pith. sign in

arxiv: 2606.03243 · v1 · pith:2PS4CWYWnew · submitted 2026-06-02 · 💻 cs.CV

MemoGen: Can Past Experience Improve Future Text-to-Image Generation?

Wenshuo Chen , Kuimou Yu , Bowen Tian , Jianfei Song , Shaofeng Liang , Haozhe Jia , Kan Cheng , Haosen Li
show 5 more authors
Kaishen Yuan Lei Wang Jiemin Wu Songning Lai Yutao Yue
This is my paper

Pith reviewed 2026-06-28 10:46 UTC · model grok-4.3

classification 💻 cs.CV
keywords text-to-image generationexperience memoryagentic evolutioncontinual improvementtraining-free frameworkreasoning benchmarksself-evolution
0
0 comments X

The pith

Storing and reusing experience from past generations lets text-to-image models improve on complex prompts without retraining.

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

The paper asks whether a text-to-image system can get better at difficult requests by keeping records of what worked and what failed before. MemoGen adds an agent layer that figures out hidden visual needs, gathers references, generates an image, judges the result, and saves the understanding, choices, feedback, strategies, and lessons as reusable memory. On later similar tasks the agent pulls the right memories to guide the generator, fixing earlier mistakes while leaving good outputs alone. The process runs without changing the underlying model weights and shows gains after only two rounds on benchmarks that test knowledge and reasoning.

Core claim

MemoGen is a training-free framework that augments any image generator with an agentic evolution layer. For each task the agent infers visual requirements, retrieves evidence when needed, translates it into generation constraints, evaluates the output, and stores task understanding, reference choices, visual feedback, successful strategies, and failure lessons as experience memory. Across evolution rounds the agent retrieves relevant stored experience to improve new generations, selectively repairing previously failed cases while preserving successful ones.

What carries the argument

Experience memory, which records inferred visual requirements, chosen references, evaluations, successful strategies, and failure lessons so the agent can retrieve and apply them to future similar tasks.

If this is right

  • An open-source backbone reaches higher scores than proprietary systems on WISE and Mind-Bench after only two rounds.
  • The system improves failed cases from earlier rounds while leaving successful generations unchanged.
  • Continual improvement occurs through memory retrieval alone, with no parameter updates to the generator.
  • The approach directly addresses prompts that need implicit constraints, relational reasoning, or external knowledge.

Where Pith is reading between the lines

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

  • The same memory-retrieval pattern could be added to other generative models such as video or 3-D synthesis.
  • If memory entries grow large, compression or relevance-ranking rules would become necessary to keep retrieval accurate.
  • One could measure whether the stored lessons remain useful when the generator backbone itself is swapped for a newer model.
  • The framework suggests a route to test-time adaptation that avoids the cost of periodic full retraining.

Load-bearing premise

The agent can reliably infer visual requirements, judge image quality, and extract transferable strategies and lessons without introducing systematic errors or needing human correction.

What would settle it

A controlled test on the same benchmarks where performance after two or more evolution rounds is no better than the first round or where the agent repeatedly selects harmful memory entries for similar prompts.

Figures

Figures reproduced from arXiv: 2606.03243 by Bowen Tian, Haosen Li, Haozhe Jia, Jianfei Song, Jiemin Wu, Kaishen Yuan, Kan Cheng, Kuimou Yu, Lei Wang, Shaofeng Liang, Songning Lai, Wenshuo Chen, Yutao Yue.

Figure 1
Figure 1. Figure 1: Overview of MemoGen. MemoGen formulates image generation as a test-time self-evolution process. This framework enables the system to continuously improve future generations by storing and reusing task understanding, retrieved evidence, reference choices, visual feedback, and success or failure experiences, all without updating the underlying model. into future visual constraints—missing objects, incorrect … view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative comparison. The results in illustrate that MemoGen ("Ours") effectively follows complex prompts and adheres to implicit constraints where other models struggle, such as correctly depicting an unlit candle in space or specific product endorsements. IP, and WK (0.60, 0.72, and 0.82), and Reasoning-Driven tasks, such as SL, Life Reasoning, and GU (0.74, 0.42, and 0.20). This indicates that memory … view at source ↗
Figure 3
Figure 3. Figure 3: Category-wise WISE pass rates from the Qwen Image baseline to three evolving rounds. The dashed gray line indicates the overall score trend. underlying Internet-meme association, rendering the result in a rec￾ognizable meme-like visual form. In the chemistry example, only MemoGen correctly captures the color and physical state of potassium permanganate dissolved in water, producing a homogeneous purple sol… view at source ↗
read the original abstract

