Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

PROCLAIM claims that structuring LLM fact-checking as a courtroom-style adversarial debate with progressive retrieval raises Check-COVID accuracy to 81.7%, 10 points above standard multi-agent debate, with the dynamic retrieval module drivi

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

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

T0 review · deepseek-v4-flash

2026-08-04 05:37 UTC pith:SX2HOT3C

load-bearing objection A genuinely integrated courtroom-style debate framework with open code and a rich audit trail, but the headline accuracy numbers rest on a non-standard scoring rule and a confounded P-RAG ablation, so treat the +10pp as not yet established. the 4 major comments →

arxiv 2603.28488 v3 pith:SX2HOT3C submitted 2026-03-30 cs.CL cs.AIcs.MA

Courtroom-Style Multi-Agent Debate with Progressive RAG and Role-Switching for Controversial Claim Verification

Masnun Nuha Chowdhury , Nusrat Jahan Beg , Umme Hunny Khan , Syed Rifat Raiyan , Md Kamrul Hasan , Hasan Mahmud This is my paper
classification cs.CL cs.AIcs.MA
keywords claim verificationmulti-agent debateprogressive retrievalretrieval-augmented generationfact-checkingCOVID-19 misinformationLLM-as-judgerole-switching
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

PROCLAIM tests a specific hypothesis: fact-checking controversial claims becomes more reliable when the LLM pipeline mirrors a courtroom, with separate counsel arguing support and refutation, a judge refining retrieval queries, and independent judges rendering verdicts. The paper reports 81.7% accuracy on the binary-label subset of Check-COVID, 10.0 percentage points above standard multi-agent debate, with the progressive retrieval module P-RAG contributing 7.5 points of that gain. It also finds that removing P-RAG raises agreement among judges while lowering accuracy—a signal of confident convergence on wrong answers—and that heterogeneous judge models correct each other's consistent biases. The point is that reliability in this design comes from deliberation structure, evidence dynamics, and model diversity rather than from any single answer.

Core claim

On its own terms, the paper claims that a structured, adversarial courtroom protocol reshapes LLM claim verification: instead of one retrieval pass and a free-form exchange, each claim is decomposed into premises, support and opposition counsel retrieve and negotiate evidence, a critic and self-reflection expose gaps, P-RAG iteratively expands the evidence pool with novelty-filtered documents, and a three-judge panel delivers a verdict. The measurable discovery is that the dynamic retrieval loop is doing the heavy lifting: ablating P-RAG costs 7.5 accuracy points and makes the judges agree more often, while ablating role-switching, the multi-judge panel, or self-reflection costs less. The pa

What carries the argument

Progressive Retrieval-Augmented Generation (P-RAG): an iterative retrieval module that, each debate round, combines the last four debate messages, the counsel's self-identified evidence gap, and the prior round's reflection-driven discovery need into a single query, sends it through a presiding judge for refinement, then admits only documents whose embedding novelty relative to the current pool is at least 0.20. It drives the primary performance gain (+7.5 pp) and prevents 'epistemic bubble' convergence by forcing genuinely new evidence to enter the debate.

Load-bearing premise

The load-bearing premise is that mapping every 'inconclusive' judge verdict to 'supported' is a fair scoring rule on a subset of claims that already has definitive binary labels; if judges abstain on a biased slice of claims, the 81.7% accuracy and the 10-point gain over standard multi-agent debate are inflated.

What would settle it

On the 120 claims used for ablations and the Check-COVID binary subset, recompute accuracy under three verdict mappings—INCONCLUSIVE counted as SUPPORT, as REFUTE, and excluded—and report per-class results. If the INCONCLUSIVE-to-SUPPORT mapping accounts for most of the 10-point gap over standard MAD, or if true REFUTE claims are over-represented among the inconclusive verdicts that get converted, the central claim fails.

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

If this is right

  • If P-RAG is removed, accuracy falls by 7.5 pp while inter-judge agreement rises from kappa 0.468 to 0.599; the authors read this as confident convergence on wrong conclusions.
  • Role-switching contributes 4.2 pp and adds 20% more evidence, indicating that the second debate pass uncovers missed evidence rather than serving only as a consistency check.
  • A heterogeneous three-judge panel beats any single judge by 3.3 pp; the individual judges show complementary error profiles that majority voting corrects.
  • Self-reflection has minimal impact on accuracy (-0.8 pp) but cuts debate rounds by 29% and token usage by 17%, acting as an economic governor.
  • Accuracy stays within roughly 9 pp across HealthVer, FEVEROUS, and Check-COVID with no domain-specific tuning, suggesting the deliberation architecture transfers.

