Pith. sign in

REVIEW 4 major objections 5 minor 31 references

Language Games as the Pathway to Artificial Superhuman Intelligence

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper claims that language games—structured linguistic interactions with shifting roles, rewards, and rules—keep LLM training data sufficiently novel to avoid stagnation and thus open the route to artificial superhuman intelligence.

desk verdict A clear, well-written position paper that names a real problem and proposes a plausible agenda, but the central causal claim is asserted, not demonstrated, and the formal definitions do not carry the weight put on them. read the letter →

arxiv 2501.18924 v1 pith:WJLB6RI5 submitted 2025-01-31 cs.AI cs.CLcs.MA

Ying Wen , Ziyu Wan , Shao Zhang This is my paper · ORCID
classification cs.AIcs.CLcs.MA
keywords languagegamesdatareproductiontrapexpandedartificialsuperhumanintelligencerolefluidityrewardvarietyruleplasticityopen-endedlearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that large language models stall when they are repeatedly trained on data drawn from the same human-generated distributions they have already mastered, a condition it calls the data reproduction trap. It proposes dynamic language games—structured, open-ended linguistic interactions among humans, models, and bots—as a mechanism for escaping this trap by continuously producing novel training data. Three design principles are said to do the work: role fluidity (agents switch between teacher, student, critic, and other roles), reward variety (multiple, sometimes conflicting feedback criteria), and rule plasticity (the game's constraints evolve over time). The paper claims these jointly keep the effective data distribution shifting, so each training round injects enough novelty to avoid stagnation, and that scaling such games to global human-AI interaction turns data reproduction into an engine for superhuman intelligence.

What carries the argument

The central object is the language game, defined as a tuple of agents, a shared semantic context, per-agent action spaces, a reward function, and a state-evolution rule. It carries the argument by supplying three continuously adjustable degrees of freedom—who plays (roles), what is rewarded, and what rules constrain play—that the paper identifies as the mechanisms keeping the curated data distribution moving. The formal target is the inequality defining expanded data reproduction, that the time-averaged distributional shift between successive curated datasets stays above a positive threshold.

What would settle it

Run a multi-agent language-game training loop for dozens of retraining rounds, measure the distributional shift $D(p_t, p_{t-1})$ between consecutive curated datasets, and also test the final model on held-out novel tasks; the central claim is falsified if the shift decays to near zero or the model's novel-task performance does not improve even while the shift metric stays high.

Watch

Extended reading notes

Core claim

The paper's central claim is that language games are the pivotal mechanism for escaping the data reproduction trap in large language models, enabling open-ended conceptual exploration on the way to artificial superhuman intelligence. Data reproduction is formalized as the cycle in which a model generates data, a curation step selects a subset, and retraining updates the model; a trap occurs when the distribution of curated data stops changing, measured by a small threshold on divergence between successive iterations. The paper defines expanded data reproduction as the opposite condition—average divergence above a positive threshold—and asserts that language games achieve it through three mutually reinforcing mechanisms: role fluidity, reward variety, and rule plasticity. It further claims that scaling these games to global human-AI ecosystems turns data reproduction into a self-sustaining engine of co-evolution rather than a closed loop.

Load-bearing premise

The claim stands or falls on whether role fluidity, reward variety, and rule plasticity are sufficient to keep the curated data distribution shifting indefinitely without the data collapsing or becoming unlearnable.