Modern text-to-image models have achieved strong visual synthesis, yet remain unreliable when prompts require implicit visual constraints, relational reasoning, or external knowledge. Existing retrieval-augmented and agentic generation methods mitigate this issue by acquiring external knowledge, references, or refined prompts for the current request, yet they typically treat each generation as an isolated episode and do not systematically preserve past successes or failures for future use. In this work, we ask whether a text-to-image system can continually improve from its own generation experience without updating the underlying generator. We propose MemoGen, a training-free framework that augments existing image generators with an agentic evolution layer. For each task, MemoGen explicitly infers visual requirements, retrieves external evidence and references when necessary, translates them into executable generation constraints, evaluates the generated result, and stores task understanding, reference choices, visual feedback, successful strategies, and failure lessons as reusable experience memory. Across evolution rounds, the agent retrieves relevant experience to improve similar future generations, selectively repairing previously failed cases while preserving successful ones, thereby enabling test-time self-evolution without parameter updates. Extensive experiments on knowledge-intensive and reasoning-oriented benchmarks demonstrate the effectiveness of this paradigm: after only two evolution rounds, MemoGen built upon the open-source Qwen-Image backbone surpasses strong proprietary systems such as Nano Banana Pro and GPT-Image-1 on WISE and Mind-Bench, showing that explicit experience memory can serve as a powerful continual learning signal for reliable text-to-image generation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes MemoGen, a training-free agentic framework that augments text-to-image generators with an evolution layer. For each task the agent infers visual requirements, retrieves references, translates them into constraints, evaluates the output, and stores task understanding, strategies, and failure lessons as reusable experience memory. Across rounds the memory is retrieved to improve similar future generations. The central empirical claim is that after only two evolution rounds the Qwen-Image backbone augmented by MemoGen surpasses proprietary systems (Nano Banana Pro, GPT-Image-1) on the WISE and Mind-Bench benchmarks.

Significance. If the reported gains are robustly supported by detailed experimental protocols, the work would be significant: it demonstrates that explicit, reusable experience memory can serve as a continual-learning signal for T2I models without any parameter updates, addressing a recognized weakness of current generators on relational, knowledge-intensive, and constraint-heavy prompts.

major comments (2)
  1. [Abstract / §3–4] Abstract (and §3–4, per the described pipeline): the headline result—that two evolution rounds suffice to surpass proprietary models—rests on the unverified assumption that the agent produces reliable scalar/textual evaluations and extracts transferable lessons. No mechanism, prompt template, inter-annotator agreement, or correlation study with human judgments on relational/knowledge-intensive prompts is supplied; if LLM-based evaluation correlates poorly, retrieved memories will reinforce rather than correct errors, rendering the evolution loop non-functional.
  2. [Abstract] Abstract: the claim of surpassing Nano Banana Pro and GPT-Image-1 on WISE and Mind-Bench after two rounds is presented without any description of evaluation procedure, exact metrics, data splits, number of runs, or statistical significance testing. This information is load-bearing for the central empirical claim and must be supplied before the result can be assessed.
minor comments (1)
  1. The abstract refers to “extensive experiments” but supplies no table or figure numbers; the full manuscript should include a dedicated results section with per-benchmark breakdowns and ablations on the memory-retrieval and evaluation components.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the evaluation procedures and empirical claims. We address each point below and will revise the manuscript to supply the requested details on mechanisms, prompts, and experimental protocols.

read point-by-point responses

  1. Referee: [Abstract / §3–4] Abstract (and §3–4, per the described pipeline): the headline result—that two evolution rounds suffice to surpass proprietary models—rests on the unverified assumption that the agent produces reliable scalar/textual evaluations and extracts transferable lessons. No mechanism, prompt template, inter-annotator agreement, or correlation study with human judgments on relational/knowledge-intensive prompts is supplied; if LLM-based evaluation correlates poorly, retrieved memories will reinforce rather than correct errors, rendering the evolution loop non-functional.