Where Pith is reading between the lines

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

  • Our inference: because P-RAG accounts for 7.5 of the 10-point gain, a cheaper system that keeps P-RAG and drops role-switching plus the second debate might retain much of the accuracy at a fraction of the token cost; the paper does not test this configuration.
  • Our inference: the 'inconclusive becomes support' rule, combined with evaluation only on claims that have definite binary labels, means the reported 81.7% is not directly comparable to standard fact-checking accuracy and could shift materially under different verdict-mapping conventions.
  • Our inference: the reported instability of reflection trajectories on wrong answers suggests a deployable early-warning signal, but only for systems that log per-round internal reasoning; single-call models cannot generate that signal.
  • Our inference: a testable extension would apply the same progressive retrieval protocol to claim verification outside health domains, where evidence corpora are less centralized, to see whether the P-RAG gain persists when retrieval quality drops.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes PROCLAIM, a courtroom-style multi-agent framework for claim verification that combines progressive RAG, role-switching consistency tests, self-reflection, and a heterogeneous three-judge panel. On a binary-labeled subset of Check-COVID, it reports 81.7% majority-vote accuracy, a +10.0 pp gain over a standard multi-agent debate baseline, and attributes the largest share of the gain (+7.5 pp) to P-RAG. The paper also presents analyses of debate dynamics, termination behavior, token cost, and panel bias, and releases code and prompts.

Significance. If the headline comparisons were established, the paper would make a useful contribution: structured adversarial deliberation with dynamic retrieval is a plausible route to more auditable and robust claim verification. The paper is unusually transparent in shipping prompts, pseudo-code, per-run results, and an ablation matrix, and the 'epistemic bubble' observation (removing P-RAG raises inter-judge agreement while lowering accuracy) is a genuinely interesting phenomenon. However, the accuracy claims rest on a non-standard evaluation mapping and on a confounded ablation, and the strongest reported baseline is a single-call RAG model that already exceeds PROCLAIM. The core contribution is therefore not yet convincingly quantified.

major comments (4)
  1. [Section 3 / Appendix C] The evaluation metric is not standard claim-verification accuracy. The binary Check-COVID subset excludes non-definitive labels, and every INCONCLUSIVE judicial verdict is recoded as SUPPORT under the 'Burden of Refutation'. This makes 'not refuted' equivalent to 'verified'. No SUPPORT/REFUTE balance or confusion matrix for the 120-claim subset is reported, and Section 4.1 itself describes Run-2 as having class-asymmetric degradation suppressing REFUTE recall — exactly the failure mode this mapping can hide. The 81.7% figure and the +10.0 pp gain over MAD are therefore not comparable to conventional fact-checking results until standard-scoring accuracy and label-wise breakdowns are reported.
  2. [Section 4.4 / Table 6, A3] The +7.5 pp attribution to P-RAG is confounded. The w/o-P-RAG ablation (A3) simultaneously sets Max Rounds to 3 (from 10), disables Reflection->P-RAG, and disables Adaptive Convergence. Any of these changes alone could explain part or all of the accuracy drop. To substantiate the abstract's claim that 'P-RAG driv[es] the primary performance gains', an ablation must isolate the novelty-filtered progressive retrieval mechanism while keeping max rounds and stopping rules fixed, or must vary these factors independently.
  3. [Section 4.3 / Table 4] The paper's headline 'outperform[ing] standard multi-agent debate by 10.0 percentage points' is technically true, but the same table shows a single-call GPT-5-mini + RAG baseline at 85.8% accuracy — 4.1 pp above the full PROCLAIM pipeline. The claim that the courtroom architecture is a strong accuracy driver is undercut by the fact that its own simplest baseline beats it. The narrative in the abstract and Section 5 should be reconciled with this result, or additional stronger baselines and a significance test should be provided.
  4. [Section 4.1 / Table 3] Run-level variability is large and unreported in the summary statistics: Run-0 accuracy is 95.0%, Run-1 is 81.7%, and Run-2 is 79.0%. The majority-vote figure of 81.7% equals one of the three runs and can vary with stochastic seed selection. No confidence intervals or significance tests are given for the 81.7% figure or for the 10.0 pp/7.5 pp deltas. Given the 120-claim sample, these differences may not be statistically robust; the paper should report variance, significance (e.g., bootstrap or paired tests), and the number of runs used for the headline comparison.
minor comments (5)
  1. [Appendix L.1] The confidence-weight calibration (W_consensus=0.8) is performed on a held-out development subset and the ECE improvement is reported; please clarify whether the final test-set ECE of 0.034 is computed on claims used for weight selection or on fully held-out claims.
  2. [Section B.4] The novelty threshold tau=0.20 was selected via grid search on only 20 development claims. This is a very small tuning set; the sensitivity of downstream accuracy to tau should be reported.
  3. [Appendix M] The 'Chief Justice priority-weighting' edge-case rule (DeepSeek-R1 breaks a three-way split) is presented as a design decision but no sensitivity analysis is given. Since this rule can decide the final verdict, its effect on accuracy should be quantified.
  4. [Figure 4] The 'Trajectory Instability Signal' claim is supported only by representative trajectories with n=10 per panel. A quantitative measure (e.g., variance of reflection scores for correct vs. incorrect predictions) would strengthen the claim.
  5. [General] Several references are cited with future/in-preparation dates (e.g., 2026) that may not be verifiable; please update or mark them appropriately.