Editorial extensions

If this is right

  • If the claim holds, recursive self-training with role switching, plural rewards, and evolving rules should sustain capability growth where naive self-training on generated data collapses.
  • Deployment itself becomes a training signal: at global scale, each interaction of humans and models feeds the next generation, so progress toward superhuman intelligence becomes a sociotechnical process, not a benchmark race.
  • Reward hacking and distributional collapse are treated as symptoms of insufficient role, reward, or rule variation, implying that fixing those three levers should suppress the failure modes documented in model-collapse studies.
  • The framework predicts that models exposed to culturally and epistemically diverse language games will develop reasoning blends—such as combining dialectical and deductive inference—that exceed what any single tradition produces.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • We note that the paper's formal definitions make the 'trap' measurable: one could instrument a live system by logging $D(p_t, p_{t-1})$ across retraining rounds, giving an early-warning signal the paper does not specify.
  • An implication the authors leave implicit is that interface design becomes a core AI-research variable: making role and rule changes cheap for users would directly increase the novelty of the data stream.
  • A natural extension is to test the framework against a baseline by tracking the rate of acquisition of new skills rather than final benchmark scores; the paper predicts the language-game pipeline keeps acquiring skills longer.
  • The argument also suggests a governance consequence: if sustained novelty is the goal, then diversity of participants is not an equity add-on but a training necessity, which the paper gestures at but does not fully develop.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper argues that large language models risk a 'data reproduction trap' in which closed-loop optimization on fixed human-generated distributions leads to stagnation, and proposes 'language games' as a mechanism for 'expanded data reproduction' that would sustain continuous distributional shift and drive progress toward artificial superhuman intelligence (ASI). It formalizes data reproduction (Definition 2.1), the trap (Definition 2.2), expanded data reproduction (Definition 2.3), and language games as a tuple G=(A,S,{U_a},R,E) (Definition 3.1). The central claim is that three mechanisms—role fluidity, reward variety, and rule plasticity—ensure a steady influx of novel, learnable data, and that scaling these games globally creates a sociotechnical ecosystem supporting open-ended conceptual exploration. The paper is structured as a position/vision paper: it offers definitions, qualitative examples, a comparative table, a discussion of global-scale pathways, and an extensive limitations section, but no experiments, formal proofs, or quantitative validation.

Significance. If the central claim were substantiated, the paper would provide a useful data-centric framework for open-ended learning and a concrete proposal for how LLMs might escape recursive training collapse. It rightly connects to important existing work on model collapse, open-endedness, multi-agent learning, and human-AI co-evolution, and it is commendable for explicitly acknowledging many failure modes, including reward hacking, oligopoly risks, and epistemic concerns. The formal definitions of the data reproduction trap and expanded data reproduction are a valuable starting point for discussion. However, the paper's current contribution is largely conceptual and programmatic: the key link from 'sustained distributional shift' to 'capability growth toward ASI' is asserted rather than demonstrated, and the formal apparatus is not yet operational. The significance is therefore conditional on future work that makes the framework testable and provides evidence for its core sufficiency claim.

major comments (4)
  1. [Section 2.3, Definition 2.3] The criterion for expanded data reproduction, (1/T) Σ D(p_t, p_{t-1}) > Δ, measures only the magnitude of distributional shift and says nothing about its direction, quality, or learnability. A sequence of distributions driven by random rule changes, adversarial perturbations, or even progressive degradation can satisfy this inequality without producing any capability gain. Since the paper's central claim requires that language games produce the 'right kind' of shift—one that supports open-ended conceptual exploration—the formal definition is insufficient. The paper needs an additional condition linking D to a learning-progress measure (for example, decreasing loss on a target task, increasing held-out success, or a formal learnability criterion) and an argument that language games satisfy that condition.
  2. [Section 3.1] The paper asserts that role fluidity, reward variety, and rule plasticity 'ensure a steady influx of novelty' and 'maintain learnable gradients,' but no definition of learnability is provided and no theoretical or empirical argument rules out the possibility that these mechanisms generate high-D data that is unlearnable or misaligned with capability. The thresholds δ and Δ in Definitions 2.2 and 2.3 are also left unspecified, so the formal claims are not falsifiable as stated. A concrete proposal for measuring D, setting thresholds, and testing whether language games actually produce expanded data reproduction would be needed to make the core claim testable.
  3. [Section 4.3] The 'three promising pathways' to ASI—cross-cultural concept fusion, distributed proof markets, and consensus reality engineering—are each described in a few sentences with no supporting evidence or mechanistic explanation for why they would lead to superhuman capability. For a paper that repeatedly emphasizes formal definitions and systematic analysis, this section relies on assertion rather than argument. The authors should either provide preliminary evidence from existing systems, clearly mark these as open hypotheses, or explain more concretely how each pathway operationalizes expanded data reproduction into measurable capability growth.
  4. [Sections 5.6 and 7] There is a direct tension between the conclusion's claim that language games 'maintain learnable gradients' and Section 5.6's admission that 'an echo chamber of superficial novelty may emerge' without appropriate validation. The paper does not explain which design features prevent the echo-chamber failure mode or how the proposed governance mechanisms would be sufficient to rule it out. Since the central claim depends on language games producing beneficial rather than merely novel data, this failure mode must be addressed with a concrete mechanism or a precise boundary condition, not just listed as a limitation.
minor comments (5)
  1. [Section 3.3] There is a typo in 'crucial toexpanded data reproduction' and in 'The Socratic Gamedemonstrates' at the start of Section 3.2; both should be corrected.
  2. [Section 6] The phrase 'similar to that that shaped organic minds' contains a doubled 'that' and should be edited.
  3. [Table 1 and Appendix A] The bullet-count scoring in Table 1 uses uncalibrated ordinal levels; the metric definitions in Table 2 are qualitative and do not explain how any rater would assign counts, which makes the comparative claim difficult to evaluate.
  4. [Section 4.4] The term 'Societal Turing Test' is introduced without a definition; it would benefit from a precise description of the evaluation procedure.
  5. [General] The paper would be stronger if it explicitly stated at the outset that it is a position paper and that the formal definitions are intended as a conceptual scaffold rather than as a complete mathematical theory, to set appropriate expectations for the reader.