    Authors: We agree that the manuscript would benefit from explicit documentation of the evaluation mechanism. MemoGen's agent employs LLM-based scoring and lesson extraction via structured prompts that output both scalar scores and textual feedback; these prompts were omitted from the initial submission. In revision we will include the full prompt templates in an appendix and add a correlation study with human judgments on 50 relational/knowledge-intensive prompts, reporting agreement statistics. This directly addresses the risk of error reinforcement. revision: yes

  2. Referee: [Abstract] Abstract: the claim of surpassing Nano Banana Pro and GPT-Image-1 on WISE and Mind-Bench after two rounds is presented without any description of evaluation procedure, exact metrics, data splits, number of runs, or statistical significance testing. This information is load-bearing for the central empirical claim and must be supplied before the result can be assessed.

    Authors: We acknowledge that the abstract and experimental section lack these protocol details. The reported results follow the official WISE and Mind-Bench evaluation pipelines using their published metrics; experiments were run three times with different random seeds, and we will report means, standard deviations, and statistical significance (paired t-tests) in the revised manuscript. Data splits are exactly those released with the benchmarks. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical framework without derivations or self-referential reductions

full rationale

The paper describes MemoGen as a training-free agentic framework that infers requirements, evaluates outputs, stores memories, and retrieves them across rounds to improve generations on benchmarks like WISE and Mind-Bench. No equations, fitted parameters, predictions derived from inputs, uniqueness theorems, or self-citations appear in the provided text. The central result is an empirical claim of improvement after two rounds on an open-source backbone, which does not reduce to its own inputs by construction and remains independent of any internal derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The framework rests on the domain assumption that an external agent can accurately perform inference, retrieval, constraint translation, evaluation, and memory curation without introducing bias; the experience memory itself is an invented entity with no independent evidence supplied in the abstract.

axioms (1)
  • domain assumption Existing text-to-image generators can be prompted and evaluated by an external agent without internal parameter changes.
    This is required for the training-free claim and is invoked in the description of the evolution layer.
invented entities (1)
  • Experience memory store no independent evidence
    purpose: Holds task understanding, references, visual feedback, successful strategies, and failure lessons for future retrieval.
    New component introduced by the paper; no independent evidence of its correctness or transferability is given.

pith-pipeline@v0.9.1-grok · 5843 in / 1300 out tokens · 30879 ms · 2026-06-28T10:46:06.944202+00:00 · methodology

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

58 extracted references · 15 canonical work pages · 6 internal anchors

  1. [1]

    WIT:Wikipedia-basedimagetextdatasetformultimodal multilingualmachinelearning

    K. Srinivasan, K. Raman, J. Chen, M. Bendersky, and M. Na- jork,“WIT:Wikipedia-basedimagetextdatasetformultimodal multilingualmachinelearning”,inProceedingsofthe44thInter- nationalACMSIGIRConferenceonResearchandDevelopment inInformationRetrieval,2021,pp.2443–2449

    2021

  2. [2]

    Re-imagen: Retrieval- augmented text-to-image generator.arXiv preprint arXiv:2209.14491, 2022

    W. Chen, H. Hu, C. Saharia, and W. W. Cohen, “Re-imagen: Retrieval-augmentedtext-to-imagegenerator”,arXivpreprint arXiv:2209.14491,2022

    work page arXiv 2022

  3. [3]

    Traininglanguagemodelstofollowinstruc- tionswithhumanfeedback

    L.Ouyangetal.,“Traininglanguagemodelstofollowinstruc- tionswithhumanfeedback”,Advancesinneuralinformation processingsystems,vol.35,pp.27730–27744,2022

