Pith. sign in

REVIEW 4 major objections 6 minor 16 references

Separating controlled divergence from evidence-governed absorption can reduce persona-environment self-locking while preserving identity continuity.

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 · grok-4.5

2026-07-10 10:34 UTC pith:OXYEXMIW

load-bearing objection Strong multi-model evidence for persona mode-lock, plus a real same-runtime diversity gain; the soft spot is that the A/B measures theme variety more than OSO absorption or identity continuity. the 4 major comments →

arxiv 2607.08252 v1 pith:OXYEXMIW submitted 2026-07-09 cs.AI cs.CLcs.HC

AutoPersonas: A Multi-Timescale Loop Engine for Open-Ended Persona Evolution

Mengchen Li This is my paper
classification cs.AI cs.CLcs.HC
keywords persona agentsself-lockinglife-environmentOSO loopopen-ended evolutioncontext gravitydiversity collapsemulti-timescale simulation
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.

Long-term persona agents are meant to keep living—entering new situations, forming relationships, revising concerns—while remaining recognizable. This paper identifies self-locking: a runtime failure in which locally plausible events keep appearing while the generated life collapses into familiar environments, weak relationships, suspended decisions, and stale life stages. The authors trace the failure to two coupled pressures: model convergence onto high-probability behavioral channels, and system-level context gravity from State, memory, history, and environment summaries. AutoPersonas is a multi-timescale life-environment engine whose OSO loop separates environment-side Occurrences, accumulated Observations, and persona State, so divergent future-facing material can enter while only evidence-governed absorption may revise State or reachability. Across eight models, direct recursive loops showed extreme action- and theme-channel repetition; a same-runtime A/B found that context-slice masking plus per-sample divergence targeting cut macro-theme repetition from 61.8% to 36.3% and roughly doubled cumulative themes. Readers who build continuing companions should care because memory alone does not solve this, and uncontrolled novelty breaks identity.

Core claim

Self-locking is a distinct runtime failure of continuing persona-life loops, not merely forgetting or low local text quality: the loop can produce varied prose while functionally returning to the same places, unresolved routes, relationship roles, and life stage. Direct self-orchestrated loops across eight models generated 1,600 events with mean rolling 5-day action-category repetition of 95.2%–97.6% (all models past 90% by day 11) and 79.0%–88.0% macro-theme repetition. The paper’s bounded systems claim is that separating controlled divergence from evidence-governed absorption—via an OSO architecture plus context-slice masking and per-sample divergence targeting—can reverse that fixation: i

What carries the argument

The OSO loop (Occurrence → Observation → State revision → future possibility space): a multi-timescale life-environment engine that keeps environment-side future-facing material, accumulated evidence, and continuity-bearing State as separate authorities. Controlled divergence (conditional variation with context-slice masking and per-sample targeting) opens plausible non-identical signals; evidence-governed absorption decides whether they harden into State or reachability. Without the divergence source the loop self-locks around old State; without OSO governance, novelty becomes drift.

Load-bearing premise

That action-category and macro-theme repetition on one primary complex persona canon, with one run per redesigned condition, are adequate proxies for whether a persona’s functional life trajectory is actually diversifying rather than only looking varied in labels.

What would settle it

A multi-persona, multi-run full-runtime A/B in which enabling context-slice masking and per-sample divergence targeting fails to cut macro-theme repetition or raise cumulative themes relative to the same-runtime baseline, or raises theme counts without corresponding diagnostic movement in State, relationship function, occurrence hardening, or later reachability.

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

If this is right

  • Direct recursive persona loops without life-environment architecture will rapidly close over a small action and life-theme repertoire across current foundation models.
  • Memory expansion and higher sampling temperature alone will not break self-locking; divergence must be architecturally separated from State absorption.
  • Day-level simulation is a phase boundary for audits: coarser weekly or yearly summaries can hide deferred decisions, decorative opportunities, and environment watermark shells.
  • Deployed companions with independent lives require dual-stream recall—persona self-life versus user-specific relationship memory—rather than user-only personalization.
  • Sandbox society simulators and open-environment persona engines solve different authority problems and should not be treated as direct benchmark substitutes.