Circularity Check

0 steps flagged

No circular derivation; evaluation is empirical, though the non-standard Burden-of-Refutation scoring and confounded P-RAG ablation are validity concerns, not circularity.

full rationale

The paper proposes an empirical pipeline and evaluates it on Check-COVID; there is no derivation chain in which a predicted quantity is defined by its input. The closest definitional choice is the 'Burden of Refutation' mapping (Section 4 and Appendix C) that recodes INCONCLUSIVE judge verdicts as SUPPORT. This is a transparent scoring rule rather than a fitted parameter or an equation-level identity: the system still must produce REFUTE verdicts and can be wrong on both classes, so accuracy is not forced by construction. The P-RAG ablation (Table 6) is confounded — w/o P-RAG also lowers Max Rounds to 3 and disables reflection-driven retrieval — so the +7.5 pp attribution is not cleanly isolated, but this is an experimental soundness issue, not circular reasoning. Self-citations (CircuitLM, PhysicsEval) appear only in related-work lists and are not load-bearing. No uniqueness theorems or ansatz-smuggling citations are used. Thus no circular step meets the evidence bar.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 3 invented entities

The central accuracy claims rest on several hand-set/fitted hyperparameters (novelty threshold, admissibility thresholds, stopping criteria, confidence adjustments) and on domain assumptions about label quality, embedding similarity, and the burden-of-refutation mapping. The invented entities are interpretive constructs rather than independently validated scientific objects.

