through support and attack edges, and a human care planner can edit the argument graph with deterministic recomputation of the final plan. The paper asserts that this structure provides faithful explanation—each recommendation linked to the roles, evidence, and arguments that survived—and human contestability, directly addressing the transparency and safety limits of monolithic LLM pipelines. Evaluated on Discharge Me! and MedicalRAG with three backbones, the medically fine-tuned MedGemma 1.5 achieves the best AlignScore, MEDCON F1, and LLM-judge safety and validity scores, while the larger GPT-OSS leads on ROUGE-L and coherence.

Load-bearing premise

The load-bearing premise is that the LLM-generated argument strengths, arena win rates, and QBAF propagation produce acceptability scores that are clinically meaningful without any calibration against real clinical outcomes or clinician judgment.

Editorial extensions

If this is right

Every recommendation in a CANOE care plan carries a traceable chain: which roles argued for it, which evidence sources support it, and which arguments survived human contestation.
Human edits modify the formal reasoning object, and the deterministic propagation guarantees that a rejected or weakened argument changes the plan in a reproducible way.
Medically fine-tuned backbones are the safer choice for care-plan generation, since MedGemma 1.5 led on factual consistency, clinical concept coverage, and the judge's safety and validity scores.
Larger general-purpose models trade clinical grounding for coherence and completeness, so backbone selection becomes a safety-versus-style decision.
The same argumentative machinery can be lifted to other open-ended, multi-stakeholder coordination tasks where monolithic generation is considered unsafe.

Reading between the lines

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

A direct head-to-head comparison against a monolithic LLM pipeline—which the paper does not run—would test whether the argumentative structure itself is what improves safety and transparency over a single generation pass.
The acceptability scores are uncalibrated against clinician judgments or patient outcomes; a prospective study correlating $F(o_i)$ with readmission or adverse-event rates would test whether the framework's safety claims are empirical rather than formal.
The arena clash-resolution step could be ablated by replacing it with plain averaging of near-tied arguments; if final rankings do not change, the added adjudication complexity may not be earning its place.
Supporting the four human actions with direct strength overrides would test whether the propagation semantics respect qualitative clinical edits in the intended direction.

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.

Desk Editor's Note

read the letter

Colleague,

The one thing to know: CANOE is a clearly written, well-structured framework paper whose central selling point — faithful, contestable argumentative reasoning for care plans — is never actually tested. The experiments only compare three LLM backbones inside the fixed CANOE skeleton. That comparison is internally plausible, but it says nothing about whether the A-QBAF layer, the arena clash resolution, or the human-in-the-loop recomputation do any useful work.

What is new: the specific combination of complexity-gated team recruitment, arena-based QBAF with near-tied clash resolution, and editable argument graphs applied to open-ended care plan generation is new. The backbone comparison on Discharge Me! and MedicalRAG is also new. The framework description is genuinely clear — Algorithm 1 is readable, the QBAF machinery (Eqs. 10–12) is standard, and the human contestation loop is concretely specified. If I were building a contestable clinical reasoning agent, this would be a useful roadmap.

The soft spots are load-bearing, not cosmetic. First, there is no monolithic-pipeline baseline. The abstract claims CANOE addresses the transparency and safety limits of monolithic LLM pipelines, yet no such pipeline is evaluated. Second, no ablation isolates the argumentative machinery: the differences in Tables 1–2 could come entirely from the backbone. Third, there is no test of whether the explanation trace reflects the actual computation — the final coordinator is an LLM that consumes scores but could in principle ignore them, so the 'faithful explanation' attached at line 32 is not verified. Fourth, parameters δ and β in Algorithm 1 are never given values, so the supposedly deterministic recompute is not reproducible as reported. Fifth, there is no human evaluation of contestability in use; the LLM-as-a-judge protocol scores the output plan, not the argumentation process. The A-QBAF core is self-cited from the authors' own prior work; that alone is not a flaw, but it does mean the novelty lies in the application, not the formalism.

None of this is fatal to the framework as a research direction. The central claim just is not supported by the current evidence. The path forward is clear: add a monolithic baseline, ablate the A-QBAF and arena modules, report δ/β, release code, and ideally run a small study with clinicians using the contestation interface.

