Pith. sign in

REVIEW 4 major objections 4 minor 51 references

Agents of Diffusion shows that a frozen diffusion language model, steered by two cooperating LLM agents through natural-language feedback, outperforms autoregressive and diffusion baselines on schema-consistent, diverse JSON generation.

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

2026-08-03 11:09 UTC pith:FHUIEPBK

load-bearing objection The real contribution is a prompt-optimizer/judge loop around a frozen diffusion language model, but the multi-agent RL framing is internally contradicted by the implementation: no policy parameter is ever updated. the 4 major comments →

arxiv 2601.07152 v1 pith:FHUIEPBK submitted 2026-01-12 cs.MA

Agents of Diffusion: Enhancing Diffusion Language Models with Multi-Agent Reinforcement Learning for Structured Data Generation (Extended Version)

Aja Khanal , Kaushik T. Ranade , Rishabh Agrawal , Kalyan S. Basu , Apurva Narayan This is my paper
classification cs.MA
keywords diffusion language modelsmulti-agent reinforcement learningstructured data generationJSON synthesisprompt optimizationnatural language feedbackschema consistencydiversity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that a diffusion language model, left frozen and unmodified, can be made to generate structured JSON data that is both schema-valid and semantically diverse if two LLM agents talk to it through natural language. A prompt-optimizer agent proposes edits to the prompt, a judge agent critiques the generated samples, and the feedback loop iteratively steers the diffusion generator toward the desired format and content. The authors claim this method beats both autoregressive LLMs and earlier diffusion or prompt-search baselines on four structured-data benchmarks, achieving the highest task success rate and the lowest duplication of reference fields. If true, it would mean that controllable structured generation does not require fine-tuning or handcrafted constraints, only language-based supervision around a frozen diffusion backbone.

Core claim

The paper introduces a multi-agent reinforcement learning loop in which a stochastic prompt-optimizer policy πθ edits a text prompt, a frozen diffusion language model gφ (LLaDA-8B) samples candidate JSON records conditioned on that prompt, and a judge cluster converts five quantitative metrics—similarity, diversity, novelty, entropy, perplexity—into natural-language feedback that becomes a surrogate reward. The claim is that this closed loop drives the DLM's conditional distribution toward the valid schema subset while preserving diversity, without modifying generator weights or using explicit validation constraints. On MultiWOZ, Super-NaturalInstructions, Self-Instruct, and TruthfulQA, AoD

What carries the argument

The central object is the language-mediated reinforcement loop: a prompt-optimizer agent (an autoregressive LLM) defines a stochastic policy over prompt edits, and a judge cluster (an LLM judge plus a natural-language evaluator) converts scalar metric scores into interpretable feedback that serves as the reward signal. The diffusion language model acts as a frozen generative environment, and the edit operator U updates prompts as P_{t+1}=U(P_t,ΔP_t). The paper's theoretical support is Theorem 1, which asserts that under Lipschitz continuity and bounded edit variance this update contracts in expectation to a prompt fixed point that maximizes expected reward.

Load-bearing premise

The claimed RL update assumes the prompt-optimizer policy is actually trainable and its parameters change via gradient steps, but the paper also states the optimization is parameter-free and acts through discrete language edits rather than backpropagation, leaving ambiguous whether any learning beyond iterative prompting occurs.

What would settle it

Capture the prompt-optimizer LLM's weights (or API state) before and after a full AoD run. If the weights and internal parameters are byte-for-byte identical and no gradient-based update is applied, then the policy-gradient step in Algorithm 1, θ←θ+η∇θ Eπθ[r_t], is not executed and the claimed reinforcement learning component is not present.

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

If this is right

  • If the central claim holds, diffusion language models become controllable for structured output without any parameter updates, which would lower the compute barrier for adapting DLMs to schema-heavy domains.
  • Natural-language feedback can act as a dense, interpretable reward signal that avoids scalar reward hacking and gives human-readable reasons for each prompt change.
  • The reported combination of high task success and low field overlap suggests that agent-guided diffusion could produce synthetic training data that is both useful and less likely to leak private or memorized fields.
  • The method's model-agnostic agent loop (tested across several LLMs as optimizer and judge) implies the control layer transfers across different autoregressive backbones while keeping the diffusion generator fixed.

