Pith. sign in

REVIEW 4 major objections 5 minor 147 references

Masked diffusion language models beat autoregressive LLMs over four times their size as text-based world models, and RL agents trained on their rollouts gain up to 47% on unseen environments.

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

2026-08-02 14:44 UTC pith:KDGLKJ2G

load-bearing objection First systematic MDLM-as-world-model study with a useful dataset and a clean diffusion-objective control, but the headline causal claim is built on synthetic hindsight grounding and partially scaffolded RL rollouts. the 4 major comments →

arxiv 2607.16204 v1 pith:KDGLKJ2G submitted 2026-05-07 cs.AI cs.LG

Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RL

Darshan Deshpande This is my paper
classification cs.AI cs.LG
keywords masked diffusion language modelsworld modelsagentic reinforcement learningGRPOtext-based environment simulationautoregressive left-to-right biasstate-action trajectorieszero-shot transfer
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.

The paper sets out to establish that masked diffusion language models (MDLMs)—which predict masked tokens using both left and right context—are better world models for text-based agentic reinforcement learning than autoregressive LLMs several times their size. Environment states are globally interdependent, with tool schemas, prior turns, and expected outcomes constraining one another, and the paper argues that left-to-right causal decoding cannot condition on trailing anchors, producing prefix-consistent but globally incoherent rollout states. On a curated corpus of 239,403 state-action trajectories, it reports that MDLMs beat AR baselines up to 35B parameters on coherence, groundedness, and diversity at comparable latency, and that GRPO training on MDLM rollouts yields up to 47% absolute task-success gains on three held-out environments without environment-specific fine-tuning. If correct, the limiting factor in LLM-based world modeling is the causal inductive bias, not model capacity, opening a practical route to on-demand, steerable training environments for RL.

Core claim

The paper's central claim is that masked diffusion language models, by conditioning on anchors in both directions while denoising, learn an any-order likelihood over environment-state token sequences and therefore produce rollouts that are more coherent, more grounded in tool schemas and prior turns, and more diverse than autoregressive LLMs more than four times their size. The authors formalize text-based world modeling as steerable transition dynamics p(c_{t+1} | e0, c_<=t, a_<=t; T, R, D), decomposing each state into initial environment state, task context, tool schemas, domain rules, and steering directives, and they curate 239,403 grounded trajectories from nine environments and twelve

What carries the argument

The load-bearing mechanism is anchor-aware masked denoising. An MDLM is trained to recover randomly masked tokens in a next environment state from the unmasked context; at inference it fills masked positions in parallel or block-wise, so every generated token can attend to anchors fixed by the schema, tool definitions, prior turns, and trailing fields. This is the object that carries the argument: it removes the left-to-right commitment of an autoregressive factorization, lets the model condition on globally interdependent anchors such as a trailing status field, and adds a second axis of stochasticity—which positions are committed in which order—that yields rollout diversity beyond temperat

Load-bearing premise

The load-bearing premise is that the synthesized training references are faithful to true environment dynamics: roughly 80% of the corpus's grounding context was written by a frontier LLM instructed to invent hindsight environment state, domain rules, and deliberate distractors, so if those references are noisy or hallucinated, the groundedness and downstream RL gains may partly measure imitation of the annotator rather than real simulation.

What would settle it

Take a held-out set of real trajectories from one of the nine source environments, execute each recorded action in the actual environment to obtain true next states, and compare MDLM versus AR predictions against these real references instead of the synthesized ones; if the 8B MDLM's MAUVE and Self-BLEU advantage over AR baselines shrinks or vanishes, the claimed bidirectional-anchor advantage is in part an artifact of LLM-generated evaluation data. In parallel, ablate the hand-written world-model prompts and deterministic local responders that precompute pagination, validate tool calls, and s

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

If this is right

  • An 8B MDLM fine-tuned on world-model data can replace AR LLMs of 20-35B parameters for rollout generation, lowering the compute cost of environment simulation while keeping latency competitive.
  • GRPO agents trained entirely on imagined rollouts from an MDLM world model improve on real held-out environments without environment-specific fine-tuning, suggesting world models can serve as adaptive curricula as task difficulty scales.
  • Rollout diversity is higher under MDLMs (lower Self-BLEU, higher Distinct-N and MAUVE), so RL training on them should be less prone to prefix-induced mode collapse into a few workflows.
  • Known failure modes—pagination drift, block-level repetition under verbose tool outputs, and brittle API-key generation—are explicit limits that the paper says degrade training signal and need steering or post-processing.

Where Pith is reading between the lines

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

  • Not tested in the paper: any structured-generation task with global constraints—database state simulation, JSON schema completion, tool-output synthesis—should show a similar MDLM advantage if the causal bias is the bottleneck; a cheap experiment is to compare MDLM and AR next-state prediction on raw execution logs from a real simulator rather than LLM-synthesized references.
  • The paper's groundedness references are largely synthesized by a frontier LLM with deliberate distractor rows, so 'groundedness' may partly measure fidelity to the annotator's hindsight model of the environment; re-validating the best MDLM against true environment rollouts would tell whether the bidirectional advantage persists.
  • If the RL improvements transfer to larger backbones and longer horizons, world models trained this way could let agentic RL scale environments without hand-engineering rewards—but the same mechanism could train an agent against a world model that has drifted from the true environment, so online calibration against real rollouts is a natural safety check the paper does not address.

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 treating text-based world modeling as a steerable transition-dynamics problem and introduces a 239,403-trajectory dataset spanning nine environments and twelve frontier model families. It compares autoregressive LLMs with masked diffusion language models (MDLMs), reporting that MDLMs achieve better coherence, groundedness, and rollout diversity than AR models more than 4× their size, at comparable latency. A plug-and-play GRPO training framework with deterministic state checks is introduced, and zero-shot transfer experiments on ScienceWorld, ALFWorld, and AppWorld report absolute gains up to 47% over baselines across 1.2B–7B agent backbones. Human evaluation and behavioral analyses of failure modes are also presented.