    2022

  4. [4]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Om- mer,High-resolutionimagesynthesiswithlatentdiffusionmodels, 2022

    2022

  5. [5]

    Saharia et al.,Photorealistic text-to-image diffusion models withdeeplanguageunderstanding,2022

    C. Saharia et al.,Photorealistic text-to-image diffusion models withdeeplanguageunderstanding,2022

    2022

  6. [6]

    kNN-Diffusion:Imagegenerationvialarge- scaleretrieval

    S.Sheyninetal.,“kNN-Diffusion:Imagegenerationvialarge- scaleretrieval”,arXivpreprintarXiv:2204.02849,2022

    work page arXiv 2022

  7. [7]

    J.Yuetal.,Scalingautoregressivemodelsforcontent-richtext-to- imagegeneration,2022

    2022

  8. [8]

    J.Betkeretal.,Improvingimagegenerationwithbettercaptions, https://cdn.openai.com/papers/dall-e-3.pdf, Accessed: 2026-05-21,2023

    2026

  9. [9]

    Bernstein,Generative agents: Interactive simulacra of human behavior,2023

    J.S.Park,J.C.O’Brien,C.J.Cai,M.R.Morris,P.Liang,andM.S. Bernstein,Generative agents: Interactive simulacra of human behavior,2023

    2023

  10. [10]

    D.Podelletal.,Sdxl:Improvinglatentdiffusionmodelsforhigh- resolutionimagesynthesis,2023

    2023

  11. [11]

    Reflexion: Language Agents with Verbal Reinforcement Learning

    N.Shinn,F.Cassano,A.Gopinath,K.Narasimhan,andS.Yao, “Reflexion:Languageagentswithverbalreinforcementlearn- ing”,arXivpreprintarXiv:2303.11366,2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  12. [12]

    Voyager: An Open-Ended Embodied Agent with Large Language Models

    G.Wangetal.,“Voyager:Anopen-endedembodiedagentwith largelanguagemodels”,arXivpreprintarXiv:2305.16291,2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  13. [13]

    Retrieval-augmentedmultimodallanguage modeling

    M.Yasunagaetal.,“Retrieval-augmentedmultimodallanguage modeling”,inInternationalConferenceonMachineLearning, 2023

    2023

  14. [14]

    H.Ye,J.Zhang,S.Liu,X.Han,andW.Yang,Ip-adapter:Text compatible image prompt adapter for text-to-image diffusion models,2023

    2023

  15. [15]

    L.Zhang,A.Rao,andM.Agrawala,Addingconditionalcontrol totext-to-imagediffusionmodels,2023

    2023

  16. [16]

    Black Forest Labs,FLUX.1, https://github.com/black-forest- labs/flux,Accessed:2026-05-21,2024

    2026

  17. [17]

    Sato: Stable text-to-motion framework

    W. Chen et al., “Sato: Stable text-to-motion framework”, in Proceedingsofthe32ndACMInternationalConferenceonMulti- media,2024,pp.6989–6997.doi:10.1145/3664647.3681034

    work page doi:10.1145/3664647.3681034 2024

  18. [18]

    Y.Liangetal.,Richhumanfeedbackfortext-to-imagegeneration, 2024

    2024

  19. [19]

    Mou et al.,T2i-adapter: Learning adapters to dig out more controllableabilityfortext-to-imagediffusionmodels,2024

    C. Mou et al.,T2i-adapter: Learning adapters to dig out more controllableabilityfortext-to-imagediffusionmodels,2024

    2024

  20. [20]

    StabilityAI,Stablediffusion3.5large,https://stability.ai/news- updates/introducing-stable-diffusion-3-5,Accessed:2026-05- 21,2024

    2026

  21. [21]

    co/stabilityai/stable-diffusion-3.5-medium,Accessed:2026-05- 21,2024

    StabilityAI,Stablediffusion3.5medium,https://huggingface. co/stabilityai/stable-diffusion-3.5-medium,Accessed:2026-05- 21,2024

    2026

  22. [22]