Where Pith is reading between the lines

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

  • If the RL framing is not literal—if the prompt-optimizer policy's parameters never actually update—the results may be explainable as iterative prompt refinement rather than policy-gradient learning; a direct weight-difference test would settle which story is true.
  • The same language-mediated loop might extend beyond JSON to tabular records, code, or any constrained output space where an autoregressive judge can identify structural errors, though the paper only demonstrates JSON.
  • Because the five reward metrics are computed on the generated sample itself, the approach could be sensitive to the choice of encoder or n-gram definition; a different evaluator might change which outputs are considered novel or similar.
  • The convergence theorem assumes a well-defined gradient for the prompt-optimizer policy, but the implementation admits discrete, language-conditioned edits; reconciling these would require either a relaxation of the theory or a re-description of the method as stochastic search.

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 / 4 minor

Summary. The paper proposes Agents of Diffusion (AoD), a framework that couples a frozen diffusion language model (LLaDA-8B) with two autoregressive LLM agents — a prompt optimizer and a judge — to generate schema-consistent, diverse JSON data. The authors frame the prompt-optimizer as a stochastic policy π_θ trained with REINFORCE/PPO-style updates over natural-language feedback, and report state-of-the-art results across MultiWOZ, Super-NaturalInstructions, Self-Instruct, and TruthfulQA, including the highest Task Success Rate (0.79) and lowest Field Overlap (0.29). The main claimed contribution is that a DLM can be supervised by cooperative agents through language-mediated RL without modifying the DLM's parameters.

Significance. If the central mechanism were supported, the paper would offer a practical way to obtain controllable structured generation from frozen diffusion language models, which is a relevant goal for multi-agent and controllable-generation communities. The paper has several strengths: it provides full prompts and a complete experiment log in the appendix, reports an ablation over agentic components, includes model-transferability experiments, and emphasizes reproducibility on consumer-grade hardware. These are concrete and useful. However, the core RL claim is not backed by the implementation described in the same manuscript: the policy parameter θ is never shown to exist, to be updated, or to influence the prompt-rewriting LLM. The formal results either assume the conclusion (Proposition 1) or assume the very aligned-ascent property the algorithm is supposed to provide (Theorem 1). The empirical tables also lack error bars and statistical tests. Given that the advertised contribution is 'multi-agent reinforcement learning,' these gaps are load-bearing rather than cosmetic.

major comments (4)
  1. [Algorithm 1 (§3.3) and §4.1] Algorithm 1 line 10 updates θ ← θ + η ∇_θ E_{π_θ}[r_t], and Appendix A.10 derives REINFORCE/PPO updates with clipped ratios and KL regularization. But §3.3 states that π_θ 'approximates the gradient ... through discrete, language-conditioned updates rather than backpropagation,' §4.1 and Table 2 describe the loop as running without fine-tuning and with 'limited backpropagation overhead,' and Appendix A.5 shows the reward is a natural-language sentence, not a scalar. No mechanism is provided by which θ exists, is updated, or changes the behavior of the prompt-optimizer LLM. If θ is never updated, the method reduces to iterative prompt editing akin to PromptBreeder/EvoPrompt, and the multi-agent RL claim is unsupported. This is the paper's central thesis, not a presentation issue.
  2. [A.11, Theorem 1] The convergence proof assumes A4: E[ΔP_t | P_t] = η G(P_t), ⟨∇T(P_t), G(P_t)⟩ ≥ c‖∇T(P_t)‖², and bounded variance. This is essentially the assumption that the policy's edits already perform aligned gradient ascent on T(P). The algorithm itself gives no reason this holds for language-model prompt rewrites driven by natural-language feedback. Moreover, the proof concludes that the contraction 'extends to the exact update' with no additional argument. Thus Theorem 1 does not establish convergence of AoD; it restates the needed condition as an axiom and then derives its consequences. The assumptions A1–A3 are also unverified for the actual T(P) induced by LLaDA and the LLM judge/scorer.
  3. [A.12, Proposition 1] Proposition 1 claims that a diffusion model approximates q_real better than an autoregressive model under bounded reconstruction error. The proof's key assumption A3 is ε_φ + δ_φ ≤ ε_AR − δ_AR, which is precisely the inequality needed to obtain KL(q_real‖p_φ) ≤ KL(q_real‖p_AR) from the preceding bounds. Since none of ε_φ, δ_φ, ε_AR, δ_AR is measured or bounded from data, the proposition is a conditional statement that assumes its conclusion. The empirical comparison in Table 1 cannot rescue this because the proposition is stated as an unconditional theoretical result. The same pattern appears in Proposition 2 (A.13), where the ordering-preserving property is built into the Lipschitz/monotonicity assumptions on R and the local-neighborhood argument.
  4. [Table 1 and §4.2] Table 1 reports no error bars, confidence intervals, or significance tests even though the caption says each experiment was repeated 15 times. The primary in-loop metrics (Similarity, Diversity, Novelty, Perplexity) are used as rewards during the iterative loop and then reported as headline evaluations; the separate 'independent' metrics (BLEU, ROUGE-L, METEOR, TSR, Field Overlap) are also presented without variance. Moreover, no agentic autoregressive baseline is included — a same-loop AR generator with the same judge and prompt-optimizer feedback but no DLM — which would be needed to isolate the contribution of the diffusion generator. Without such comparisons and without standard errors, the claim that AoD 'outperforms' baselines is not established.
minor comments (4)
  1. [Conclusion and A.16] Typographical errors: 'Futhermore' in the conclusion, and 'Baseines Choice' in the A.16 heading.
  2. [A.11, A.13] The appendix contains formatting artifacts such as 'Compactitem 1.' and 'Compactitem 2.' that interrupt the proof text. Please clean up the LaTeX and restate the assumptions as numbered items.
  3. [§3.3 and A.10] The paper oscillates between 'no scalar rewards / no reward modeling' (abstract, §3.3) and the scalar-reward variant in the ablation (Figure 3: FRL–AR–S, FRL–DLM–S). Please clarify which components are used in the final AoD configuration and whether scalar rewards are ever part of the proposed method.
  4. [Figure 2 and Figure 3] The radar plots and bar charts are referenced but their axes and statistical content are not described in the text. Please add figure captions that state what is plotted and whether bars/points show means and standard errors.