free parameters (7)
  • P-RAG novelty threshold τ = 0.20
    Selected by grid search over {0.10, 0.15, 0.20} on a 20-claim held-out development subset (Appendix B.4); controls which retrieved documents enter the evidence pool and directly shapes the P-RAG contribution claim.
  • Consensus weight W_consensus = 0.8
    Chosen via 5-fold cross-validation grid search over [0.5, 0.9] to minimize ECE on a held-out dev subset (Appendix L.1); affects confidence scores, not verdicts.
  • Admissibility thresholds = admit >0.5; disputed 0.1–0.5; discard ≤0.1
    Hand-set product rule w = relevance × credibility (Section 2.3, Appendix G.2); determines the initial evidence pool before debate.
  • Self-reflection weights = logic 0.4, novelty 0.3, rebuttal 0.3
    Hand-chosen weights in Eq. 2; affect confidence adjustment, termination signals, and the 'logic lie detector' interpretation.
  • Self-reflection adjustment multiplier and floor = multiplier 0.6; floor -0.15
    Hand-set in Eq. 3 and Appendix L.3 to bound agent influence on confidence; not accuracy, but part of the reported framework.
  • Role-switch consistency mapping thresholds = +0.10 if γ≥7; 0.0 if 5≤γ<7; -0.05 if γ<5
    Hand-set piecewise mapping (Appendix L.2) converts consistency score to a confidence adjustment.
  • Stopping criteria thresholds = redundancy ratio 0.70; relevance gain 0.05; iteration cap 10; plateau 0.05
    Hand-set in Table 1 and Section 2.5; determine debate length and termination, affecting accuracy and token cost.
axioms (6)
  • domain assumption Ground-truth labels in the binary Check-COVID subset are correct and the subset is representative of controversial claim verification.
    The evaluation and all accuracy claims assume the selected 120 claims with definitive labels are a fair test set; no human re-verification or label distribution is reported.
  • ad hoc to paper INCONCLUSIVE judicial verdicts should be counted as SUPPORT under the 'Burden of Refutation' standard.
    This is a paper-specific evaluation convention (Appendix C); if the dataset labels were created without this legal presumption, the mapping can systematically inflate accuracy.
  • domain assumption Cosine similarity in all-MiniLM-L6-v2 embedding space is a valid measure of evidential novelty and relevance for biomedical abstracts.
    P-RAG's novelty filter and retrieval rely on this embedding similarity; no domain-specific retrieval evaluation is provided.
  • domain assumption Majority voting across three heterogeneous LLM judges corrects complementary errors rather than compounding bias.
    Section 4.5 infers this from correlational judge-bias profiles; there is no intervention demonstrating error correction.
  • ad hoc to paper Reflection scores and their trajectories reflect argument quality and can serve as a lie detector.
    The trajectory-instability signal is defined and observed on the same runs used to develop the framework; no out-of-sample validation is provided.
  • standard math Standard statistical measures (Cohen's κ, Fleiss' κ, ECE) are appropriate for these categorical judgments.
    Used in Tables 3, 9, 10; standard but depends on label prevalence and judge-agreement assumptions.
invented entities (3)
  • Trajectory Instability Signal ('Logic Lie Detector') no independent evidence
    purpose: Diagnostic that flags incorrect predictions from oscillating self-reflection scores.
    Defined post hoc on the same evaluation runs; no pre-registered threshold, ROC analysis, or independent dataset; Figure 4 shows n=10 per panel.
  • Epistemic bubble effect no independent evidence
    purpose: Interpretive label for higher inter-judge agreement with lower accuracy when P-RAG is removed.
    Single ablation with confounded changes (max rounds 3, no reflection queries); not independently verified.
  • Chief Justice priority-weighting no independent evidence
    purpose: Breaks three-way judge deadlocks by deferring to DeepSeek-R1.
    Hand-designed rule (Appendix M) with no empirical comparison to other deadlock-breaking methods.

pith-pipeline@v1.3.0-alltime-deepseek · 26601 in / 16125 out tokens · 162860 ms · 2026-08-04T05:37:13.014565+00:00 · methodology

0 comments
read the original abstract

Large language models (LLMs) remain unreliable for high-stakes claim verification due to hallucinations and shallow reasoning. While retrieval-augmented generation (RAG) and multi-agent debate (MAD) address this, they are limited by one-pass retrieval and unstructured debate dynamics. We propose a courtroom-style multi-agent framework, PROClaim, that reformulates verification as a structured, adversarial deliberation. Our approach integrates specialized roles (e.g., Plaintiff, Defense, Judge) with Progressive RAG (P-RAG) to dynamically expand and refine the evidence pool during the debate. Furthermore, we employ evidence negotiation, self-reflection, and heterogeneous multi-judge aggregation to enforce calibration, robustness, and diversity. In zero-shot evaluations on the Check-COVID benchmark, PROClaim achieves 81.7% accuracy, outperforming standard multi-agent debate by 10.0 percentage points, with P-RAG driving the primary performance gains (+7.5 pp). We ultimately demonstrate that structural deliberation and model heterogeneity effectively mitigate systematic biases, providing a robust foundation for reliable claim verification. Our code and data are publicly available at https://github.com/mnc13/PROClaim.

Figures

Figures reproduced from arXiv: 2603.28488 by Hasan Mahmud, Masnun Nuha Chowdhury, Md Kamrul Hasan, Nusrat Jahan Beg, Syed Rifat Raiyan, Umme Hunny Khan.