Who this is for: people working on argumentation-based clinical decision support and contestable AI. It deserves a serious referee, but the referee should push for major revision. My recommendation: send to review, but treat this as a system description with very provisional evaluation, and require the missing experiments before acceptance.

Referee Report

5 major / 5 minor

Summary. The paper proposes CANOE, a five-module multi-agent neuro-symbolic framework for care plan coordination. Patient information P and retrieved evidence D are used to propose candidate interventions O, recruit role-specialized agents A, and generate supporting and attacking arguments whose strengths τ(α) are calibrated by an arena-based clash resolution mechanism and then propagated through a quantitative bipolar argumentation graph (A-QBAF). Human planners can accept, reject, edit, or add arguments, after which the framework deterministically recomputes option acceptability scores F^H(oi). A coordinator LLM Π then synthesizes the final care plan Y from the revised argumentative structure. The paper evaluates CANOE on Discharge Me! and MedicalRAG using three LLM backbones, reporting ROUGE-L, AlignScore, MEDCON F1, FKGL, and an LLM-as-a-judge protocol, and concludes that medically fine-tuned models give the strongest clinical correctness and safety and that CANOE's argumentative structure provides faithful explanation and human contestability.

Significance. If the central claims were validated, CANOE would be a meaningful step toward contestable, explainable multi-agent clinical decision support: the formal QBAF machinery in Section 4.3 is clearly specified, the deterministic recomputation after human edits is a concrete operationalization of contestability, and the evaluation metrics are all external to the framework, avoiding equation-level circularity. The framework is also fairly general and could be reused outside care planning. However, the significance is currently limited because the experiments only compare LLM backbones inside a fixed CANOE pipeline; they do not isolate the contribution of the argumentative machinery and do not measure explanation faithfulness or contestability in use. As a result, the paper's headline claims about transparency, safety, and the superiority of argumentative structure over monolithic pipelines are not empirically supported by the data presented.

major comments (5)

[Abstract; §§5.4–5.5] The central claim that CANOE's argumentative structure provides faithful explanation, human contestability, and addresses the limitations of monolithic LLM pipelines is not tested. All experiments vary only the LLM backbone while holding the CANOE skeleton fixed, so Tables 1 and 2 are attributable to backbone capability rather than to the A-QBAF layer, arena clash resolution, or contestation module. There is no monolithic-pipeline baseline, no ablation that disables the arena mechanism (Eqs. 7–8) or propagation (Eqs. 10–12), and no user study or simulation that exercises the accept/reject/edit/add actions of Module 4. A reader cannot determine from the reported results whether the argumentative computation changes the output at all.
[Algorithm 1, lines 31–32; §4.5] The claimed 'faithful explanation' is not verified. The coordinator Π is an LLM that receives P, D, O, ΓH, and {F^H(oi)} and generates Y; nothing in the experiments checks that the attached explanation trace (roles, evidence, survived arguments, scores) corresponds to the computation that produced each y_t. Since Π could in principle generate text from P, D, and O while ignoring the acceptability scores, the trace may be as post-hoc as the chain-of-thought rationalizations the paper criticizes in its introduction. A faithfulness metric or a direct intervention study is needed before the explanation claim can be accepted.
[Algorithm 1, lines 14–17; §4.3, Eqs. (7)–(8)] The conflict threshold δ and calibration factor β are never assigned values in the experimental setup of Section 5.2. These parameters directly control the arena-based calibration step, so the reported results are not reproducible as stated, and the 'deterministic recompute' promised for human edits is not fully specified. The paper should either state the values used for all experiments or provide a sensitivity analysis showing that the conclusions are robust across a range of δ and β.
[Tables 1 and 2] All reported metrics are point estimates without error bars, confidence intervals, significance tests, or per-example variance, and the dataset preprocessing does not state the number of test examples used. Claims such as MedGemma 1.5 providing the 'strongest clinical correctness and safety' rest on small differences (e.g., AlignScore 0.451 vs. 0.403 on Discharge Me!; judge average 3.85 vs. 3.75) that could easily change with a different sample. Without a measure of variability, the comparisons between backbones are not statistically grounded.
[§5.1; §5.2] The evaluation benchmarks only weakly instantiate the claimed multidisciplinary, open-ended care-planning setting. The Discharge Me! adaptation uses discharge-instruction sections as ground truth, which is related to but not identical to care plan coordination across multiple disciplines. The MedicalRAG experiments filter to the 'Medical Practitioners' role only, so they do not exercise the role-diverse argumentative comparisons that are central to the framework's motivation. A small-scale clinician or care-coordinator evaluation, or at least a benchmark that includes multiple professional perspectives, would be needed to substantiate the multidisciplinary claims.