Circularity Check

3 steps flagged

Two supporting formal results reduce to their own assumptions (Prop. 1's A3 is the conclusion; Theorem 1 assumes the policy already performs aligned ascent), and part of Table 1 reports in-loop reward metrics as quality evidence; the independent TSR/BLEU/Field-Overlap comparison is not circular.

specific steps
  1. self definitional [Proposition 1 and its proof, Appendix A.12]
    "Suppose: ... (A3) (Optimization noise) Training induces suboptimality gaps δφ ≥ 0 and δAR ≥ 0 so that each model attains a divergence within its bound: KL(q_real∥p_φ) ≤ εφ+δφ and KL(q_real∥p_AR) ≥ εAR−δAR. If εφ+δφ ≤ εAR−δAR, then KL(q_real∥p_φ) ≤ KL(q_real∥p_AR)."

    The proposition is supposed to show that diffusion approximates q_real at least as well as autoregressive models. But condition A3 asserts exactly the target inequality: the diffusion upper bound is no larger than the AR lower bound. Lemma 1 and Lemma 2 only establish the sandwich KL_diff ≤ εφ+δφ and KL_AR ≥ εAR−δAR; the proof completes by transitivity of the assumed inequality. Thus the 'result' is equivalent to its input condition, not a derived property of diffusion versus autoregressive modeling.

  2. self definitional [Theorem 1 proof, Appendix A.11, assumption (A4)]
    "(A4) The policy satisfies aligned ascent and bounded variance: E[ΔP_t|P_t] = ηG(P_t), ⟨∇T(P_t), G(P_t)⟩ ≥ c‖∇T(P_t)‖², E[‖ΔP_t−EΔP_t‖²|P_t] ≤ σ²."

    The theorem claims prompt edits sampled from πθ converge to a maximizer of the reward objective T. Assumption A4, however, states that the conditional expectation of the sampled edit is η times an ascent direction of T with positive alignment c. Lemma 2 ('Expected ascent') then merely substitutes this assumption into Lemma 1; the remaining argument is the standard gradient-ascent contraction analysis. The reward-aligned behavior of the policy is therefore assumed, not derived. Section 3.3 says πθ acts 'through discrete, language-conditioned updates rather than backpropagation', so no mechanism is given by which the LLM prompt-editor realizes this aligned-ascent condition.