Circularity Check

1 steps flagged · score 6.0 of 10

The claimed link from language games to expanded data reproduction is built into Definitions 2.3 and 3.1, making the central mechanism a definitional restatement.

  1. self definitional [Definition 2.3; Definition 3.1; Section 3.1 (Rule Plasticity)]
    "Using the same measure of distributional shift D(pt, pt−1), we say the model exhibits expanded data reproduction if there exists a positive threshold ∆ such that: 1/T Σ D(pt, pt−1) > ∆ ... E defines how states evolve and how new roles, rules, or information enter the game. ... The environment of a language game remains malleable, allowing the introduction of new constraints, tasks, or cultural references over time. These 'rule injections' prevent any single strategy or distribution from dominating ..."

    Expanded data reproduction (Def. 2.3) is defined as average distributional shift above a threshold. A language game (Def. 3.1) is defined with E, which 'defines how new roles, rules, or information enter the game.' Section 3.1 then 'derives' that rule plasticity 'prevent[s] any single strategy or distribution from dominating' and that interactions 'generate off-distribution samples,' i.e., the very D > Δ condition. The novelty-producing mechanism is an input to Definition 3.1, not a demonstrated consequence, so the link from language games to expanded data reproduction is true by construction.

full rationale

This is a position/framework paper rather than an empirical study, so there are no fitted parameters, benchmark predictions, or external validations to check. The self-citations present (Yang et al. 2021; Wen et al. 2024) are used as supporting references for multi-agent curriculum and RL integration, but they are not load-bearing for the central argument, so they do not themselves create circularity. The main circularity is definitional: the paper defines expanded data reproduction in Definition 2.3 purely as a distributional-shift threshold, and it defines language games in Definition 3.1 with an element E whose role is to introduce new roles, rules, or information. Section 3.1 then presents the resulting off-distribution samples and sustained distributional shift as a discovered property of language games, when it is a restatement of the stipulated definition. This makes the paper's central claim that language games achieve expanded data reproduction true by construction. However, the paper's stronger ASI claim does not reduce to the definitions: nothing in Definitions 2.3 or 3.1 entails that the distributional shift is learnable or capability-relevant, and Section 5.6 explicitly concedes that superficial novelty can emerge. That remaining gap is an unsupported sufficiency assumption and a correctness risk, not an additional circular step, so the overall score is 6 rather than 8-10.