Significance. If the central comparison is valid, the result is significant: it isolates the causal inductive bias of AR LMs as the primary bottleneck in text-based world modeling, rather than model capacity. The paper ships a large open dataset, an open-source training/evaluation pipeline, and a controlled Qwen3-8B baseline that isolates the diffusion objective from base-model capacity. The three-seed protocol is a good start, but the lack of error bars and the use of test-set-selected hyperparameters weaken the quantitative claims. The synthetic grounding of the benchmark and the deterministic scaffolding in the RL experiments are the main threats to the paper's central causal claim; both need to be addressed before the results can be taken as evidence about learned dynamics.

major comments (4)
  1. [§4 Training, Tables 1–2] No confidence intervals or error bars are reported despite three seeds, and the text says 'We perform an automated sweep over learning rates ... and batch sizes ... and report only the best scores.' There is no statement that the sweep selection was performed on a held-out split. With multiple models and metrics, the headline margins (e.g., SDAR-8B +0.283 in-domain MAUVE over Qwen-3-8B) may be within noise or reflect test-set selection. Please report mean ± std across seeds and clarify how hyperparameters were selected (e.g., validation split).
  2. [Appendix C, Dataset Grounding] The benchmark's 'grounded' references are not raw environment outputs: Claude-4.6-Sonnet synthesizes hindsight instructions for ~80% of trajectories, including extra database states, domain rules, and deliberate distractors, and the same model family was also used for trajectory generation. The 87% approval from two annotators does not verify factual agreement with the actual environment state, and no error analysis of hallucinated database states is provided. This threatens the groundedness and fidelity comparisons in Table 1, since the models may be fitting a Claude-generated target distribution rather than the true transition function. Please add an ablation on the unaugmented 17.2% split and a factual-consistency check against real environment outcomes.
  3. [Appendix D, RL World-Model Setup] The downstream RL experiments use hand-crafted WM prompts with baked room maps (ALFWorld, D.2) and local deterministic responders (AppWorld, D.3) that pre-compute query results, handle invalid tool calls, and respond to mutation tools without invoking the trained WM. As a result, Table 3's SDAR-WM vs Qwen-WM gap may reflect prompt formatting and oracle record retrieval rather than learned transition dynamics. Please quantify how many rollout transitions are actually produced by the trained WM, or ablate the deterministic components, to support the causal claim that MDLM rollouts improve downstream RL.
  4. [Abstract and §6, RQ2 Results] The 'up to 47% absolute gain' is measured against the untrained base (LFM2.5-1.2B on ALFWorld: 5.7%→53.6%), not against the strongest baseline. Versus SFT-only the gain is +10.7 points, and versus Qwen-WM +5.0 points. The abstract and conclusion should state the appropriate comparison baseline; as written, the claim overstates the MDLM-specific contribution.
minor comments (5)
  1. [§2 Related Work] Typos: 'techniuqes' and 'alogrithm' should be corrected.
  2. [Table 1] Several cells have formatting issues (e.g., '712 .781.840'), making the table difficult to read. Please fix spacing.
  3. [§5 RQ3] The human evaluation covers only SDAR-generated states; no AR-generated states are rated by the same annotators. Claims about MDLM superiority in human-judged realism should be scoped accordingly.
  4. [Table 6 / §5 latency claim] The abstract claims 'comparable inference latency,' but Table 6 reports SDAR-8B at 380s vs Qwen3.5-35B-A3B at 211s (1.80× slower); only WeDLM is faster. Please qualify the latency claim in the main text.
  5. [Appendix E] Diffusion steps are fixed at 50 for SDAR, with a note that 25–100 were similar; a sensitivity analysis with seeds would strengthen the claim that step count does not affect quality.

Circularity Check

0 steps flagged

No circularity found: the MDLM-vs-AR comparison is an empirical evaluation against real environment references; synthesized hindsight is a data-quality concern, not a derivation loop.

full rationale

No step in the paper derives its central result from a fitted parameter, a definitional identity, or a load-bearing self-citation. The headline claim—that MDLMs outperform AR LMs on world-modeling fidelity—is supported by MAUVE/BLEU/ROUGE computed against held-out next-state references from real environments. The paper states that trajectories were produced by setting up 'the original environment as designed by the authors' and generating rollouts with a variety of models, so the prediction targets are genuine environment outputs, not the model's own outputs or a fitted parameter. Downstream RL results are likewise evaluated against the true environment: 'Reward is computed from ground-truth environment state.' The only substantive concern raised—that Claude-4.6-Sonnet synthesizes hindsight context for ~80% of training data—is a potential train/evaluation contamination issue that could affect external validity, but it does not make the MDLM-vs-AR comparison reduce to its inputs by construction: both model families are trained and evaluated on the same references, and the comparison is an empirical measurement rather than a definitional identity. There are no self-citations by the present author that are load-bearing, no uniqueness theorem imported from prior work by the same authors, and no fitted parameter renamed as a prediction. The paper also explicitly acknowledges that surface-form metrics do not verify semantic equivalence and supplements them with human evaluation; that is a stated limitation, not a circular step. The derivation chain is therefore self-contained with respect to circularity, though external-validity questions about data contamination remain for the correctness pass.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central empirical comparison does not depend on a fitted free parameter, but several hand-chosen hyperparameters, reward shaping constants, and synthetic grounding choices affect the headline numbers. The axioms are mostly standard ML assumptions or scaffolding choices introduced specifically for this paper; the most fragile is the assumption that Claude-synthesized hindsight instructions are faithful enough to be called ground truth.