  3. other [Section 4.1 Evaluation Metrics / Table 1 / Section 4.2 Discussion]
    "During training, the judge agent leverages five grounded metrics: perplexity PPL(x), semantic similarity Sim(x,X), diversity Distinct-n(x), token entropy H(x), and novelty Nov-n(x), to generate interpretable natural language feedback for the prompt optimizer. ... Table 1: ... The first five metrics correspond to those used in training, while the last five serve as independent evaluation metrics."

    The five metrics used as in-loop rewards are then reported in Table 1 and interpreted in Section 4.2 as evidence of quality ('High Similarity (0.88) combined with strong Diversity (0.72) and Novelty (0.83) demonstrates that AoD generates data that remains semantically faithful...'). Since AoD's optimization loop directly maximizes these quantities, high values partly record the objective being optimized rather than an independent test. The paper does disclose the distinction and does provide non-reward metrics (BLEU, ROUGE, METEOR, TSR, Field Overlap), so this is partial self-referentiality rather than the sole basis of the empirical claim.

full rationale

The independent empirical core is not circular: TSR, Field Overlap, BLEU, ROUGE, and METEOR are not used as in-loop rewards, so the headline 'highest TSR (0.79) and lowest Field Overlap (0.29)' is a genuine external comparison against baselines. However, the paper's first-principles formal support contains two steps that reduce to their own inputs. Proposition 1's assumption A3 is literally the inequality being proved, and Theorem 1's A4 assumes the policy already performs reward-aligned ascent, which is the substantive behavior the RL framework claims to establish. Additionally, the first five columns of Table 1 are in-loop reward components reported as quality evidence, though the paper marks them as training metrics and supplies independent metrics alongside. There is no load-bearing self-citation chain; the issues are definitional assumptions and partially self-referential evaluation, not citation circularity. Overall score 5 reflects a mix of two circular formal results and one partially self-referential empirical presentation, while the main benchmark comparison retains independent content.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The method's central claims rest on unspecified reward weights/thresholds, unverified regularity assumptions in Theorems/Propositions, and the assumption that prompt edits can steer LLaDA's distribution; no new physical entities are introduced.

free parameters (6)
  • Terminal objective weights α, β, δ = unspecified
    Eq. (3) maximizes αE[V_S]+βE[Sim]+δE[Distinct-n]; weights are never given or swept, and results could change with their values.
  • Validity/similarity thresholds τ_valid, τ_sim = unspecified
    Eq. (3) constrained form requires thresholds; no values are reported, so the optimization target is underdetermined.
  • TSR thresholds τ_s, τ_d = unspecified, 'chosen a priori' (A.16)
    TSR definition depends on thresholds; values are not given, so TSR 0.79 cannot be independently reproduced or interpreted.
  • n-gram order n for Distinct-n and Nov-n = unspecified
    Diversity and novelty scores depend on n; no n reported.
  • Semantic similarity encoder E(·) = unspecified
    Sim is defined as max cosine similarity under 'an encoder such as a sentence transformer' (A.16), but the specific encoder is not named; PPL also needs a reference LM, unspecified.
  • Learning rate η and discount γ = unspecified
    Algorithm 1 uses η and the MDP uses γ; neither is specified or tied to implementation.
axioms (6)
  • ad hoc to paper T(P) is L-smooth, μ-strongly concave and policy edits satisfy aligned ascent/bounded variance (A1–A4)
    Theorem 1's convergence depends on these unverified properties of an LLM prompt-edit distribution; A4 essentially assumes the update is already improving T.
  • ad hoc to paper εφ+δφ ≤ εAR−δAR (Prop. 1, A.12)
    This inequality is the conclusion of the proposition; no derivation or empirical estimate is provided.
  • ad hoc to paper NLE→judge map is L_J-Lipschitz and R is strictly monotone in Sim with sensitivity α>0 (Prop. 2, A.13)
    Assumed regularity of unspecified LLM judgment and reward components; not verified.
  • domain assumption Prompt-conditioned DLM expectations are locally Lipschitz in P for any bounded measurable f (A.13 Lemma 1)
    Standard regularity hope for prompt-conditioned samplers, asserted by 'dominated convergence plus local smooth dependence' without proof.
  • domain assumption DLM LLaDA with T=12 and no classifier-free guidance produces diverse but structure-brittle outputs that prompts can steer
    The whole control mechanism assumes prompt edits reliably shift the DLM's conditional distribution toward schema validity; no analysis of LLaDA's prompt sensitivity is provided.
  • domain assumption Random subsampling of datasets eliminates memorization risk
    Paper claims subsampling 'minimizes risk of copying' but no memorization tests beyond post-hoc Field Overlap are specified.