Assumptions & free parameters 1 free parameters · 5 assumptions · 2 invented entities

The paper's central claim rests on several unverified domain assumptions about data-driven improvement, distributional shift as a proxy for intelligence, and the feasibility of global human-AI co-evolution. The only explicit free parameters are the divergence thresholds δ and Δ, which are left undefined. The formal language game object and the Societal Turing Test are introduced without independent evidence.

free parameters (1)
  • Divergence thresholds δ and Δ = unspecified
    Definitions 2.2 and 2.3 depend on thresholds δ and Δ but provide no values, estimation procedure, or sensitivity analysis, making the trap/expansion conditions untestable as stated.
assumptions (5)
  • domain assumption LLM capabilities grow primarily by ingesting and training on newly generated data.
    Section 2.1 defines data reproduction as the engine of model evolution; no proof is offered that this is sufficient or necessary for ASI.
  • domain assumption Distributional shift D(p_t, p_{t-1}) is a valid proxy for capability growth.
    Section 2.3 equates average divergence exceeding Δ with expanded reproduction and assumes this drives open-ended learning, but no evidence connects this divergence to intelligence.
  • domain assumption Language game mechanisms produce learnable novelty rather than degenerate or repetitive data.
    Section 3.1 asserts role fluidity, reward variety, and rule plasticity 'ensure a steady influx of novelty' without experiments; the paper's own Section 5 notes expressive limits and manipulation risks.
  • domain assumption Global human-AI co-evolution generates unbounded, useful data streams.
    Section 4 assumes scaling to billions of users yields open-ended data and continued novelty; no measurement or argument supports unboundedness.
  • domain assumption The 'reward is enough' hypothesis holds at the metacognitive level.
    Section 1 invokes Silver et al. (2021) as a foundation; this is a contested philosophical premise, assumed without critical scrutiny.
invented entities (2)
  • Formal language game tuple G=(A,S,{U_a},R,E)
    purpose: Formal model of training interactions used to define the proposed mechanism for expanded data reproduction.
    Introduced in Definition 3.1; no experimental or observational validation of the formal object as a predictor of intelligence.
  • Societal Turing Test
    purpose: Evaluative concept for continuous global assessment of AI by human communities.
    Mentioned in Section 4.4 as a governance device; no operational definition or measurement procedure is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Language Games as the Pathway to Artificial Superhuman Intelligence." pith.science (2026). https://pith.science/paper/WJLB6RI5

@misc{pith2026250118924,
  author       = {Pith},
  title        = {Pith review of: Language Games as the Pathway to Artificial Superhuman Intelligence},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WJLB6RI5}},
  note         = {Machine review of arXiv:2501.18924}
}
read the original abstract

The evolution of large language models (LLMs) toward artificial superhuman intelligence (ASI) hinges on data reproduction, a cyclical process in which models generate, curate and retrain on novel data to refine capabilities. Current methods, however, risk getting stuck in a data reproduction trap: optimizing outputs within fixed human-generated distributions in a closed loop leads to stagnation, as models merely recombine existing knowledge rather than explore new frontiers. In this paper, we propose language games as a pathway to expanded data reproduction, breaking this cycle through three mechanisms: (1) \textit{role fluidity}, which enhances data diversity and coverage by enabling multi-agent systems to dynamically shift roles across tasks; (2) \textit{reward variety}, embedding multiple feedback criteria that can drive complex intelligent behaviors; and (3) \textit{rule plasticity}, iteratively evolving interaction constraints to foster learnability, thereby injecting continual novelty. By scaling language games into global sociotechnical ecosystems, human-AI co-evolution generates unbounded data streams that drive open-ended exploration. This framework redefines data reproduction not as a closed loop but as an engine for superhuman intelligence.

Figures

Figures reproduced from arXiv: 2501.18924 by the authors.

Figure 1
Figure 1. An iterative data reproduction framework powered by language games progressively refines model capabilities toward artificial superhuman intelligence. From left to right: (1) Data reproduction via closed-loop optimization, (2) Expanded data reproduction via language games with evolving roles, rules and rich rewards, (3) Global language games driving continual adaptation and surpassing human-level capabilities. open-… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 2 canonical work pages

  1. [1]

    L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al

    Achiam, J., Adler, S., Agarwal, S., Ahmad, L., Akkaya, I., Aleman, F. L., Almeida, D., Altenschmidt, J., Altman, S., Anadkat, S., et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774,

    arXiv

  2. [8]

    Self-playing adversarial language game enhances llm reasoning

    Cheng, P., Hu, T., Xu, H., Zhang, Z., Dai, Y ., Han, L., and Du, N. Self-playing adversarial language game enhances llm reasoning. arXiv preprint arXiv:2404.10642,

    arXiv

  3. [10]