    Z.Wang,A.Li,Z.Li,andX.Liu,Genartist:Multimodalllmas anagentforunifiedimagegenerationandediting,2024

    2024

  23. [23]

    Yang et al.,Idea2img: Iterative self-refinement with gpt- 4v(ision)forautomaticimagedesignandgeneration,2024

    Z. Yang et al.,Idea2img: Iterative self-refinement with gpt- 4v(ision)forautomaticimagedesignandgeneration,2024

    2024

  24. [24]

    BlackForestLabs,FLUX.1Kontext,https://bfl.ai/models/flux- kontext,Accessed:2026-05-21,May2025

    2026

  25. [25]

    BlackForestLabs,FLUX.2[max],https://bfl.ai/blog,Accessed: 2026-05-21,Dec.2025

    2026

  26. [26]

    BlackForestLabs,FLUX.2[pro],https://bfl.ai/blog,Accessed: 2026-05-21,Nov.2025

    2026

  27. [27]

    Black Forest Labs and Krea AI,FLUX.1 Krea, https://bfl.ai/ blog/flux-1-krea-dev,Accessed:2026-05-21,2025

    2026

  28. [28]

    Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer

    H.Caietal.,“Z-Image:Anefficientimagegenerationfounda- tion model with single-stream diffusion transformer”,arXiv preprintarXiv:2511.22699,2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  29. [29]

    C.-Y.Chen,M.Shi,G.Zhang,andH.Shi,T2i-copilot:Atraining- freemulti-agenttext-to-imagesystemforenhancedpromptinter- pretationandinteractivegeneration,2025

    2025

  30. [30]

    Ant:Adaptiveneuraltemporal-awaretext-to- motionmodel

    W.Chenetal.,“Ant:Adaptiveneuraltemporal-awaretext-to- motionmodel”,inProceedingsofthe33rdACMInternational ConferenceonMultimedia,2025,pp.9852–9861.doi:10.1145/ 3746027.3755168

    work page arXiv 2025

  31. [31]

    Chen et al.,Free-t2m:Robusttext-to-motiongenerationfor humanoidrobotsviafrequency-domain,2025

    W. Chen et al.,Free-t2m:Robusttext-to-motiongenerationfor humanoidrobotsviafrequency-domain,2025

    2025

  32. [32]

    Emerging Properties in Unified Multimodal Pretraining

    C. Deng et al., “Emerging properties in unified multimodal pretraining”,arXivpreprintarXiv:2505.14683,2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  33. [33]

    GoogleDeepMind,Gemini3proimage:Nanobananapro,https: //deepmind.google/models/gemini-image/pro/, Accessed: 2026-05-21,2025

    2026

  34. [34]

    Google DeepMind,Gemini image: Nano banana, https:// deepmind.google/models/gemini-image/, Accessed: 2026- 05-21,2025

    2026

  35. [35]

    arXiv preprint arXiv:2512.05112 (2025)

    D.Jiangetal.,“DraCo:DraftasCoTfortext-to-imagepreview andrareconceptgeneration”,arXivpreprintarXiv:2512.05112, 2025

    work page arXiv 2025

  36. [36]

    M.Ningetal.,Dctdiff:Intriguingpropertiesofimagegenerative modelinginthedctspace,2025

    2025

  37. [37]

    Y.Niuetal.,WISE:Aworldknowledge-informedsemanticeval- uationfortext-to-imagegeneration,2025

    2025

  38. [38]

    OpenAI,GPTImage1.5, https://platform.openai.com/docs/ models/gpt-image-1.5,Accessed:2026-05-21,Dec.2025

    2026

  39. [39]

    OpenAI,GPT-5.5, https://openai.com/index/gpt-5-5/, Ac- cessed:2026-06-02,2025

    2026

  40. [40]

    OpenAI,Introducingourlatestimagegenerationmodelintheapi, https://openai.com/index/image-generation-api/,Accessed: 2026-05-21,Apr.2025

    2026

  41. [41]

    Serper,SerperAPI,https://serper.dev/,Accessed:2026-06-02, 2025

    2026

  42. [42]