Figure 1
Figure 1. Figure 1: Overview of the pipeline Before retrieval, the raw claim is decom￾posed into atomic, independently testable premises (Hu et al., 2025a; Lawrence & Reed, 2017). This serves two purposes: first, de￾composing complex claims allows the re￾trieval system to cast a wider and more tar￾geted net; second, the resulting premises act as an explicit checklist for scoring ar￾gument completeness during self-reflection a… view at source ↗
Figure 2
Figure 2. Figure 2: Termination distribution and con￾vergence speed across 360 debate instances. Dataset. To evaluate the framework’s capac￾ity for adversarial resolution, we focus on the subset of the Check-COVID (Wang et al., 2023a) test set possessing definitive binary ground-truths (SUPPORT or REFUTE). This task formulation, which we term Adversar￾ial Resolution of Hard-Binary Claims, ensures that the system is tested on … view at source ↗
Figure 3
Figure 3. Figure 3: P-RAG evidence novelty across debate rounds Representative reflection score trajectories (n 10 per panelall runs) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Reflection score trajectories across plateau, judicial, and critic resolution patterns [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cost–accuracy Pareto front across system configurations. [PITH_FULL_IMAGE:figures/full_fig_p033_5.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cost–accuracy Pareto front across system configurations. [PITH_FULL_IMAGE:figures/full_fig_p032_5.png] view at source ↗

Review history (2 revisions) →

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Automatic Hard Example Synthesis with Multi-Level Agentic Data Curation

    cs.AI 2026-07 conditional novelty 5.0

    Test-time retrieval of committee-disagreement-mined synthetic images cuts a safety classifier's false-negative rate on a hard HoliSafe subset from 41.2% to 24.5%.

Reference graph

Works this paper leans on

24 extracted references · 4 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Relevance: How directly does this evidence address the premises of the claim? (0.0 - 1.0)

  2. [2]

    Shuzhi Gong, Richard O Sinnott, Jianzhong Qi, Cecile Paris, Preslav Nakov, and Zhuohan Xie

    URLhttps://arxiv.org/abs/2505.17762. Shuzhi Gong, Richard O Sinnott, Jianzhong Qi, Cecile Paris, Preslav Nakov, and Zhuohan Xie. Multi-sourced, multi-agent evidence retrieval for fact-checking.arXiv preprint arXiv:2603.00267, 2026. Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, S...

    Pith/arXiv arXiv 2026

  3. [3]

    unresolved

    Rebuttal Coverage: Did they address the opponent’s strongest points? Identify any premises that remain "unresolved" or under-supported. Provide actionable recommendations for both sides to improve their discovery and arguments. Respond ONLY in valid JSON format: { "plaintiff": { "logic": 0.0, "evidence": 0.0, "rebuttal": 0.0, "reasoning": "..." }, "defens...

  4. [4]

    23 Preprint

    Clinical studies on hospitalized COVID-19 patients do not report significant occur- rences of heart muscle cell damage. 23 Preprint. Under review

  5. [5]

    There is no statistical association between COVID-19 infection severity and markers of heart muscle cell damage in hospitalized patients

  6. [6]

    Autopsy findings of deceased hospitalized COVID-19 patients do not show evidence of heart muscle cell damage

  7. [7]

    thinking

    Credibility: Does the evidence come from a reliable scientific context or contain high-quality data? (0.0 - 1.0) G.3 Plaintiff Counsel Prompt Agent:GPT-5-mini System Prompt: You are the Plaintiff Counsel in a legal proceeding. Your role is to present arguments supporting the claim, interpret evidence favorably, challenge opposing arguments, and conduct ex...

  8. [8]

    Logical Coherence: Argument flow and structure

  9. [9]

    Evidence Coverage: How well they used admitted exhibits

  10. [11]

    Logical Coherence: Evaluate the flow and structural integrity of your arguments

  11. [12]

    Evidence Novelty: Have you introduced truly new information or just repeated old points?

  12. [13]

    scores": { 20 Preprint. Under review

    Rebuttal Coverage: How effectively did you address the{opp side} counsel’s latest points? Identify: - Critical gaps in your current evidence base. - Premises you haven’t sufficiently supported. Respond ONLY in valid JSON format: { "scores": { 20 Preprint. Under review. "logic": 0.0-1.0, "novelty": 0.0-1.0, "rebuttal": 0.0-1.0 }, "flaws_identified": ["..."...