    URL https://doi.org/ 10.1145/1553374.1553380

    1145/1553374.1553380. URL https://doi.org/ 10.1145/1553374.1553380. Bergemann, D. and Bonatti, A. Markets for information: An introduction. Annual Review of Economics , 11(1): 85–107,

    arXiv

  4. [11]

    Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L

    URL https: //arxiv.org/abs/2012.08630. Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., and Fei-Fei, L. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248–255. Ieee,

    arXiv 2012

  5. [18]

    LeCun, Y

    URL https://arxiv.org/ abs/2402.06782. LeCun, Y . A path towards autonomous machine intelligence. OpenReview,

    arXiv

  6. [21]

    Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al

    URL https://arxiv.org/abs/1903.00742. Liu, A., Feng, B., Xue, B., Wang, B., Wu, B., Lu, C., Zhao, C., Deng, C., Zhang, C., Ruan, C., et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437,

    arXiv 1903

  7. [23]

    and Mumuni, F

    Mumuni, A. and Mumuni, F. Large language models for artificial general intelligence (agi): A survey of foundational principles and approaches. arXiv preprint arXiv:2501.03151,

    arXiv

  8. [26]

    URL https: //doi.org/10.24963/ijcai.2020/671

    doi: 10.24963/ijcai.2020/671. URL https: //doi.org/10.24963/ijcai.2020/671. Qian, C., Liu, M. X., Reif, E., Simon, G., Hussein, N., Clement, N., Wexler, J., Cai, C. J., Terry, M., and Kahng, M. The evolution of llm adoption in industry data curation practices. arXiv preprint arXiv:2412.16089,

    arXiv 2020

Show all 31 references
  1. [28]

    10 Language Games as the Pathway to Artificial Superhuman Intelligence Shah, C

    URL https: //arxiv.org/abs/2411.16905. 10 Language Games as the Pathway to Artificial Superhuman Intelligence Shah, C. and White, R. W. Agents are not enough. ArXiv preprint, abs/2412.16241,

    arXiv

  2. [29]

    Shneiderman, B

    URL https: //arxiv.org/abs/2412.16241. Shneiderman, B. Human-centered ai: A new synthesis. In Human-Computer Interaction–INTERACT 2021: 18th IFIP TC 13 International Conference, Bari, Italy, August 30–September 3, 2021, Proceedings, Part I 18, pp. 3–8. Springer,

    2021 arXiv

  3. [31]

    Si, C., Yang, D., and Hashimoto, T

    URL https://arxiv.org/ abs/2305.17493. Si, C., Yang, D., and Hashimoto, T. Can llms generate novel research ideas? a large-scale human study with 100+ nlp researchers. arXiv preprint arXiv:2409.04109,

    arXiv

  4. [34]

    Sutton, R

    URL https://arxiv.org/ abs/2501.05707. Sutton, R. S. and Barto, A. G. Reinforcement learning: past, present and future. In IEEE International Symposium on Approximate Dynamic Programming and Reinforcement Learning, pp. 195–197,

    arXiv

  5. [37]

    Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., and Zhou, D

    URL https://arxiv.org/ abs/2412.14872. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A., and Zhou, D. Self-consistency im- proves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171,

    arXiv

  6. [38]

    Densing law of llms

    Xiao, C., Cai, J., Zhao, W., Zeng, G., Han, X., Liu, Z., and Sun, M. Densing law of llms. arXiv preprint arXiv:2412.04315,

    arXiv

  7. [39]

    Yang, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu, B., Li, C., Liu, D., Huang, F., Wei, H., et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024a. Yang, J., Yang, S., Gupta, A. W., Han, R., Fei-Fei, L., and Xie, S. Thinking in space: How multimodal large...

    arXiv

  8. [40]

    Zhang, A

    URL https://arxiv.org/abs/2102.07659. Zhang, A. et al. On the phenomenon of self-collapse in language model fine-tuning,

    arXiv

  9. [41]

    Zuboff, S

    URL https: //arxiv.org/abs/2301.00000. Zuboff, S. The age of surveillance capitalism: The fight for a human future at the new frontier of power, edn. PublicAffairs, New York,

  10. [1859]

    R., Sohl-Dickstein, J., Fiedel, N., Warkentin, T., Dafoe, A., Faust, A., Farabet, C., and Legg, S