    B.Tian,W.Chen,Z.Li,S.Lai,J.Wu,andY.Yue,Text2weight: Bridging natural language and neural network weight spaces, https://doi.org/10.1145/3746027.3755441,2025

    work page doi:10.1145/3746027.3755441 2025

  43. [43]

    X.Wanetal.,Maestro:Self-improvingtext-to-imagegeneration viaagentorchestration,2025

    2025

  44. [44]

    Qwen-Image Technical Report

    C. Wu et al., “Qwen-image technical report”,arXiv preprint arXiv:2508.02324,2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  45. [45]

    Echo-4o: Harnessing the power of GPT-4o syn- theticimagesforimprovedimagegeneration

    J. Ye et al., “Echo-4o: Harnessing the power of GPT-4o syn- theticimagesforimprovedimagegeneration”,arXivpreprint arXiv:2508.09987,2025

    work page arXiv 2025

  46. [46]

    10–11 MemoGen: Can Past Experience Improve Future Text-to-Image Generation?

    S.Chenetal.,Genevolve:Self-evolvingimagegenerationagents viatool-orchestratedvisualexperiencedistillation,2026. 10–11 MemoGen: Can Past Experience Improve Future Text-to-Image Generation?

    2026

  47. [47]

    Towards betterevaluationmetricsfortext-to-motiongeneration

    W.Chen,H.Jia,K.Yu,S.Lai,L.Wang,andY.Yue,“Towards betterevaluationmetricsfortext-to-motiongeneration”,inThe SecondInternationalWorkshoponTransformativeInsightsin MultifacetedEvaluationatTheWebConference2026,2026

    2026

  48. [48]

    Gen-Searcher: Reinforcing Agentic Search for Image Generation

    K. Feng et al., “Gen-searcher: Reinforcing agentic search for imagegeneration”,arXivpreprintarXiv:2603.28767,2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  49. [49]

    Mind-Brush: Integrating agentic cognitive search and reasoning into image generation

    J. He et al., “Mind-Brush: Integrating agentic cognitive search and reasoning into image generation”,arXiv preprint arXiv:2602.01756,2026

    work page arXiv 2026

  50. [50]

    He et al.,Gems: Agent-native multimodal generation with memoryandskills,2026

    Z. He et al.,Gems: Agent-native multimodal generation with memoryandskills,2026

    2026

  51. [51]

    Genagent: Scaling text-to-image gener- ation via agentic multimodal reasoning

    K. Jiang et al., “Genagent: Scaling text-to-image gener- ation via agentic multimodal reasoning”,arXiv preprint arXiv:2601.18543,2026

    work page arXiv 2026

  52. [52]

    Kou et al.,Think-then-generate: Reasoning-aware text-to- imagediffusionwithllmencoders,2026

    S. Kou et al.,Think-then-generate: Reasoning-aware text-to- imagediffusionwithllmencoders,2026

    2026

  53. [53]

    H. Li, W. Chen, L. Wang, S. Liang, H. Jia, and Y. Yue,Oracle noise:Fastersemanticsphericalalignmentforinterpretablelatent optimization,2026

    2026

  54. [54]

    H.Lietal.,Deltascorematters!spatialadaptivemultiguidance indiffusionmodels,2026

    2026

  55. [55]

    S. Mun, S. Cho, and C. Kim,Epic: Efficient predicate-guided inference-timecontrolforcompositionaltext-to-imagegeneration, 2026

    2026

  56. [56]

    H.Wang,C.Wei,W.Ren,J.Liu,F.Lin,andW.Chen,Rational- rewards:Reasoningrewardsscalevisualgenerationbothtraining andtesttime,2026

    2026

  57. [57]

    S.Wu,M.Sun,W.Wang,Y.Wang,andJ.Liu,Visualprompter: Semantic-awarepromptoptimizationwithvisualfeedbackfor text-to-imagesynthesis,2026

    2026

  58. [58]

    Zhang et al.,Rewardharness: Self-evolving agentic post- training,2026

    Y. Zhang et al.,Rewardharness: Self-evolving agentic post- training,2026. 11–11

    2026