Where Pith is reading between the lines

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

  • If self-locking is mainly a context-authority problem, long-horizon agent progress may come more from runtime governance than from larger base models alone.
  • The same separation of controlled divergence from evidence-governed absorption may transfer to other recursive generation settings where summaries re-enter generation, such as long-running story engines.
  • Because relationship persistence remained a hard failure even when theme variety improved, product systems that only measure event novelty may still be shipping decorative lives that never change obligations or reachability.
  • A natural next measurement is multi-persona variance bands that track whether route-authority and relationship-function changes move with theme counts, not only whether new themes appear.

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 / 6 minor

Summary. The paper identifies self-locking as a runtime failure in long-term persona-life loops: locally plausible events continue while trajectories collapse into familiar environments, weak relationships, deferred decisions, and stale stages. It attributes this to model-side channel convergence and system-side context gravity, and proposes AutoPersonas—an OSO (Occurrence–Observation–State) multi-timescale life-environment engine that separates controlled divergence from evidence-governed absorption. Evaluation is diagnostic rather than benchmark-competitive: a three-year compressed simulation surfaces failure modes; an eight-model 40-day direct-loop stress test (1,600 events) reports 95.2%–97.6% mean rolling 5-day action-category repetition and 79.0%–88.0% macro-theme repetition; a same-runtime A/B with context-slice masking plus per-sample divergence targeting reduces macro-theme repetition from 61.8% to 36.3% and roughly doubles theme count; a juvenile-goblin fictional-world run reproduces the anti-fixation regime. The bounded claim is that separating controlled divergence from evidence-governed absorption can reduce persona-environment self-locking while preserving identity continuity.

Significance. If the result holds, the paper supplies a useful systems framing for open-ended persona agents distinct from task agents and closed sandbox societies, plus a concrete, reproducible demonstration that direct recursive persona loops rapidly mode-lock across current foundation models. Strengths include the eight-model stress test, temperature probe, semantic re-keeping, same-runtime A/B (better than cross-setting anecdote), public-safe aggregate artifacts, and an action-repetition evaluator. The life-environment / OSO separation is a clear architectural contribution for reward-free open evolution under context gravity. The main significance risk is that the strongest quantitative mitigation evidence measures generation-side theme diversity rather than the paper’s own causal criteria for anti-self-locking (hardening into State and reachability change with identity continuity).

major comments (4)
  1. [§7.2 Quantitative anti-fixation validation] §7.2 and Abstract: the load-bearing claim is that separating controlled divergence from evidence-governed absorption reduces self-locking while preserving identity continuity. The same-runtime A/B and isolated-generator ablation report cumulative macro-theme count and repeat ratio under a reviewer reuse rule on generated events (61.8%→36.3%, 55→102 themes). They do not score the paper’s own OSO criteria in §§3.6 and 4.5—Occurrence→Observation→State revision→changed future possibility space—nor identity-continuity checks under redesigned conditions. Masking + targeting can raise theme variety without proving novelty hardens rather than remaining decorative (the occurrence-hardening gap of §7.3). Please report, for the A/B arms, at least proxy rates of State/reachability movement, hardening success/failure, and a continuity audit (e.g., canon-consistency or stale-state regression), or narr
  2. [§7.2 / §14 Limitations] §7.2 and §14: the redesigned A/B and isolated ablation use one primary complex persona canon with one run per condition; the paper correctly flags missing variance bands and multi-persona replication. A single-run effect size this large is suggestive but not yet adequate for the systems claim as stated. At minimum, report 2–3 independent seeds or a second real-world canon under the same A/B protocol, with uncertainty on repeat ratio and theme count, or restate results as a single-trajectory mechanism demonstration.
  3. [§6.1 Experimental setup; §7.1] §6.1 and §7.1: action-channel metrics rely on a rule-based keyword taxonomy over title/action fields (with Other residual), and theme metrics on a semantic keeper under a reuse rule. Stricter variants still show early convergence, which is reassuring for the baseline, but the mitigation A/B is only reported in the theme vocabulary. Please (i) apply the same action-channel metric to the redesigned full-runtime A/B arms, and (ii) state inter-rater or stability checks for the macro-theme grouping, so the anti-fixation gain is not keeper-specific.
  4. [§4 AutoPersonas; §7.2 Stage 2] §4 and §7.2: AutoPersonas is defined as a multi-mechanism engine (conditional variation, context governance, information orthogonality, progressive causal propagation, trajectory monitoring, multi-timescale revision). The quantitative intervention tests only context-slice masking plus per-sample divergence targeting (and a narrative-arc targeting-only lane). That is a valid mechanism A/B for divergence under context gravity, but it does not validate the full OSO absorption stack. Either add at least one absorption-side ablation (e.g., hardening/review disabled vs enabled with divergence fixed) or explicitly bound the claim to the divergence/governance pair rather than the full engine.
minor comments (6)
  1. [Abstract] Abstract vs §7.1: abstract says all models crossed 90% by day 11; body also states all crossed 80% by day 9. Keep thresholds consistent in the abstract.
  2. [§1 Introduction] §1 mentions a ~60% history/state token-budget threshold as a design finding from researcher judgment. Flag more clearly as non-benchmarked and avoid implying a universal law (the paper already hedges; a single sentence in Limitations would help).
  3. [§7.4–7.6] Figures 3–6 are public-safe storyboards; ensure captions state that panels are illustrative compressions, not raw logs, so readers do not over-read them as quantitative evidence.
  4. [§8 Comparison] Related work on Generative Agents / Agentopia is appropriately complementary; a short explicit non-claim that AutoPersonas is not evaluated on society-level believability metrics would reduce misreading as a sandbox competitor.
  5. [§3.4] Terminology density (self-locking, watermark shell, occurrence hardening, current-state authority) is high; a one-page glossary early (partially present in §3.4) would help non-systems readers.
  6. [§7.1 Table] Kimi temperature forced to 1.0 and mixed OpenRouter/Moonshot route is disclosed; consider a footnote in the main results table so readers do not treat that row as fully protocol-matched.