  13. [14]

    Hospitalized COVID-19 patients have detectable levels of cardiac biomarkers in- dicative of heart muscle cell damage

  14. [15]

    The prevalence of elevated cardiac biomarkers in hospitalized COVID-19 patients is comparable to a control group without COVID-19

  15. [16]

    Incidence rates of heart muscle cell damage in hospitalized COVID-19 patients are not higher than in patients with other viral respiratory infections

  16. [20]

    not an associated condition

    Hospitalized COVID-19 patients with pre-existing cardiac conditions do not have higher rates of heart muscle cell damage compared to those without pre-existing conditions. Evidence Negotiation & Admission Initial RAG retrieved 5 candidate documents; negotiation and arbitration admitted21 ex- hibits(weights ranging 0.54–0.81), including: Source 32517963 (J...

    2022

  17. [21]

    Role-Play Consistency (0–10) During the role-switching consistency test (Section 2.7), an independent consistency ana- lyzer evaluates whether an agent successfully argues the opposing position using identical evidence without logically contradicting its prior arguments. The score reflects adherence to the persona constraints on a 10-point scale; lower sc...

  18. [22]

    I concede,

    Concession Rate We programmatically track explicit linguistic markers of concession and conversational yielding (e.g.,“I concede,” “you make a good point,” “I partially agree”) within the counsel transcripts. To normalize for varying debate lengths, the metric is reported as the frequency of such triggers per 1,000 generated words. A near-zero rate indica...

  19. [23]

    The early-stopping criterion conservatively halts the debate if ∆S< 0.05 (stagnation)

    Reflection Plateau (∆S) It is computed as the average absolute change in the cumulative self-reflection score (Stotal) between consecutive debate rounds: ∆S=|S (t) total −S (t−1) total | For a given round, the maximum possible change is ∼ 1.0 (depending on reflection ad- justments). The early-stopping criterion conservatively halts the debate if ∆S< 0.05 ...

  20. [24]

    rubber- stamping

    Judicial Conformity (Fleiss’κ) To measure whether the three structurally heterogeneous LLM judges exhibit “rubber- stamping” or independent evaluation, we calculate Fleiss’ Kappa (κ) over their final verdicts (SUPPORTED, NOT SUPPORTED, INCONCLUSIVE). A κ≈ 0.4513 indicates moderate, au- thentic agreement. While confirming they reach consensus on clear-cut ...

  21. [297]

    Zhihong Shao, Yeyun Gong, Yelong Shen, Minlie Huang, Nan Duan, and Weizhu Chen

    URLhttps://aclanthology.org/2021.findings-emnlp.297/. Zhihong Shao, Yeyun Gong, Yelong Shen, Minlie Huang, Nan Duan, and Weizhu Chen. Enhancing retrieval-augmented large language models with iterative retrieval-generation synergy. In Houda Bouamor, Juan Pino, and Kalika Bali (eds.),Findings of the Association for Computational Linguistics: EMNLP 2023, pp....

    Pith/arXiv arXiv 2021

  22. [2021]

    doi: 10.18653/v1/2021.findings-emnlp

    Association for Computational Linguistics. doi: 10.18653/v1/2021.findings-emnlp

    doi:10.18653/v1/2021.findings-emnlp 2021

  23. [2023]

    innocent until proven guilty

    URLhttps://arxiv.org/abs/2212.10509. Pat Verga, Sebastian Hofstatter, Sophia Althammer, Yixuan Su, Aleksandra Piktus, Arkady Arkhangorodsky, Minjie Xu, Naomi White, and Patrick Lewis. Replacing judges with juries: Evaluating llm generations with a panel of diverse models, 2024. URL https: //arxiv.org/abs/2404.18796. Gengyu Wang, Kate Harwood, Lawrence Chi...

    Pith/arXiv arXiv 2024

  24. [2025]

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch

    URLhttps://arxiv.org/abs/2401.08281. Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. InForty-first international conference on machine learning, 2024. Wei Fan, JinYi Yoon, and Bo Ji. imad: Intelligent multi-agent debate for efficient and accurate...

    Pith/arXiv arXiv 2024