free parameters (5)
  • AR decoding temperature = 0.5
    Chosen by sweeping {0.5, 0.7, 0.9}; AR baselines are evaluated at 0.5 because it maximized MAUVE balance. MDLM sampling temperature is not reported, so the comparison may not be symmetric.
  • Diffusion inference steps and block size = 50 denoising steps; block size 4 (SDAR), 32 (LLaDA); W=6, lambda=0.10 (WeDLM)
    Selected by hand/prior recommendations (Appendix E); step counts 25 vs 100 shown equivalent for SDAR, so 50 is a latency/quality choice.
  • Reward shaping coefficients = ScienceWorld [-0.2, 0.1, 0.15]; ALFWorld 0.25/0.35/0.03/-0.05/0.3x; AppWorld deterministic
    Hand-designed in Appendix D to make small models learn; these constants shape the RL curves in Table 3 and may inflate absolute success rates.
  • Instruction augmentation rate = 80% augmented; 15% sub-trajectory truncation probability
    Chosen by hand in Appendix C to 'preserve approximative capacity'; no ablation isolates the augmentation rate on final metrics.
  • Data truncation threshold T_max = 16,384 tokens; Nmax formula Eq. 3
    Hand-set context budget; 2% of trajectories affected. Affects which trajectories are in the dataset.
axioms (5)
  • standard math MDLM loss in Eq. 2 is a valid upper bound on next-token NLL and the MDLM objective is equivalent to an any-order AR bound (citations [70,73]).
    Used to claim bidirectional conditioning and any-order steering; not re-derived in this paper.
  • domain assumption MAUVE with GPT-2 embeddings and canonicalized BLEU/ROUGE capture state-simulation fidelity.
    Used as primary evidence for coherence/groundedness; no human-verified semantic equivalence metric despite the paper's own caveat.
  • ad hoc to paper Claude-4.6-Sonnet-generated hindsight instructions and noise injection produce valid 'ground-truth' grounding for training/evaluation.
    Appendix C; if synthesized context is hallucinated, 'groundedness' claims are contaminated.
  • ad hoc to paper Hand-crafted WM prompts with baked room maps and deterministic local responders do not dominate the MDLM-vs-AR comparison.
    Appendix D reduces AppWorld WM to formatting and precomputes query state; the 'zero-shot' claim is conditional on this scaffolding.
  • domain assumption Four Upwork annotators rating 100 SDAR outputs on a 1-5 Likert scale are representative of training utility/realism.
    RQ3 evidence; small sample, no AR baseline, no independent validation of annotator expertise.
invented entities (1)
  • Hindsight instruction (grounding/steering components E, C, T, R, D) no independent evidence
    purpose: Augment each trajectory with synthesized environment context to train steerable prediction.
    A formatting/annotation construct, not a new physical entity; its validity rests entirely on Claude-generated content, so no independent falsifiable handle.

pith-pipeline@v1.3.0-alltime-deepseek · 33822 in / 21849 out tokens · 212631 ms · 2026-08-02T14:44:49.883151+00:00 · methodology

0 comments
read the original abstract

Recent growth in reinforcement learning (RL) has surfaced a need for diverse, specialized training environments. Hand-curated environments with fixed task and reward difficulties become ineffective signals as model performance improves, and sparse rewards over long horizons induce mode collapse on specific workflows or tool structures. World models that simulate environment states have matched pure rollout performance, making them promising for scaling diversity on-demand. However, autoregressive (AR) world models suffer from a left-to-right bias preventing conditioning on globally interdependent state anchors such as tool schemas, prior turns, and expected outcomes. We (i) formalize text-based world modeling as a steerable transition-dynamics problem decomposed into initial state, task context, tool schemas, domain rules, and steering directives, and (ii) curate 239,403 grounded state-action trajectories spanning nine open-source environments and twelve frontier model families. We compare AR LMs and masked diffusion language models (MDLMs), showing MDLMs, via bidirectional anchor-aware denoising, achieve better coherence, groundedness, and empirically validated rollout diversity than LLMs over 4x their parameter size, at comparable inference latency. We introduce a plug-and-play GRPO training framework with deterministic state checks, and perform zero-shot transfer ablations on three OOD environments (ScienceWorld, ALFWorld, AppWorld) across three 1.2B-7B agent backbones (LFM2.5, Qwen3, Mistral), achieving up to 47% absolute gains over baselines without environment-specific fine-tuning. We further conduct behavioral analysis of failure modes under adversarial scenarios and human evaluation on realism, outcome correctness, and training utility. We open-source our work to encourage research in this direction.

Figures

Figures reproduced from arXiv: 2607.16204 by Darshan Deshpande.

Figure 1
Figure 1. Figure 1: Anchor-aware structured generation: autoregressive vs. masked diffusion LMs. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Training split counts and token length distributions per dataset. The sampled number of [PITH_FULL_IMAGE:figures/full_fig_p028_2.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

147 extracted references · 55 linked inside Pith

  1. [1]

    Lfm2 technical report.arXiv preprint arXiv:2511.23404, 2025

    Liquid AI. Lfm2 technical report.arXiv preprint arXiv:2511.23404, 2025

    arXiv 2025

  2. [2]

    Large language models hallucination: A comprehensive survey, 2026

    Aisha Alansari and Hamzah Luqman. Large language models hallucination: A comprehensive survey, 2026. URLhttps://arxiv.org/abs/2510.06265

    arXiv 2026

  3. [3]

    A survey of llm-based methods for synthetic data generation and the rise of agentic workflows

    Ahmad Alismail and Carsten Lanquillon. A survey of llm-based methods for synthetic data generation and the rise of agentic workflows. InInternational Conference on Human-Computer Interaction, pages 119–135. Springer, 2025

    2025

  4. [4]

    Polaris: A post-training recipe for scaling reinforcement learning on advanced reasoning models, 2025

    Chenxin An, Zhihui Xie, Xiaonan Li, Lei Li, Jun Zhang, Shansan Gong, Ming Zhong, Jingjing Xu, Xipeng Qiu, Mingxuan Wang, and Lingpeng Kong. Polaris: A post-training recipe for scaling reinforcement learning on advanced reasoning models, 2025. URL https: //hkunlp.github.io/blog/2025/Polaris

    2025

  5. [5]

    Many-shot jailbreaking.Advances in Neural Information Processing Systems, 37:129696–129742, 2024

    Cem Anil, Esin Durmus, Nina Panickssery, Mrinank Sharma, Joe Benton, Sandipan Kundu, Joshua Batson, Meg Tong, Jesse Mu, Daniel Ford, et al. Many-shot jailbreaking.Advances in Neural Information Processing Systems, 37:129696–129742, 2024

    2024

  6. [6]

    Coderforge-preview: Sota open dataset for training efficient agents, February

    Alpay Ariyak, Junda Zhang, Junxiong Wang, Shang Zhu, Federico Bianchi, Sanjana Sri- vastava, Ashwinee Panda, Siddhant Bharti, Chenfeng Xu, John Heo, Xiaoxia Shirley Wu, James Zhou, Percy Liang, Leon Song, Ce Zhang, Ben Athiwaratkun, Zhongzhu Zhou, and Qingyang Wu. Coderforge-preview: Sota open dataset for training efficient agents, February

  7. [7]

    Block diffusion: Interpolating between autoregressive and diffusion language models