pith-pipeline@v1.3.0-alltime-deepseek · 29541 in / 15606 out tokens · 153831 ms · 2026-08-03T11:09:42.298155+00:00 · methodology

0 comments
read the original abstract

Generating high-quality structured data such as JSON records, remains a fundamental challenge for large language models (LLMs), particularly when semantic richness must coexist with strict schema adherence. While autoregressive LLMs offer strong structural consistency, they often struggle with semantic variation and output diversity. In contrast, diffusion language models (DLMs) introduce powerful mechanisms for semantic richness and bidirectional decoding, yet lack the inductive biases needed for reliable structure preservation. We present Agents of Diffusion (AoD), a novel framework that unifies the generative flexibility of DLMs with the reasoning capabilities of autoregressive models through language-mediated reinforcement learning. AoD frames structured text generation as a multi-agent alignment process, where a prompt optimization agent collaborates with a judge agent to iteratively guide a DLM using natural language feedback. This approach enables controllable, schema-consistent generation without modifying model parameters or relying on handcrafted constraints. AoD advances the state of controllable generation by demonstrating that diffusion models, when supervised by cooperative agents, can achieve both high semantic novelty and structural fidelity. Across multiple structured data benchmarks, AoD consistently outperforms diffusion and autoregressive baselines, establishing a new path forward for structure-aware, diversity-enhanced text synthesis.

Figures

Figures reproduced from arXiv: 2601.07152 by Aja Khanal, Apurva Narayan, Kalyan S. Basu, Kaushik T. Ranade, Rishabh Agrawal.

Figure 1
Figure 1. Figure 1: Agents of Diffusion: Overview of the multi-agent training framework. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of normalized metrics across datasets. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation of agents, averaged across datasets and Optimizer–Judge pairs. SP = Static Prompt, PO = Prompt Optimizer, [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

51 extracted references · 45 linked inside Pith

  1. [1]

    Ateret Anaby-Tavor, Boaz Carmeli, Esther Goldbraich, Amir Kantor, George Kour, Segev Shlomov, Naama Tepper, and Naama Zwerdling. 2019. Not Enough Data? Deep Learning to the Rescue! https://arxiv.org/abs/1911.03118

    Pith/arXiv arXiv 2019

  2. [2]

    André Bauer, Simon Trapp, Michael Stenger, Robert Leppich, Samuel Kounev, Mark Leznik, Kyle Chard, and Ian Foster. 2024. Comprehensive Exploration of Synthetic Data Generation: A Survey. https://arxiv.org/abs/2401.02524

    Pith/arXiv arXiv 2024

  3. [3]

    Luca Beurer-Kellner, Marc Fischer, and Martin Vechev. 2024. Guiding LLMS THE RIGHT WAY: Fast, non-invasive constrained generation. https://arxiv.org/abs/ 2403.06988

    Pith/arXiv arXiv 2024

  4. [4]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin...

    Pith/arXiv arXiv 2020

  5. [5]

    Paweł Budzianowski, Tsung-Hsien Wen, Bo-Hsiang Tseng, Iñigo Casanueva, Stefan Ultes, Osman Ramadan, and Milica Gašić. 2018. MultiWOZ - A Large- Scale Multi-Domain Wizard-of-Oz Dataset for Task-Oriented Dialogue Modelling. https://arxiv.org/abs/1810.00278

    Pith/arXiv arXiv 2018

  6. [6]

    Yuji Cao, Huan Zhao, Yuheng Cheng, Ting Shu, Yue Chen, Guolong Liu, Gaoqi Liang, Junhua Zhao, Jinyue Yan, and Yun Li. 2024. Survey on Large Language Model-Enhanced Reinforcement Learning: Concept, Taxonomy, and Methods. IEEE Transactions on Neural Networks and Learning Systems(01 2024), 1–21. https://doi.org/10.1109/tnnls.2024.3497992

    arXiv 2024

  7. [7]

    Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebas- tian Gehrmann, and et al. 2022. PaLM: Scaling Language Modeling with Pathways. arXiv preprint arXiv:2204.02311(April 2022). https://arxiv.org/abs/2204.02311

    Pith/arXiv arXiv 2022

  8. [8]

    DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, and et al. 2025. DeepSeek- R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. https://arxiv.org/abs/2501.12948 arXiv preprint arXiv:2501.12948

    Pith/arXiv arXiv 2025

  9. [9]

    Jasper Dekoninck, Marc Fischer, Luca Beurer-Kellner, and Martin Vechev. 2023. Controlled Text Generation via Language Model Arithmetic. https://arxiv.org/ abs/2311.14479

    Pith/arXiv arXiv 2023

  10. [10]

    Ruan, Yaxing Cai, Ruihang Lai, Ziyi Xu, Yilong Zhao, and Tianqi Chen

    Yixin Dong, Charlie F. Ruan, Yaxing Cai, Ruihang Lai, Ziyi Xu, Yilong Zhao, and Tianqi Chen. 2025. XGRAMMAR: Flexible and efficient structured generation engine for large language models. https://arxiv.org/abs/2411.15100

    Pith/arXiv arXiv 2025

  11. [11]

    Tenenbaum, and Igor Mor- datch

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mor- datch. 2023. Improving Factuality and Reasoning in Language Models through Multiagent Debate. https://doi.org/10.48550/arXiv.2305.14325

    Pith/arXiv arXiv doi:10.48550/arxiv.2305.14325 2023

  12. [12]

    Chrisantha Fernando, Dylan Banarse, Henryk Michalewski, Simon Osindero, and Tim Rocktäschel. 2023. Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution. https://doi.org/10.48550/arXiv.2309.16797

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.16797 2023

  13. [13]

    Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and Lingpeng Kong. 2022. DiffuSeq: Sequence to Sequence Text Generation with Diffusion Models. https: //arxiv.org/abs/2210.08933

    Pith/arXiv arXiv 2022

  14. [14]

    Qingyan Guo, Rui Wang, Junliang Guo, Bei Li, Kaitao Song, Xu Tan, Guoqing Liu, Jiang Bian, and Yujiu Yang. 2023. Connecting Large Language Models with Evolutionary Algorithms Yields Powerful Prompt Optimizers.arXiv (Cornell University)(01 2023). https://doi.org/10.48550/arxiv.2309.08532

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.08532 2023

  15. [15]

    Chawla, Olaf Wiest, and Xiangliang Zhang

    Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V. Chawla, Olaf Wiest, and Xiangliang Zhang. 2024. Large Language Model based Multi-Agents: A Survey of Progress and Challenges. https://doi.org/10.48550/ arXiv.2402.01680

    Pith/arXiv arXiv doi:10.48550/arxiv.2402.01680 2024

  16. [16]

    Emiel Hoogeboom, Alexey A Gritsenko, Jasmijn Bastings, Ben Poole, van , and Tim Salimans. 2021. Autoregressive Diffusion Models. https://arxiv.org/abs/ 2110.02037

    Pith/arXiv arXiv 2021

  17. [17]

    Jiaxin Huang, Shixiang Gu, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu, and Jiawei Han. 2023. Large Language Models Can Self-Improve.ACL Anthology Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (01 2023). https://doi.org/10.18653/v1/2023.emnlp-main.67

    doi:10.18653/v1/2023.emnlp-main.67 2023

  18. [18]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. 2022. Survey of Hallucination in Natural Language Generation.Comput. Surveys55 (11 2022). https://doi.org/10. 1145/3571730

    2022

  19. [19]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, De- vendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. Mistral 7B. https://doi.or...

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.06825 2023

  20. [20]

    Martin Josifoski, Marija Sakota, Maxime Peyrard, and Robert West. 2023. Exploit- ing Asymmetry for Synthetic Training Data Generation: SynthIE and the Case of Information Extraction. https://arxiv.org/abs/2303.04132

    Pith/arXiv arXiv 2023

  21. [21]

    Varshney, Caiming Xiong, and Richard Socher

    Nitish Shirish Keskar, Bryan McCann, Lav R. Varshney, Caiming Xiong, and Richard Socher. 2019. CTRL: A Conditional Transformer Language Model for Controllable Generation.arXiv:1909.05858 [cs](09 2019). https://arxiv.org/abs/ 1909.05858