minor comments (5)

[§5.1] There is a typo in 'Brief Hospitalßl' (should be 'Brief Hospital Course').
[§5.2] The description 'all models are evaluated in a zero-shot prompted setting' is ambiguous about which LLM calls use the backbone: does each of the five modules (complexity assessment, option proposal, argument generation, strength scoring, adjudication, coordination) run the same backbone, or is the backbone used only for the final coordinator Π? Please clarify the exact model assignment per module.
[Template 1, §5.5] The paper states that full scoring rubrics 'will be released with the code upon acceptance'; for reproducibility, the rubrics and the judge model version should be included in the paper itself or in an available supplement.
[References] Reference [8] is cited as the source of the MedicalRAG dataset, but the reference title is 'Controlnet: A firewall for rag-based llm system', which appears to be a different work. Please correct the citation for the MedicalRAG dataset.
[§4.3, Eq. (11)] The impact function h(z) in Eq. (12) is nonnegative by construction, so the term h(−Ei(x)) is nonzero only when the net energy is negative. A brief explanation of how this implements bidirectional attack/support would help readers not already familiar with this specific QBAF semantics.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the argumentative scoring chain is fully specified in the paper and the reported metrics are external; the untested transparency and contestability claims are an evaluation gap, not a circular derivation.

full rationale

The core computation is defined in Sections 4.3-4.5. Argument strength tau(alpha) is an LLM-generated input; arena clash resolution (Eqs. 7-8) modifies tau using pairwise adjudicator win rates; propagation (Eqs. 10-12) deterministically produces F(oi); and the final plan is synthesized by coordinator Pi from P, D, O, Gamma_H, and the scores. None of these steps fits a parameter to the evaluation targets (ROUGE-L, AlignScore, MEDCON F1, FKGL, or the LLM judge), so no reported metric reduces by construction to an input or fitted value. The paper cites self-authored prior work [4,15] for the Arena-based QBAF mechanism, but Section 4.3 restates the full equations rather than importing an unverified black-box result, so the self-citation is not load-bearing in a circular way. The central claims that CANOE provides 'faithful explanation and human contestability' are not adequately tested by the backbone-only comparison in Sections 5.4-5.5, and the unreported values of delta and beta prevent full reproducibility; however, these are evidence and validation gaps, not equation-level circularity. Under the hard rule that circularity requires a quoted reduction of a prediction to its input, no such step is present.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The framework's behavior rests on two unreported hyperparameters (delta, beta), a hand-chosen neutral prior, and several domain assumptions about LLM argument scoring and QBAF semantics that are taken from prior work without independent validation in this paper. No new physical or ontological entities are introduced, so the invented_entities list is empty.

free parameters (3)

conflict threshold delta
Appears in Eq. 7 and Algorithm 1 line 14; controls which opposing argument pairs are sent to arena clash resolution. Value not reported in the paper.
calibration factor beta
Appears in Eq. 8 and Algorithm 1 line 15; scales the arena win-rate adjustment to intrinsic strengths. Value not reported in the paper.
intervention node prior = 0.5
Algorithm 1 line 20 sets tau_tilde_i(o_i) = 0.5; a hand-chosen neutral prior that anchors the propagated acceptability score for every intervention.