    Marianne Arriola, Aaron Gokaslan, Justin T Chiu, Zhihan Yang, Zhixuan Qi, Jiaqi Han, Subham Sekhar Sahoo, and V olodymyr Kuleshov. Block diffusion: Interpolating between autoregressive and diffusion language models. InThe Thirteenth International Conference on Learning Representations, 2025. URLhttps://arxiv.org/abs/2503.09573

    Pith/arXiv arXiv 2025

  8. [8]

    Abidegym: Turning static rl worlds into adaptive challenges.arXiv preprint arXiv:2509.21234, 2025

    Abi Aryan, Zac Liu, and Aaron Childress. Abidegym: Turning static rl worlds into adaptive challenges.arXiv preprint arXiv:2509.21234, 2025

    arXiv 2025

  9. [9]

    From masks to worlds: A hitchhiker’s guide to world models.arXiv preprint arXiv:2510.20668, 2025

    Jinbin Bai, Yu Lei, Hecong Wu, Yuchen Zhu, Shufan Li, Yi Xin, Xiangtai Li, Molei Tao, Aditya Grover, and Ming-Hsuan Yang. From masks to worlds: A hitchhiker’s guide to world models.arXiv preprint arXiv:2510.20668, 2025

    arXiv 2025

  10. [10]

    Philip J. Ball, Jakob Bauer, Frank Belletti, Bethanie Brownfield, Ariel Ephrat, Shlomi Fruchter, Agrim Gupta, Kristian Holsheimer, Aleksander Holynski, Jiri Hron, Christos Kaplanis, Mar- jorie Limont, Matt McGill, Yanko Oliveira, Jack Parker-Holder, Frank Perbet, Guy Scully, Jeremy Shar, Stephen Spencer, Omer Tov, Ruben Villegas, Emma Wang, Jessica Yung, ...

    2025

  11. [11]

    τ 2-bench: Evaluating conversational agents in a dual-control environment, 2025

    Victor Barres, Honghua Dong, Soham Ray, Xujie Si, and Karthik Narasimhan. τ 2-bench: Evaluating conversational agents in a dual-control environment, 2025. URL https://arxiv. org/abs/2506.07982

    Pith/arXiv arXiv 2025

  12. [12]

    a is b" fail to learn

    Lukas Berglund, Meg Tong, Max Kaufmann, Mikita Balesni, Asa Cooper Stickland, Tomasz Korbak, and Owain Evans. The reversal curse: Llms trained on "a is b" fail to learn "b is a",

  13. [13]

    Tiwei Bie, Maosong Cao, Kun Chen, Lun Du, Mingliang Gong, Zhuochen Gong, Yanmei Gu, Jiaqi Hu, Zenan Huang, Zhenzhong Lan, et al. Llada2. 0: Scaling up diffusion language models to 100b.arXiv preprint arXiv:2512.15745, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Tiwei Bie, Maosong Cao, Xiang Cao, Bingsen Chen, Fuyuan Chen, Kun Chen, Lun Du, Daozhuo Feng, Haibo Feng, Mingliang Gong, et al. Llada2. 1: Speeding up text diffusion via token editing.arXiv preprint arXiv:2602.08676, 2026

    arXiv 2026

  15. [15]

    Language models are few-shot learners.Advances in neural information processing systems, 33:1877– 1901, 2020

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

    1901

  16. [16]

    Sheetagent: towards a generalist agent for spreadsheet reasoning and manipulation via large language models

    Yibin Chen, Yifu Yuan, Zeyu Zhang, Yan Zheng, Jinyi Liu, Fei Ni, Jianye Hao, Hangyu Mao, and Fuzheng Zhang. Sheetagent: towards a generalist agent for spreadsheet reasoning and manipulation via large language models. InProceedings of the ACM on Web Conference 2025, pages 158–177, 2025

    2025

  17. [17]

    Scaling agent learning via experience synthesis, 2025

    Zhaorun Chen, Zhuokai Zhao, Kai Zhang, Bo Liu, Qi Qi, Yifan Wu, Tarun Kalluri, Sara Cao, Yuanhao Xiong, Haibo Tong, Huaxiu Yao, Hengduo Li, Jiacheng Zhu, Xian Li, Dawn Song, Bo Li, Jason Weston, and Dat Huynh. Scaling agent learning via experience synthesis, 2025. URLhttps://arxiv.org/abs/2511.03773

    arXiv 2025

  18. [18]

    Sdar: A synergistic diffusion-autoregression paradigm for scalable sequence generation.arXiv preprint arXiv:2510.06303, 2025

    Shuang Cheng, Yihan Bian, Dawei Liu, Linfeng Zhang, Qian Yao, Zhongbo Tian, Wenhai Wang, Qipeng Guo, Kai Chen, Biqing Qi, et al. Sdar: A synergistic diffusion-autoregression paradigm for scalable sequence generation.arXiv preprint arXiv:2510.06303, 2025

    arXiv 2025

  19. [19]

    Killian, Mikhail Yurochkin, Zhengzhong Liu, Eric P

    Zhoujun Cheng, Shibo Hao, Tianyang Liu, Fan Zhou, Yutao Xie, Feng Yao, Yuexin Bian, Yonghao Zhuang, Nilabjo Dey, Yuheng Zha, Yi Gu, Kun Zhou, Yuqi Wang, Yuan Li, Richard Fan, Jianshu She, Chengqian Gao, Abulhair Saparov, Haonan Li, Taylor W. Killian, Mikhail Yurochkin, Zhengzhong Liu, Eric P. Xing, and Zhiting Hu. Revisiting reinforcement learning for llm...

    Pith/arXiv arXiv 2025

  20. [20]

    Sft memorizes, rl generalizes: A comparative study of foundation model post-training.arXiv preprint arXiv:2501.17161, 2025

    Tianzhe Chu, Yuexiang Zhai, Jihan Yang, Shengbang Tong, Saining Xie, Dale Schuurmans, Quoc V Le, Sergey Levine, and Yi Ma. Sft memorizes, rl generalizes: A comparative study of foundation model post-training.arXiv preprint arXiv:2501.17161, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Lmdeploy: A toolkit for compressing, deploying, and serving llm

    LMDeploy Contributors. Lmdeploy: A toolkit for compressing, deploying, and serving llm. https://github.com/InternLM/lmdeploy, 2023

    2023

  22. [22]

    Process reinforcement through implicit rewards, 2025

    Ganqu Cui, Lifan Yuan, Zefan Wang, Hanbin Wang, Yuchen Zhang, Jiacheng Chen, Wendi Li, Bingxiang He, Yuchen Fan, Tianyu Yu, Qixin Xu, Weize Chen, Jiarui Yuan, Huayu Chen, Kaiyan Zhang, Xingtai Lv, Shuo Wang, Yuan Yao, Xu Han, Hao Peng, Yu Cheng, Zhiyuan Liu, Maosong Sun, Bowen Zhou, and Ning Ding. Process reinforcement through implicit rewards, 2025. URLh...