    Pith/arXiv arXiv 2019

  22. [22]

    Chungpa Lee, Jongho Im, and Kim Joseph. 2025. A Generalized Theory of Mixup for Structure-Preserving Synthetic Data. https://arxiv.org/abs/2503.02645

    Pith/arXiv arXiv 2025

  23. [23]

    Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, and Tatsunori B Hashimoto. 2022. Diffusion-LM Improves Controllable Text Generation. https: //arxiv.org/abs/2205.14217

    Pith/arXiv arXiv 2022

  24. [24]

    Stephanie Lin, Jacob Hilton, and Owain Evans. 2021. TruthfulQA: Measuring How Models Mimic Human Falsehoods.arXiv:2109.07958 [cs](09 2021). https: //arxiv.org/abs/2109.07958

    Pith/arXiv arXiv 2021

  25. [25]

    Shengcai Liu, Caishun Chen, Xinghua Qu, Ke Tang, and Yew-Soon Ong. 2023. Large Language Models as Evolutionary Optimizers. https://arxiv.org/abs/2310. 19046

    2023

  26. [26]

    Calvin Luo. 2022. Understanding Diffusion Models: A Unified Perspective. arXiv:2208.11970 [cs](08 2022). https://arxiv.org/abs/2208.11970

    Pith/arXiv arXiv 2022

  27. [27]

    Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. 2025. Large Language Diffusion Models. https://arxiv.org/abs/2502.09992

    Pith/arXiv arXiv 2025

  28. [28]

    OpenAI. 2023. GPT-4 Technical Report.arXiv:2303.08774 [cs](03 2023). https: //doi.org/10.48550/arXiv.2303.08774

    Pith/arXiv arXiv doi:10.48550/arxiv.2303.08774 2023

  29. [29]

    Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schul- man, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Training lan- guage models to follow instructions with human...

    Pith/arXiv arXiv 2022

  30. [30]

    Bo Pan, Jiaying Lu, Ke Wang, Li Zheng, Zhen Wen, Yingchaojie Feng, Minfeng Zhu, and Wei Chen. 2024. AgentCoord: Visually Exploring Coordination Strategy for LLM-based Multi-Agent Collaboration. https://arxiv.org/abs/2404.11943

    Pith/arXiv arXiv 2024

  31. [31]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. https: //arxiv.org/abs/1910.10683

    Pith/arXiv arXiv 2019

  32. [32]

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Rad- ford, Mark Chen, and Ilya Sutskever. 2021. Zero-Shot Text-to-Image Generation. arXiv:2102.12092 [cs](02 2021). https://arxiv.org/abs/2102.12092

    Pith/arXiv arXiv 2021

  33. [33]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. arXiv:2112.10752 [cs](04 2022). https://arxiv.org/abs/2112.10752

    Pith/arXiv arXiv 2022

  34. [34]

    Noah Shinn, Beck Labash, and Ashwin Gopinath. 2023. Reflexion: an autonomous agent with dynamic memory and self-reflection.arXiv:2303.11366 [cs](03 2023). https://arxiv.org/abs/2303.11366

    Pith/arXiv arXiv 2023

  35. [35]

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhu- patiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juli- ette Love, and et al. 2024. Gemma: Open Models Based on Gemini Research and Technology. https://doi.org/10.48550/arXiv.2403.08295 arXiv preprint arXiv:2403.08295

    Pith/arXiv arXiv doi:10.48550/arxiv.2403.08295 2024

  36. [36]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guil- laume Lample. 2023. LLaMA: Open and Efficient Foundation Language Models. arXiv:2302.13971 [cs](02 2023)

    Pith/arXiv arXiv 2023

  37. [37]

    Khanh-Tung Tran, Dung Dao, Minh-Duong Nguyen, Quoc-Viet Pham, Barry O’Sullivan, and Hoang D Nguyen. 2025. Multi-Agent Collaboration Mechanisms: A Survey of LLMs. https://arxiv.org/abs/2501.06322

    Pith/arXiv arXiv 2025

  38. [38]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention Is All You Need. https://arxiv.org/abs/1706.03762

    Pith/arXiv arXiv 2017

  39. [39]

    Veniamin Veselovsky, Manoel Horta Ribeiro, Akhil Arora, Martin Josifoski, Ash- ton Anderson, and Robert West. 2023. Generating Faithful Synthetic Data with Large Language Models: A Case Study in Computational Social Science. https://doi.org/10.48550/arXiv.2305.15041