assumptions (5)

standard math QBAF gradual semantics (Eqs. 10-12) from Baroni et al. yield valid acceptability scores for clinical decision support.
Invoked in Algorithm 1 lines 20-22 and Section 4.3; the paper relies on the prior framework's formal properties without re-deriving them.
domain assumption LLM-assigned intrinsic strengths tau(alpha) measure clinically meaningful argument quality.
Section 4.3 states tau(alpha) is assigned by a scoring model evaluating clinical appropriateness, factual consistency, and feasibility, but no human validation of these scores is provided.
domain assumption The arena win rate w(alpha) from LLM adjudication identifies the clinically stronger argument in near-tied pairs.
Eq. 7-8 assume adjudication outcomes reflect clinical merit; no evidence is given beyond the authors' prior papers [4,15].
domain assumption Complexity classifier C and recruitment operator S(.) improve plan quality.
Sections 4.1-4.2 claim these modules preserve relevance and reduce overhead, but no ablation tests their effect on outcomes.
domain assumption Synthesis planner Pi converts acceptability scores into clinically safe and coherent plans.
Section 4.5 defines the final plan assembly from F(oi), but the mapping from scores to plan text is not validated against clinician-built plans or safety outcomes.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Adaptive Arena-based Contestable Argumentative Network-of-Experts for Open-Ended Care Plan Coordination." pith.science (2026). https://pith.science/paper/CTRCRHJ3

@misc{pith2026260805391,
  author       = {Pith},
  title        = {Pith review of: Adaptive Arena-based Contestable Argumentative Network-of-Experts for Open-Ended Care Plan Coordination},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CTRCRHJ3}},
  note         = {Machine review of arXiv:2608.05391}
}

read the original abstract

Care plan coordination demands synthesizing heterogeneous clinical, functional, and psychosocial information across multiple professional disciplines, where monolithic LLM pipelines cannot perform in a transparent or safe manner. We present CANOE (Contestable Argumentative Network-of-Experts), a multi-agent neuro-symbolic framework that addresses these limitations through five modules: complexity assessment, adaptive team recruitment, role-based argumentative computation via an Arena-based Quantitative Bipolar Argumentation Framework (A-QBAF), human-in-the-loop contestation, and care-plan synthesis. Role-specialized agents generate supporting and attacking arguments for candidate interventions; conflicts are resolved through arena-based clash resolution before acceptability scores propagate across the argumentation graph. Care planners may accept, reject, edit, or add arguments, and the framework will deterministically recompute the final plan. Evaluation on Discharge Me! and MedicalRAG using ROUGE-L, AlignScore, MEDCON F1, FKGL, and LLM-as-a-judge shows that medically fine-tuned models achieve the strongest clinical correctness and safety, while CANOE's argumentative structure provides faithful explanation and human contestability.

Figures

Figures reproduced from arXiv: 2608.05391 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 22 canonical work pages

[1]

arXiv preprint arXiv:2508.10925 (2025)

Agarwal, S., et al.: gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925 (2025)

arXiv 2025
[2]

Minds and Machines33(4), 613–639 (2023), publisher: Springer

Alfrink, K., Keller, I., Kortuem, G., Doorn, N.: Contestable AI by design: Towards a framework. Minds and Machines33(4), 613–639 (2023), publisher: Springer

2023
[3]

International Journal of Approximate Reasoning 105(2019)

Baroni, P., et al.: From fine-grained properties to broad principles for gradual argu- mentation: A principled spectrum. International Journal of Approximate Reasoning 105(2019)

work page 2019
[4]

In: The 39th Canadian Conference on Artificial Intelligence

Cao, H.L., et al.: Neuro-symbolic adaptive collaboration of arena-based argumen- tative llms for contestable legal reasoning. In: The 39th Canadian Conference on Artificial Intelligence. pp. 895–902. PMLR (2026)

work page 2026
[5]

discharge me!