    Pith/arXiv arXiv 2025

  23. [24]

    Wang, Xiang Deng, Yuntao Ma, Nikhil Barhate, and Sean M

    Jeff Da, Clinton J. Wang, Xiang Deng, Yuntao Ma, Nikhil Barhate, and Sean M. Hendryx. Agent-rlvr: Training software engineering agents via guidance and environment rewards. ArXiv, abs/2506.11425, 2025. URL https://api.semanticscholar.org/CorpusID: 279391657

    Pith/arXiv arXiv 2025

  24. [25]

    Agentdojo: A dynamic environment to evaluate prompt injection attacks and defenses for LLM agents

    Edoardo Debenedetti, Jie Zhang, Mislav Balunovic, Luca Beurer-Kellner, Marc Fischer, and Florian Tramèr. Agentdojo: A dynamic environment to evaluate prompt injection attacks and defenses for LLM agents. InThe Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. URL https://openreview.net/forum? id=m1YYAQjO3w. 11

    2024

  25. [26]

    Chain-of-verification reduces hallucination in large language models

    Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. Chain-of-verification reduces hallucination in large language models. In Findings of the association for computational linguistics: ACL 2024, pages 3563–3578, 2024

    2024

  26. [27]

    Understanding world or predicting future? a comprehensive survey of world models.ACM Computing Surveys, 58(3):1–38, 2025

    Jingtao Ding, Yunke Zhang, Yu Shang, Yuheng Zhang, Zefang Zong, Jie Feng, Yuan Yuan, Hongyuan Su, Nian Li, Nicholas Sukiennik, et al. Understanding world or predicting future? a comprehensive survey of world models.ACM Computing Surveys, 58(3):1–38, 2025

    2025

  27. [28]

    Diffusion world model: Future modeling beyond step-by-step rollout for offline reinforcement learning.arXiv preprint arXiv:2402.03570, 2024

    Zihan Ding, Amy Zhang, Yuandong Tian, and Qinqing Zheng. Diffusion world model: Future modeling beyond step-by-step rollout for offline reinforcement learning.arXiv preprint arXiv:2402.03570, 2024

    Pith/arXiv arXiv 2024

  28. [29]

    A survey on code generation with llm-based agents.arXiv preprint arXiv:2508.00083, 2025

    Yihong Dong, Xue Jiang, Jiaru Qian, Tian Wang, Kechi Zhang, Zhi Jin, and Ge Li. A survey on code generation with llm-based agents.arXiv preprint arXiv:2508.00083, 2025

    Pith/arXiv arXiv 2025

  29. [30]

    Efficient-dlm: From autoregressive to diffusion language models, and beyond in speed, 2025

    Yonggan Fu, Lexington Whalen, Zhifan Ye, Xin Dong, Shizhe Diao, Jingyu Liu, Chengyue Wu, Hao Zhang, Enze Xie, Song Han, Maksim Khadkevich, Jan Kautz, Yingyan Celine Lin, and Pavlo Molchanov. Efficient-dlm: From autoregressive to diffusion language models, and beyond in speed, 2025. URLhttps://arxiv.org/abs/2512.14067

    Pith/arXiv arXiv 2025

  30. [31]

    Nextlong: Toward effective long-context training without long documents.arXiv preprint arXiv:2501.12766, 2025

    Chaochen Gao, Xing Wu, Zijia Lin, Debing Zhang, and Songlin Hu. Nextlong: Toward effective long-context training without long documents.arXiv preprint arXiv:2501.12766, 2025

    Pith/arXiv arXiv 2025

  31. [32]

    Vista: A generalizable driving world model with high fidelity and versatile controllability.Advances in Neural Information Processing Systems, 37:91560–91596, 2024

    Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability.Advances in Neural Information Processing Systems, 37:91560–91596, 2024

    2024

  32. [33]

    Scaling synthetic data creation with 1,000,000,000 personas.arXiv preprint arXiv:2406.20094, 2024

    Tao Ge, Xin Chan, Xiaoyang Wang, Dian Yu, Haitao Mi, and Dong Yu. Scaling synthetic data creation with 1,000,000,000 personas.arXiv preprint arXiv:2406.20094, 2024

    Pith/arXiv arXiv 2024

  33. [34]

    Synthetic data generation & multi-step rl for reasoning & tool use.arXiv preprint arXiv:2504.04736, 2025

    Anna Goldie, Azalia Mirhoseini, Hao Zhou, Irene Cai, and Christopher D Manning. Synthetic data generation & multi-step rl for reasoning & tool use.arXiv preprint arXiv:2504.04736, 2025

    Pith/arXiv arXiv 2025

  34. [35]

    Scaling diffusion language models via adaptation from autoregressive models, 2025

    Shansan Gong, Shivam Agarwal, Yizhe Zhang, Jiacheng Ye, Lin Zheng, Mukai Li, Chenxin An, Peilin Zhao, Wei Bi, Jiawei Han, Hao Peng, and Lingpeng Kong. Scaling diffusion language models via adaptation from autoregressive models, 2025. URL https://arxiv. org/abs/2410.17891

    Pith/arXiv arXiv 2025

  35. [36]

    Diffucoder: Understanding and improving masked diffusion models for code generation.arXiv preprint arXiv:2506.20639, 2025

    Shansan Gong, Ruixiang Zhang, Huangjie Zheng, Jiatao Gu, Navdeep Jaitly, Lingpeng Kong, and Yizhe Zhang. Diffucoder: Understanding and improving masked diffusion models for code generation.arXiv preprint arXiv:2506.20639, 2025

    Pith/arXiv arXiv 2025

  36. [37]

    Deepsearchqa: Bridging the comprehensiveness gap for deep research agents.arXiv preprint arXiv:2601.20975, 2026