Circularity Check

0 steps flagged

No load-bearing circular derivation: the mode-lock baselines and A/B anti-fixation gains are empirical outcome statistics, not quantities forced by definition or self-citation.

full rationale

This is an empirical systems paper, not a first-principles derivation. Self-locking is defined as runtime functional collapse (stale State/environment/relationship attractors), then operationalized with external behavioral metrics—rolling action-category repetition under a fixed keyword taxonomy and cumulative macro-theme reuse under a stated reuse rule—and those metrics are applied to direct-loop baselines (eight models, 1,600 events) and to same-runtime A/B conditions with divergence mechanisms off vs on. The A/B numbers (61.8%→36.3% macro-theme repetition; 55→102 themes) are not algebraically implied by the OSO definitions, the masking rule, or the targeting interface; masking alone left repetition high (85%), so the measured gain is not tautological. There is no fitted parameter renamed as a prediction, no uniqueness theorem imported from the same authors, and no load-bearing self-citation chain that forces the result. Related-work citations (Generative Agents, Agentopia, model-collapse literature) are used for positioning, not as the sole warrant for the quantitative claims. Residual concern is only mild operationalization risk: absolute fixation levels partly reflect author-chosen taxonomies/theme grouping, which can inflate how strongly the problem is scored—but comparative condition differences under a fixed metric are still independent evidence. Metric–claim mismatch (theme diversity vs full Occurrence→State→reachability hardening and identity continuity) is a validity gap, not circularity. Score 1 reflects that minor taxonomy-dependence only.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 5 invented entities

This is a systems architecture paper, not a first-principles derivation. The load-bearing content is a failure diagnosis plus an engineered loop whose success is measured by author-defined diversity proxies. Free parameters are operational (temperatures, windows, taxonomies, token-budget heuristic). Axioms are domain assumptions about LLM context gravity and open-environment persona requirements. Invented entities are architectural abstractions introduced to make the failure and mitigation auditable; independent external evidence for them is limited to the paper’s own runs.