Damm, H., et al.: Wispermed at “discharge me!”: Advancing text generation in healthcare with large language models, dynamic expert selection, and priming techniques on mimic-iv. In: Proceedings of the 23rd Workshop on Biomedical Natural Language Processing. pp. 105–121 (2024)

work page 2024
[6]

In: Proceedings of the AAAI Conference on Artificial Intelligence

Freedman, G., et al.: Argumentative Large Language Models for Explainable and Contestable Claim Verification. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 14930–14939 (2025), issue: 14

work page 2025
[7]

In: 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)

Hong, S., et al.: Argmed-agents: Explainable clinical decision reasoning with llm disscusion via argumentation schemes. In: 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). pp. 5486–5493. IEEE (2024)

work page 2024
[8]

arXiv preprint arXiv:2504.09593 (2025)

Hongwei, Y., et al.: Controlnet: A firewall for rag-based llm system. arXiv preprint arXiv:2504.09593 (2025)

arXiv 2025

Show all 29 references

[9]

In: ITM Web of Conferences

Huang, S.: Hallucinations of large language models in medical environments: A systematic review of risks, detection, and mitigation. In: ITM Web of Conferences. vol. 84, p. 03005. EDP Sciences (2026)

2026
[10]

The Journal for Nurse Practitioners18(8), 833–836 (2022)

Ju, H.H.: Improving care coordination of patients with chronic diseases. The Journal for Nurse Practitioners18(8), 833–836 (2022)

2022
[11]

Advances in Neural Information Processing Systems37, 79410–79452 (2024) Contestable Argumentative Network-of-Experts for Care Coordination 17

Kim, Y., et al.: Mdagents: An adaptive collaboration of llms for medical decision- making. Advances in Neural Information Processing Systems37, 79410–79452 (2024) Contestable Argumentative Network-of-Experts for Care Coordination 17

2024
[12]

In: Proceed- ings of the 21st International Conference on Principles of Knowledge Representation and Reasoning

Leofante, F., et al.: Contestable ai needs computational argumentation. In: Proceed- ings of the 21st International Conference on Principles of Knowledge Representation and Reasoning. pp. 888–896 (2024)

2024
[13]

npj Digital Medicine (2026)

Liu, Q., Hu, Z., Huang, T., Niu, Y., Zhang, X., Ma, S., Lin, C., Huat, G.K., Kwon, H.E., Gao, F., et al.: Evomdt: a self-evolving multi-agent system for structured clinical decision-making in multi-cancer. npj Digital Medicine (2026)

2026
[14]

ACM Trans

Nguyen, H., Rahimi, A., Whitford, V., Fournier, H., Kondratova, I., Richard, R., Cao, H.: Heart2mind: Human-centered contestable psychiatric disorder prediction system using wearable ecg monitors. ACM Trans. Comput. Healthcare (2026)

2026
[15]

In: Proceedings of the 2026 International Conference on Multimedia Retrieval

Nguyen, H.T.T., et al.: Contestable multi-agent debate with arena-based argu- mentative computation for multimedia verification. In: Proceedings of the 2026 International Conference on Multimedia Retrieval. pp. 2887–2891 (2026)

2026
[16]

In: Proceedings of 9th International Symposium on Chatbots and Human-centred AI (2025)

Nguyen, L.P.T., et al.: Motion2Meaning: A Clinician-Centered Framework for Contestable LLM in Parkinson’s Disease Gait Interpretation. In: Proceedings of 9th International Symposium on Chatbots and Human-centred AI (2025)

2025
[17]

In: International Workshop on Health Intelligence

Nguyen, T.T.H., et al.: Towards trust of explainable ai in thyroid nodule diagnosis. In: International Workshop on Health Intelligence. pp. 11–26. Springer (2023)

2023
[18]

Artificial intelligence in medicine107(2020)

Ploug, T., Holm, S.: The four dimensions of contestable AI diagnostics-A patient- centric approach to explainable AI. Artificial intelligence in medicine107(2020)

2020
[19]

In: Proceedings of the International Conference on Principles of Knowledge Representation and Reasoning

Rago, A., Li, H., Toni, F.: Interactive explanations by conflict resolution via argu- mentative exchanges. In: Proceedings of the International Conference on Principles of Knowledge Representation and Reasoning. vol. 19, pp. 582–592 (2023)

2023
[20]

arXiv preprint arXiv:2604.05081 (2026)

Sellergren, A., et al.: Medgemma 1.5 technical report. arXiv preprint arXiv:2604.05081 (2026)

2026 arXiv
[21]