    Nikita Gupta, Riju Chatterjee, Lukas Haas, Connie Tao, Andrew Wang, Chang Liu, Hidekazu Oiwa, Elena Gribovskaya, Jan Ackermann, John Blitzer, et al. Deepsearchqa: Bridging the comprehensiveness gap for deep research agents.arXiv preprint arXiv:2601.20975, 2026

    arXiv 2026

  37. [38]

    Mastering diverse domains through world models.arXiv preprint arXiv:2301.04104, 2023

    Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models.arXiv preprint arXiv:2301.04104, 2023

    Pith/arXiv arXiv 2023

  38. [39]

    Resyn: Autonomously scaling synthetic environments for reasoning models, 2026

    Andre He, Nathaniel Weir, Kaj Bostrom, Allen Nie, Darion Cassel, Sam Bayless, and Huzefa Rangwala. Resyn: Autonomously scaling synthetic environments for reasoning models, 2026. URLhttps://arxiv.org/abs/2602.20117

    arXiv 2026

  39. [40]

    Pre-trained video generative models as world simulators.arXiv preprint arXiv:2502.07825, 2025

    Haoran He, Yang Zhang, Liang Lin, Zhongwen Xu, and Ling Pan. Pre-trained video generative models as world simulators.arXiv preprint arXiv:2502.07825, 2025

    Pith/arXiv arXiv 2025

  40. [41]

    Scaling instruction-tuned llms to million-token contexts via hierarchical synthetic data generation

    Linda He, Jue Wang, Maurice Weber, Shang Zhu, Ben Athiwaratkun, and Ce Zhang. Scaling instruction-tuned llms to million-token contexts via hierarchical synthetic data generation. arXiv preprint arXiv:2504.12637, 2025. 12

    Pith/arXiv arXiv 2025

  41. [42]

    Llms gaming verifiers: Rlvr can lead to reward hacking, 2026

    Lukas Helff, Quentin Delfosse, David Steinmann, Ruben Härle, Hikaru Shindo, Patrick Schramowski, Wolfgang Stammer, Kristian Kersting, and Felix Friedrich. Llms gaming verifiers: Rlvr can lead to reward hacking, 2026. URL https://arxiv.org/abs/2604. 15149

    2026

  42. [43]

    Occubench: Evaluating ai agents on real-world professional tasks via language world models.arXiv preprint arXiv:2604.10866, 2026

    Xiaomeng Hu, Yinger Zhang, Fei Huang, Jianhong Tu, Yang Su, Lianghao Deng, Yuxuan Liu, Yantao Liu, Dayiheng Liu, and Tsung-Yi Ho. Occubench: Evaluating ai agents on real-world professional tasks via language world models.arXiv preprint arXiv:2604.10866, 2026

    Pith/arXiv arXiv 2026

  43. [44]

    OS agents: A survey on MLLM-based agents for computer, phone and browser use

    Xueyu Hu, Tao Xiong, Biao Yi, Zishu Wei, Ruixuan Xiao, Yurun Chen, Jiasheng Ye, Meiling Tao, Xiangxin Zhou, Ziyu Zhao, Yuhuai Li, Shengze Xu, Shenzhi Wang, Xinchen Xu, Shuofei Qiao, Zhaokai Wang, Kun Kuang, Tieyong Zeng, Liang Wang, Jiwei Li, Yuchen Eleanor Jiang, Wangchunshu Zhou, Guoyin Wang, Keting Yin, Zhou Zhao, Hongxia Yang, Fan Wu, Shengyu Zhang, a...

    2025

  44. [45]

    Mistral 7b

    Albert Q Jiang, A Sablayrolles, A Mensch, C Bamford, D Singh Chaplot, Ddl Casas, F Bressand, G Lengyel, G Lample, L Saulnier, et al. Mistral 7b. arxiv.arXiv preprint arXiv:2310.06825, 10:3, 2023

    Pith/arXiv arXiv 2023

  45. [46]

    Swe-bench: Can language models resolve real-world github issues? arXiv preprint arXiv:2310.06770, 2023

    Carlos E Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. Swe-bench: Can language models resolve real-world github issues? arXiv preprint arXiv:2310.06770, 2023

    Pith/arXiv arXiv 2023

  46. [47]

    Multi-objective instruction-aware representation learning in procedural content genera- tion rl.arXiv preprint arXiv:2508.09193, 2025

    Sung-Hyun Kim, In-Chang Baek, Seo-Young Lee, Geum-Hwan Hwang, and Kyung-Joong Kim. Multi-objective instruction-aware representation learning in procedural content genera- tion rl.arXiv preprint arXiv:2508.09193, 2025

    Pith/arXiv arXiv 2025

  47. [48]

    Computing krippendorff’s alpha-reliability

    Klaus Krippendorff. Computing krippendorff’s alpha-reliability. 2011

    2011

  48. [49]

    Fast inference from transformers via speculative decoding, 2023.URL https://arxiv

    Yaniv Leviathan, Matan Kalman, and Yossi Matias. Fast inference from transformers via speculative decoding, 2023.URL https://arxiv. org/abs/2211.17192, 1(2), 2022

    Pith/arXiv arXiv 2023

  49. [50]

    Open-world reinforcement learning over long short-term imagination.arXiv preprint arXiv:2410.03618, 2024

    Jiajian Li, Qi Wang, Yunbo Wang, Xin Jin, Yang Li, Wenjun Zeng, and Xiaokang Yang. Open-world reinforcement learning over long short-term imagination.arXiv preprint arXiv:2410.03618, 2024

    arXiv 2024

  50. [51]

    Wildlong: Synthesizing realistic long-context instruction data at scale

    Jiaxi Li, Xingxing Zhang, Xun Wang, Xiaolong Huang, Li Dong, Liang Wang, Si-Qing Chen, Wei Lu, and Furu Wei. Wildlong: Synthesizing realistic long-context instruction data at scale. arXiv preprint arXiv:2502.16684, 2025

    Pith/arXiv arXiv 2025

  51. [52]

    A diversity-promoting objective function for neural conversation models

    Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. A diversity-promoting objective function for neural conversation models. corr abs/1510.03055.arXiv preprint arxiv:5100.3055, 2015

    Pith/arXiv arXiv 2015

  52. [53]

    The tool decathlon: Benchmarking language agents for diverse, realistic, and long-horizon task execution