free parameters (5)
  • divergence token-budget threshold (~60% history/state)
    Design finding that divergence degrades once historical/current-state material exceeds about 60% of the future-facing prompt budget; treated as empirical heuristic, not a derived law.
  • daily life-generation temperature (0.75 baseline; Kimi forced 1.0; Doubao probe 1.0)
    Decoding temperatures chosen for stress tests and the temperature probe; affect measured channel weights and are not theoretically fixed.
  • rolling 5-day / 25-event repetition window
    Primary quantitative metric definition; window size is a measurement choice that shapes reported convergence timing.
  • rule-based action taxonomy and macro-theme reuse rule
    Category and theme grouping determine repetition ratios; unmatched events go to Other; semantic re-keeping depends on reviewer grouping decisions.
  • events-per-day and horizon (5 events/day, 40 days; 3-year compressed diagnostic)
    Simulation cadence and compression choices control how quickly recursive collapse appears and what failures are visible.
axioms (5)
  • domain assumption Long-term persona evolution is reward-free open-evolution under context gravity, not task optimization against an external scalar success criterion.
    Stated in Introduction and Problem Formulation; defines the evaluation target and why sandbox reward systems are treated as a different problem.
  • domain assumption Current State, memory, history, and environment summaries exert recursive authority that can dominate future generation (context gravity).
    Core causal premise for self-locking; motivates masking and authority separation throughout Sections 3–4.
  • ad hoc to paper Separating Occurrence, Observation, and State authorities is necessary for open-ended growth without identity drift.
    Architectural axiom of the OSO loop; not independently proven, but used as the design principle being tested.
  • ad hoc to paper Day-level simulation is a phase boundary that reveals self-locking hidden by weekly/yearly summaries.
    Contribution 6 and Section 3.7; justifies the 40-day high-frequency stress protocol.
  • domain assumption Base-model social-world priors can supply plausible persona-conditioned long-tail Occurrences under constraints.
    Section 4.4; underwrites conditional occurrence generation without claiming true human-life distribution recovery.
invented entities (5)
  • self-locking no independent evidence
    purpose: Name runtime functional collapse of persona-life-environment trajectories into stale attractors despite locally varied text.
    Central diagnostic construct; operationalized via action/theme repetition and qualitative failure taxonomy, not a pre-existing standard metric.
  • OSO loop (Occurrence–Observation–State) no independent evidence
    purpose: Temporal-authority partition that admits divergent future material only through evidence-governed absorption.
    Core engine abstraction; evidence is the paper’s own A/B and diagnostics.
  • life-environment layer no independent evidence
    purpose: Persona-specific middle layer between macro-world and State that defines reachable places, people, risks, and opportunities.
    Architectural object distinguishing AutoPersonas from shared sandbox world models.
  • semantic State machine no independent evidence
    purpose: Schema-bounded but open State representation updated by meaning-level evidence accumulation rather than enumerated transitions.
    Introduced to handle sparse gradual and jump changes; no external formal semantics beyond the paper.
  • environment watermark shell / occurrence hardening no independent evidence
    purpose: Failure labels for recycled living shells and decorative opportunities that never revise State or reachability.
    Audit taxonomy entities used in the three-year diagnostic; useful but paper-internal.

pith-pipeline@v1.1.0-grok45 · 33038 in / 3983 out tokens · 40318 ms · 2026-07-10T10:34:20.632010+00:00 · methodology

0 comments
read the original abstract

Long-term persona agents must remain identifiable while adapting to new events, relationships, evidence, and social conditions. We identify self-locking as a runtime failure mode in continuing persona-life loops: locally plausible events keep appearing while the generated life collapses toward familiar environments, weak relationships, suspended decisions, and stale life stages. We trace this failure to model-level convergence toward high-probability behavioral channels and system-level context gravity from State, memory, history, and environment summaries. We introduce AutoPersonas, a multi-timescale life-environment engine for bounded persona-level recursive self-evolution. It separates environment-side Occurrences, accumulated Observations, and persona State. Its OSO loop admits divergent future-facing material while requiring evidence-governed absorption before State or reachability changes. A three-year compressed simulation exposed environment watermark shells, occurrence-hardening gaps, slow-change accumulation failures, recursive indecision, and weak relationship persistence. An eight-model 40-day stress test generated 1,600 events and found mean rolling 5-day action-category repetition of 95.2%-97.6%, with all models crossing 90% by day 11. Semantic re-keeping found 79.0%-88.0% macro-theme repetition across all direct-loop runs. In a same-runtime 40-day A/B, context-slice masking plus per-sample divergence targeting reduced macro-theme repetition from 61.8% to 36.3% and roughly doubled cumulative theme count. A juvenile-goblin fictional-world run reproduced the anti-fixation regime without hard real-world intrusions. These results support a bounded claim: separating controlled divergence from evidence-governed absorption can reduce persona-environment self-locking while preserving identity continuity.

Figures

Figures reproduced from arXiv: 2607.08252 by Mengchen Li.