Annals of internal medicine165(11), 753–760 (2016)

Sinsky, C., et al.: Allocation of physician time in ambulatory practice: a time and motion study in 4 specialties. Annals of internal medicine165(11), 753–760 (2016)

2016
[22]

In: Findings of the Association for Computational Linguistics: ACL 2024 (2024)

Tang, X., et al.: Medagents: Large language models as collaborators for zero-shot medical reasoning. In: Findings of the Association for Computational Linguistics: ACL 2024 (2024)

2024
[23]

Advances in neural information processing systems35, 24824–24837 (2022)

Wei, J., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems35, 24824–24837 (2022)

2022
[24]

In: Healthcare

Xiao, L., Greer, D.: Linked argumentation graphs for multidisciplinary decision support. In: Healthcare. vol. 11, p. 585. MDPI (2023)

2023
[25]

discharge me!

Xu, J., et al.: Overview of the first shared task on clinical text generation: Rrg24 and “discharge me!”. In: Proceedings of the 23rd Workshop on Biomedical Natural Language Processing. pp. 85–98 (2024)

2024
[26]

Scientific data10(1), 586 (2023)

Yim, W.w., et al.: Aci-bench: a novel ambient clinical intelligence dataset for benchmarking automatic visit note generation. Scientific data10(1), 586 (2023)

2023
[27]

In: Proceedings of the 15th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics

Yue, L., et al.: Clinicalagent: Clinical trial multi-agent system with large language model-based reasoning. In: Proceedings of the 15th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics. pp. 1–10 (2024)

2024
[28]

In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

Zha, Y., et al.: Alignscore: Evaluating factual consistency with a unified align- ment function. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 11328–11348 (2023)

2023
[29]

In: Conference on Neurosymbolic Learning and Reasoning

Zhu, Y., et al.: Argrag: Explainable retrieval augmented generation using quan- titative bipolar argumentation. In: Conference on Neurosymbolic Learning and Reasoning. pp. 697–718. PMLR (2025)

2025

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

arXiv preprint arXiv:2508.10925 (2025)

Agarwal, S., et al.: gpt-oss-120b & gpt-oss-20b model card. arXiv preprint arXiv:2508.10925 (2025)

arXiv 2025

[2] [2]

Minds and Machines33(4), 613–639 (2023), publisher: Springer

Alfrink, K., Keller, I., Kortuem, G., Doorn, N.: Contestable AI by design: Towards a framework. Minds and Machines33(4), 613–639 (2023), publisher: Springer

2023

[3] [3]

International Journal of Approximate Reasoning 105(2019)

Baroni, P., et al.: From fine-grained properties to broad principles for gradual argu- mentation: A principled spectrum. International Journal of Approximate Reasoning 105(2019)

work page 2019

[4] [4]

In: The 39th Canadian Conference on Artificial Intelligence

Cao, H.L., et al.: Neuro-symbolic adaptive collaboration of arena-based argumen- tative llms for contestable legal reasoning. In: The 39th Canadian Conference on Artificial Intelligence. pp. 895–902. PMLR (2026)

work page 2026

[5] [5]

discharge me!

Damm, H., et al.: Wispermed at “discharge me!”: Advancing text generation in healthcare with large language models, dynamic expert selection, and priming techniques on mimic-iv. In: Proceedings of the 23rd Workshop on Biomedical Natural Language Processing. pp. 105–121 (2024)

work page 2024

[6] [6]

In: Proceedings of the AAAI Conference on Artificial Intelligence

Freedman, G., et al.: Argumentative Large Language Models for Explainable and Contestable Claim Verification. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 14930–14939 (2025), issue: 14

work page 2025

[7] [7]

In: 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)

Hong, S., et al.: Argmed-agents: Explainable clinical decision reasoning with llm disscusion via argumentation schemes. In: 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). pp. 5486–5493. IEEE (2024)

work page 2024

[8] [8]

arXiv preprint arXiv:2504.09593 (2025)

Hongwei, Y., et al.: Controlnet: A firewall for rag-based llm system. arXiv preprint arXiv:2504.09593 (2025)

arXiv 2025