    Junlong Li, Wenshuo Zhao, Jian Zhao, Weihao Zeng, Haoze Wu, Xiaochen Wang, Rui Ge, Yuxuan Cao, Yuzhen Huang, Wei Liu, Junteng Liu, Zhaochen Su, Yiyang Guo, Fan Zhou, Lueyang Zhang, Juan Michelini, Xingyao Wang, Xiang Yue, Shuyan Zhou, Graham Neubig, and Junxian He. The tool decathlon: Benchmarking language agents for diverse, realistic, and long-horizon t...

    arXiv 2025

  53. [54]

    Api-bank: A comprehensive benchmark for tool-augmented llms

    Minghao Li, Yingxiu Zhao, Bowen Yu, Feifan Song, Hangyu Li, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. Api-bank: A comprehensive benchmark for tool-augmented llms. In Proceedings of the 2023 conference on empirical methods in natural language processing, pages 3102–3116, 2023

    2023

  54. [55]

    From word to world: Can large language models be implicit text-based world models?arXiv preprint arXiv:2512.18832, 2025

    Yixia Li, Hongru Wang, Jiahao Qiu, Zhenfei Yin, Dongdong Zhang, Cheng Qian, Zeping Li, Pony Ma, Guanhua Chen, Heng Ji, et al. From word to world: Can large language models be implicit text-based world models?arXiv preprint arXiv:2512.18832, 2025. 13

    arXiv 2025

  55. [56]

    Openresearcher: A fully open pipeline for long-horizon deep research trajectory synthesis.arXiv preprint arXiv:2603.20278, 2026

    Zhuofeng Li, Dongfu Jiang, Xueguang Ma, Haoxiang Zhang, Ping Nie, Yuyu Zhang, Kai Zou, Jianwen Xie, Yu Zhang, and Wenhu Chen. Openresearcher: A fully open pipeline for long-horizon deep research trajectory synthesis.arXiv preprint arXiv:2603.20278, 2026

    arXiv 2026

  56. [57]

    Eurekaverse: Environment curriculum generation via large language models.arXiv preprint arXiv:2411.01775, 2024

    William Liang, Sam Wang, Hung-Ju Wang, Osbert Bastani, Dinesh Jayaraman, and Yecheng Ja- son Ma. Eurekaverse: Environment curriculum generation via large language models.arXiv preprint arXiv:2411.01775, 2024

    Pith/arXiv arXiv 2024

  57. [58]

    Wedlm: Reconciling diffusion language models with standard causal attention for fast inference.arXiv preprint arXiv:2512.22737, 2025

    Aiwei Liu, Minghua He, Shaoxun Zeng, Sijun Zhang, Linhao Zhang, Chuhan Wu, Wei Jia, Yuan Liu, Xiao Zhou, and Jie Zhou. Wedlm: Reconciling diffusion language models with standard causal attention for fast inference.arXiv preprint arXiv:2512.22737, 2025

    arXiv 2025

  58. [59]

    Apigen: Automated pipeline for generating verifiable and diverse function-calling datasets.Advances in Neural Information Processing Systems, 37:54463–54482, 2024

    Zuxin Liu, Thai Hoang, Jianguo Zhang, Ming Zhu, Tian Lan, Shirley Kokane, Juntao Tan, Weiran Yao, Zhiwei Liu, Yihao Feng, et al. Apigen: Automated pipeline for generating verifiable and diverse function-calling datasets.Advances in Neural Information Processing Systems, 37:54463–54482, 2024

    2024

  59. [60]

    On llms-driven synthetic data generation, curation, and evaluation: A survey

    Lin Long, Rui Wang, Ruixuan Xiao, Junbo Zhao, Xiao Ding, Gang Chen, and Haobo Wang. On llms-driven synthetic data generation, curation, and evaluation: A survey. InFindings of the Association for Computational Linguistics: ACL 2024, pages 11065–11082, 2024

    2024

  60. [61]

    Source2synth: Synthetic data generation and curation grounded in real data sources.arXiv preprint arXiv:2409.08239, 2024

    Alisia Lupidi, Carlos Gemmell, Nicola Cancedda, Jane Dwivedi-Yu, Jason Weston, Jakob Foerster, Roberta Raileanu, and Maria Lomeli. Source2synth: Synthetic data generation and curation grounded in real data sources.arXiv preprint arXiv:2409.08239, 2024

    Pith/arXiv arXiv 2024

  61. [62]

    Mock worlds, real skills: Building small agentic language models with synthetic tasks, simulated environments, and rubric-based rewards, 2026

    Yuanjie Lyu, Chengyu Wang, Lei Shen, Jun Huang, and Tong Xu. Mock worlds, real skills: Building small agentic language models with synthetic tasks, simulated environments, and rubric-based rewards, 2026. URLhttps://arxiv.org/abs/2601.22511

    arXiv 2026

  62. [63]

    Rephrasing the web: A recipe for compute and data-efficient language modeling

    Pratyush Maini, Skyler Seto, Richard Bai, David Grangier, Yizhe Zhang, and Navdeep Jaitly. Rephrasing the web: A recipe for compute and data-efficient language modeling. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14044–14072, 2024

    2024

  63. [64]

    Active domain randomization

    Bhairav Mehta, Manfred Diaz, Florian Golemo, Christopher J Pal, and Liam Paull. Active domain randomization. InConference on Robot Learning, pages 1162–1176. PMLR, 2020

    2020

  64. [65]

    Terminal-bench: Benchmarking agents on hard, realistic tasks in command line interfaces.arXiv preprint arXiv:2601.11868, 2026

    Mike A Merrill, Alexander G Shaw, Nicholas Carlini, Boxuan Li, Harsh Raj, Ivan Bercovich, Lin Shi, Jeong Yeon Shin, Thomas Walshe, E Kelly Buchanan, et al. Terminal-bench: Benchmarking agents on hard, realistic tasks in command line interfaces.arXiv preprint arXiv:2601.11868, 2026

    Pith/arXiv arXiv 2026

  65. [66]

    Agentinstruct: Toward generative teaching with agentic flows.arXiv preprint arXiv:2407.03502, 2024

    Arindam Mitra, Luciano Del Corro, Guoqing Zheng, Shweti Mahajan, Dany Rouhana, Andres Codas, Yadong Lu, Wei-ge Chen, Olga Vrousgos, Corby Rosset, et al. Agentinstruct: Toward generative teaching with agentic flows.arXiv preprint arXiv:2407.03502, 2024