Figure 1
Figure 1. Figure 1: Dual-stream recall after AutoPersonas. Each turn can trigger a persona-level self-memory [PITH_FULL_IMAGE:figures/full_fig_p019_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: CIBE cognitive pipeline. Context assembles available self, user, relationship, episodic, and [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Student route-authority storyboard. Public-safe compressed case traces are rendered as [PITH_FULL_IMAGE:figures/full_fig_p030_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Student creative-route storyboard. A weak media-club contact hardens into field recording, [PITH_FULL_IMAGE:figures/full_fig_p031_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Masked-lane non-arc storyboard. These panels are arranged as a mosaic rather than a [PITH_FULL_IMAGE:figures/full_fig_p032_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Fictional-world storyboard. The juvenile-goblin stress case uses non-modern, world [PITH_FULL_IMAGE:figures/full_fig_p033_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Causal loop of AutoPersonas. Observations, State, and Occurrences are kept as separable [PITH_FULL_IMAGE:figures/full_fig_p038_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Semantic State machine. State is schema-bounded enough for audit, but the content and [PITH_FULL_IMAGE:figures/full_fig_p039_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Multi-timescale AutoPersonas revision. Fast event capture and slower evidence review [PITH_FULL_IMAGE:figures/full_fig_p040_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: SoulOS 3M overview. AutoPersonas handles life-environment evolution, MemorOS handles [PITH_FULL_IMAGE:figures/full_fig_p041_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Life-environment layer. Macro-world signals change a persona-specific future possibility [PITH_FULL_IMAGE:figures/full_fig_p042_11.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

16 extracted references · 16 canonical work pages · 5 internal anchors

  1. [1]

    Self-Consuming Generative Models Go MAD

    Sina Alemohammad, Josue Casco-Rodriguez, Lorenzo Luzi, Ahmed Imtiaz Humayun, Hossein Babaei, Daniel LeJeune, Ali Siahkoohi, and Richard Baraniuk. Self-Consuming Generative Models Go MAD. InInternational Conference on Learning Representations, 2024. URL https://proceedings.iclr.cc/paper_files/paper/2024/hash/ebc042e767de551803cc fcc45e2454f5-Abstract-Confe...

    work page 2024

  2. [2]

    Project Sid: Many-agent simulations toward AI civilization

    Altera.AL, Andrew Ahn, Nic Becker, Stephanie Carroll, Nico Christie, Manuel Cortes, Arda Demirci, Melissa Du, Frankie Li, Shuying Luo, Peter Y. Wang, Mathew Willows, Feitong Yang, and Guangyu Robert Yang. Project Sid: Many-agent simulations toward AI civilization, 2024. URLhttps://arxiv.org/abs/2411.00114. 49

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  3. [3]

    Snap out of it: A dual-process approach to mitigating overthinking in language model reasoning

    Oron Anschel, Alon Shoshan, Adam Botach, Shunit Haviv Hakimi, Asaf Gendler, Emanuel Ben Baruch, Nadav Bhonker, Igor Kviatkovsky, Manoj Aggarwal, and Gerard Medioni. Group- Aware Reinforcement Learning for Output Diversity in Large Language Models. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 32394– 32415,...

    work page doi:10.18653/v1/2025 2025

  4. [4]

    A Theoretical Perspective: How to Prevent Model Collapse in Self-consuming Training Loops

    Shi Fu, Yingjie Wang, Yuzhu Chen, Xinmei Tian, and Dacheng Tao. A Theoretical Perspective: How to Prevent Model Collapse in Self-consuming Training Loops. InInternational Conference on Learning Representations, 2025. doi: 10.48550/arXiv.2502.18865. URLhttps://arxiv.or g/abs/2502.18865

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2502.18865 2025

  5. [5]

    I. J. Good. Speculations Concerning the First Ultraintelligent Machine. InAdvances in Computers, volume 6, pages 31–88. Academic Press, 1966. doi: 10.1016/S0065-2458(08)60418-0. URLhttps://www.sciencedirect.com/science/article/pii/S0065245808604180

    work page doi:10.1016/s0065-2458(08)60418-0 1966

  6. [6]

    Collapse or Thrive? Perils and Promises of Synthetic Data in a Self-Generating World