    Pith/arXiv arXiv doi:10.48550/arxiv.2305.15041 2023

  40. [40]

    Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2022. Self-Instruct: Aligning Language Model with Self Generated Instructions.arXiv (Cornell University)(12 2022). https://doi.org/10.48550/arxiv.2212.10560

    Pith/arXiv arXiv doi:10.48550/arxiv.2212.10560 2022

  41. [41]

    Yizhong Wang, Swaroop Mishra, Pegah Alipoormolabashi, Yeganeh Kordi, Amir- reza Mirzaei, Anjana Arunkumar, Arjun Ashok, Arut Selvan Dhanasekaran, Atharva Naik, David Stap, and et al. 2022. Super-NaturalInstructions: General- ization via Declarative Instructions on 1600+ NLP Tasks. https://arxiv.org/abs/ 2204.07705 arXiv preprint arXiv:2204.07705

    Pith/arXiv arXiv 2022

  42. [42]

    Zifeng Wang, Chun-Liang Li, Vincent Perot, Long T Le, Jin Miao, Zizhao Zhang, Chen-Yu Lee, and Tomas Pfister. 2024. CodecLM: Aligning Language Models with Tailored Synthetic Data. https://arxiv.org/abs/2404.05875

    Pith/arXiv arXiv 2024

  43. [43]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2022. Chain of Thought Prompting Elicits Reasoning in Large Language Models.arXiv:2201.11903 [cs](10 2022). https://arxiv.org/abs/2201.11903

    Pith/arXiv arXiv 2022

  44. [44]

    White, Doug Burger, and Chi Wang

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W. White, Doug Burger, and Chi Wang. 2023. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. https://doi.org/10.48550/arXiv. 2308.08155

    Pith/arXiv arXiv doi:10.48550/arxiv.2308.08155 2023

  45. [45]

    Siyuan Wu, Yue Huang, Chujie Gao, Dongping Chen, Qihui Zhang, Yao Wan, Tianyi Zhou, Xiangliang Zhang, Jianfeng Gao, Chaowei Xiao, and Lichao Sun

  46. [46]

    Jiacheng Ye, Jiahui Gao, Qintong Li, Hang Xu, Jiangtao Feng, Zhiyong Wu, Tao Yu, and Lingpeng Kong. 2022. ZeroGen: Efficient Zero-shot Learning via Dataset Generation. https://arxiv.org/abs/2202.07922

    Pith/arXiv arXiv 2022

  47. [47]

    Yue Yu, Yuchen Zhuang, Jieyu Zhang, Yu Meng, Alexander Ratner, Ranjay Krishna, Jiaming Shen, and Chao Zhang. 2023. Large Language Model as Attributed Training Data Generator: A Tale of Diversity and Bias. https://doi.org/10.48550/ arXiv.2306.15895

    Pith/arXiv arXiv doi:10.48550/arxiv.2306.15895 2023

  48. [48]

    Hengrui Zhang, Jiani Zhang, Balasubramaniam Srinivasan, Zhengyuan Shen, Xiao Qin, Christos Faloutsos, Huzefa Rangwala, and George Karypis. 2023. Mixed- Type Tabular Data Synthesis with Score-based Diffusion in Latent Space. https: //arxiv.org/abs/2310.09656

    Pith/arXiv arXiv 2023

  49. [49]

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. 2023. A Survey of Large Language Models.arXiv:2303.18223 [cs](03 2023). http...

    Pith/arXiv arXiv 2023

  50. [50]

    address":

    Ying Zhou, Xinyao Wang, Yulei Niu, Yaojie Shen, Lexin Tang, Fan Chen, Ben He, Le Sun, and Longyin Wen. 2024. DiffLM: Controllable Synthetic Data Generation via Diffusion Language Models. https://arxiv.org/abs/2411.03250 A APPENDIX This appendix provides all resources necessary to fully reproduce the experiments described in theAgents of Diffusion (AoD)stu...

    Pith/arXiv arXiv 2024

  51. [2024]

    https://arxiv.org/abs/2406.18966

    UniGen: A Unified Framework for Textual Dataset Generation Using Large Language Models. https://arxiv.org/abs/2406.18966

    arXiv