    Pith/arXiv arXiv 2024

  66. [67]

    Jointly measuring diversity and quality in text generation models, 2019

    Ehsan Montahaei, Danial Alihosseini, and Mahdieh Soleymani Baghshah. Jointly measuring diversity and quality in text generation models, 2019. URL https://arxiv.org/abs/1904. 03971

    2019

  67. [68]

    Synthetic data generation using large language models: Advances in text and code

    M Nadas, L Diosan, and A Tomescu. Synthetic data generation using large language models: Advances in text and code. arxiv 2025.arXiv preprint arXiv:2503.14023, 2025

    arXiv 2025

  68. [69]

    Scaling up masked diffusion models on text, 2025

    Shen Nie, Fengqi Zhu, Chao Du, Tianyu Pang, Qian Liu, Guangtao Zeng, Min Lin, and Chongxuan Li. Scaling up masked diffusion models on text, 2025. URL https://arxiv. org/abs/2410.18514

    Pith/arXiv arXiv 2025

  69. [70]

    Large language diffusion models, 2025

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

    Pith/arXiv arXiv 2025

  70. [71]

    Nemotron 3 Nano: Open, efficient mixture-of-experts hybrid Mamba-Transformer model for Agentic reasoning, 2025

    NVIDIA. Nemotron 3 Nano: Open, efficient mixture-of-experts hybrid Mamba-Transformer model for Agentic reasoning, 2025. URL https://arxiv.org/abs/2512.20848. Techni- cal report. 14

    arXiv 2025

  71. [72]

    gpt-oss-120b & gpt-oss-20b model card, 2025

    OpenAI. gpt-oss-120b & gpt-oss-20b model card, 2025. URL https://arxiv.org/abs/ 2508.10925

    Pith/arXiv arXiv 2025

  72. [73]

    Your absorbing discrete diffusion secretly models the conditional distributions of clean data.arXiv preprint arXiv:2406.03736, 2024

    Jingyang Ou, Shen Nie, Kaiwen Xue, Fengqi Zhu, Jiacheng Sun, Zhenguo Li, and Chongxuan Li. Your absorbing discrete diffusion secretly models the conditional distributions of clean data.arXiv preprint arXiv:2406.03736, 2024

    Pith/arXiv arXiv 2024

  73. [74]

    A survey of reinforcement learning algorithms for dynamically varying environments.ACM Computing Surveys (CSUR), 54(6):1–25, 2021

    Sindhu Padakandla. A survey of reinforcement learning algorithms for dynamically varying environments.ACM Computing Surveys (CSUR), 54(6):1–25, 2021

    2021

  74. [75]

    Patil, Huanzhi Mao, Charlie Cheng-Jie Ji, Fanjia Yan, Vishnu Suresh, Ion Stoica, and Joseph E

    Shishir G. Patil, Huanzhi Mao, Charlie Cheng-Jie Ji, Fanjia Yan, Vishnu Suresh, Ion Stoica, and Joseph E. Gonzalez. The berkeley function calling leaderboard (bfcl): From tool use to agentic evaluation of large language models. InForty-second International Conference on Machine Learning, 2025

    2025

  75. [76]

    Mauve: Measuring the gap between neural text and human text using divergence frontiers.Advances in Neural Information Processing Systems, 34: 4816–4828, 2021

    Krishna Pillutla, Swabha Swayamdipta, Rowan Zellers, John Thickstun, Sean Welleck, Yejin Choi, and Zaid Harchaoui. Mauve: Measuring the gap between neural text and human text using divergence frontiers.Advances in Neural Information Processing Systems, 34: 4816–4828, 2021

    2021

  76. [77]

    MAUVE Scores for Generative Models: Theory and Practice.JMLR, 2023

    Krishna Pillutla, Lang Liu, John Thickstun, Sean Welleck, Swabha Swayamdipta, Rowan Zellers, Sewoong Oh, Yejin Choi, and Zaid Harchaoui. MAUVE Scores for Generative Models: Theory and Practice.JMLR, 2023

    2023

  77. [78]

    Apigen-mt: Agentic pipeline for multi-turn data generation via simulated agent-human interplay.arXiv preprint arXiv:2504.03601, 2025

    Akshara Prabhakar, Zuxin Liu, Ming Zhu, Jianguo Zhang, Tulika Awalgaonkar, Shiyu Wang, Zhiwei Liu, Haolin Chen, Thai Hoang, Juan Carlos Niebles, et al. Apigen-mt: Agentic pipeline for multi-turn data generation via simulated agent-human interplay.arXiv preprint arXiv:2504.03601, 2025

    Pith/arXiv arXiv 2025

  78. [79]

    Androidworld: A dynamic benchmarking environment for autonomous agents, 2024

    Christopher Rawles, Sarah Clinckemaillie, Yifan Chang, Jonathan Waltz, Gabrielle Lau, Marybeth Fair, Alice Li, William Bishop, Wei Li, Folawiyo Campbell-Ajala, Daniel Toyama, Robert Berry, Divya Tyamagundlu, Timothy Lillicrap, and Oriana Riva. Androidworld: A dynamic benchmarking environment for autonomous agents, 2024. URL https://arxiv. org/abs/2405.14573

    Pith/arXiv arXiv 2024

  79. [80]

    Metasynth: Meta-prompting-driven agentic scaffolds for diverse synthetic data generation

    Haris Riaz, Sourav Sanjukta Bhabesh, Vinayak Arannil, Miguel Ballesteros, and Graham Horwood. Metasynth: Meta-prompting-driven agentic scaffolds for diverse synthetic data generation. InFindings of the Association for Computational Linguistics: ACL 2025, page 18770–18803. Association for Computational Linguistics, 2025. doi: 10.18653/v1/2025. findings-acl...

    doi:10.18653/v1/2025 2025

  80. [81]

    Simple and effective masked diffusion language models.Advances in Neural Information Processing Systems, 37:130136–130184, 2024

    Subham S Sahoo, Marianne Arriola, Yair Schiff, Aaron Gokaslan, Edgar Marroquin, Justin T Chiu, Alexander Rush, and V olodymyr Kuleshov. Simple and effective masked diffusion language models.Advances in Neural Information Processing Systems, 37:130136–130184, 2024

    2024

Showing first 80 references.