    Joshua Kazdan, Rylan Schaeffer, Apratim Dey, Matthias Gerstgrasser, Rafael Rafailov, David L. Donoho, and Sanmi Koyejo. Collapse or Thrive? Perils and Promises of Synthetic Data in a Self-Generating World, 2024. URL https://arxiv.org/abs/2410.16713 . Accepted at NeurIPS 2024 workshops M3L and ATTRIB

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  7. [7]

    One fish, two fish, but not the whole sea: Alignment reduces language models’ conceptual diversity

    Sonia Krishna Murthy, Tomer Ullman, and Jennifer Hu. One fish, two fish, but not the whole sea: Alignment reduces language models’ conceptual diversity. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 11241–11258, Albuqu...

    work page 2025

  8. [8]

    doi: 10.18653/v1/2025.naacl-long.561

    Association for Computational Linguistics. doi: 10.18653/v1/2025.naacl-long.561. URL https://aclanthology.org/2025.naacl-long.561/

    work page doi:10.18653/v1/2025.naacl-long.561 2025

  9. [9]

    Omohundro

    Stephen M. Omohundro. The Basic AI Drives. InArtificial General Intelligence 2008: Proceedings of the First AGI Conference, volume 171 ofFrontiers in Artificial Intelligence and Applications, pages 483–492. IOS Press, 2008. URLhttps://selfawaresystems.com/2007/ 11/30/paper-on-the-basic-ai-drives/

    work page 2008

  10. [10]

    Bernstein

    Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative Agents: Interactive Simulacra of Human Behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, UIST ’23, San Francisco, CA, USA, 2023. Association for Computing Machinery. doi: 10.1145/ 35...

    work page doi:10.1145/3586183.3606763 2023

  11. [11]

    Position: Model Collapse Does Not Mean What You Think

    Rylan Schaeffer, Joshua Kazdan, Alvan Caleb Arulandu, and Sanmi Koyejo. Position: Model Collapse Does Not Mean What You Think, 2025. URLhttps://arxiv.org/abs/2503.03150

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  12. [12]

    AI models collapse when trained on recursively generated data.Nature, 631:755–759, 2024

    Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson, and Yarin Gal. AI models collapse when trained on recursively generated data.Nature, 631:755–759, 2024. doi: 10.1038/s41586-024-07566-y. URL https://www.nature.com/articles/s41586-024-0 7566-y

    work page doi:10.1038/s41586-024-07566-y 2024

  13. [13]

    Agentopia: Long-Term Life Simulation and Learning in Agent Societies

    Xintao Wang, Sirui Zheng, Hongqiu Wu, Weiyuan Li, Jen-tse Huang, Minghao Zhu, Can Zu, Qi Deng, Jiawei Wang, Qianyu He, Heng Wang, Xiaojian Wu, and Yunzhe Tao. Agentopia: Long-Term Life Simulation and Learning in Agent Societies, 2026. URLhttps://arxiv.org/ abs/2606.07513. arXiv preprint. 50

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  14. [14]

    Humanoid Agents: Platform for Simulating Human-like Generative Agents

    Zhilin Wang, Yu Ying Chiu, and Yu Cheung Chiu. Humanoid Agents: Platform for Simulating Human-like Generative Agents. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pages 167–176, Singapore, 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.emnlp-demo.15. URL https://ac...

    work page doi:10.18653/v1/2023.emnlp-demo.15 2023

  15. [15]

    The Price of Format: Diversity Collapse in LLMs

    Longfei Yun, Chenyang An, Zilong Wang, Letian Peng, and Jingbo Shang. The Price of Format: Diversity Collapse in LLMs. InFindings of the Association for Computa- tional Linguistics: EMNLP 2025, pages 15454–15468, Suzhou, China, 2025. Association for Computational Linguistics. doi: 10.18653/v1/2025.findings-emnlp.836. URL https://aclanthology.org/2025.find...

    work page doi:10.18653/v1/2025.findings-emnlp.836 2025

  16. [16]

    Hyperagents, 2026

    Jenny Zhang, Bingchen Zhao, Wannan Yang, Jakob Foerster, Jeff Clune, Minqi Jiang, Sam Devlin, and Tatiana Shavrina. Hyperagents, 2026. URLhttps://arxiv.org/abs/2603.19461. arXiv preprint. 51

    work page arXiv 2026