    Morris, M. R., Sohl-Dickstein, J., Fiedel, N., Warkentin, T., Dafoe, A., Faust, A., Farabet, C., and Legg, S. Levels of agi: Operationalizing progress on the path to agi. arXiv preprint arXiv:2311.02462,

  11. [1954]

    Su, H., Sun, R., Yoon, J., Yin, P., Yu, T., and Arık, S. ¨O. Learn-by-interact: A data-centric framework for self- adaptive agents in realistic environments. arXiv preprint arXiv:2501.10893,

    arXiv

  12. [1999]

    Evolving deeper llm thinking

    Lee, K.-H., Fischer, I., Wu, Y .-H., Marwood, D., Baluja, S., Schuurmans, D., and Chen, X. Evolving deeper llm thinking. arXiv preprint arXiv:2501.09891,

    arXiv

  13. [2001]

    Automatic curriculum learning for deep RL: A short survey

    Portelas, R., Colas, C., Weng, L., Hofmann, K., and Oudeyer, P. Automatic curriculum learning for deep RL: A short survey. In Bessiere, C. (ed.), Proceedings of the Twenty-Ninth International Joint Conference on Arti- ficial Intelligence, IJCAI 2020, pp. 4819–4825. ijcai.org,

    2020

  14. [2002]

    D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. Language models are few-shot learners. Advances in neural information processing systems, 33: 1877–1901,

    1901

  15. [2010]

    Llama: Open and efficient foundation lan- guage models

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.-A., Lacroix, T., Rozi`ere, B., Goyal, N., Hambro, E., Azhar, F., et al. Llama: Open and efficient foundation lan- guage models. arXiv preprint arXiv:2302.13971,

    arXiv

  16. [2015]

    Complementarity in human-ai collabora- tion: Concept, sources, and evidence

    Hemmer, P., Schemmer, M., K ¨uhl, N., V ¨ossing, M., and Satzger, G. Complementarity in human-ai collabora- tion: Concept, sources, and evidence. arXiv preprint arXiv:2404.00029,

    arXiv

  17. [2018]

    Jaech, A., Kalai, A., Lerer, A., Richardson, A., El-Kishky, A., Low, A., Helyar, A., Madry, A., Beutel, A., Car- ney, A., et al

    URL https://arxiv.org/abs/1805.00899. Jaech, A., Kalai, A., Lerer, A., Richardson, A., El-Kishky, A., Low, A., Helyar, A., Madry, A., Beutel, A., Car- ney, A., et al. Openai o1 system card. arXiv preprint arXiv:2412.16720,

    arXiv

  18. [2020]

    doi: 10.18653/v1/2020.acl-main.463

    Association for Computational Linguis- tics. doi: 10.18653/v1/2020.acl-main.463. URL https: //aclanthology.org/2020.acl-main.463. Bengio, Y ., Louradour, J., Collobert, R., and Weston, J. Curriculum learning. In Danyluk, A. P., Bottou, L., and Littman, M. L. (eds.), Proceeding...

    2020 doi

  19. [2021]

    Bousmalis, K., Ding, Q., Baker, J., et al

    URL https:// arxiv.org/abs/2108.07258. Bousmalis, K., Ding, Q., Baker, J., et al. Robocat: A self- improving foundation agent for robotic manipulation. pp. 1–16,

    arXiv

  20. [2022]

    R., and Feng, S

    Pan, J., He, H., Bowman, S. R., and Feng, S. Spontaneous reward hacking in iterative self-refinement.arXiv preprint arXiv:2407.04549,

    arXiv

  21. [2023]

    Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al

    URL https://arxiv.org/ abs/2305.14325. Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783,

    arXiv

  22. [2024]

    From persona to personalization: A survey on role-playing language agents

    Chen, J., Wang, X., Xu, R., Yuan, S., Zhang, Y ., Shi, W., Xie, J., Li, S., Yang, R., Zhu, T., et al. From persona to personalization: A survey on role-playing language agents. arXiv preprint arXiv:2404.18231,

    arXiv

  23. [2025]

    Deepseek-r1: In- centivizing reasoning capability in llms via reinforcement learning

    Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. Deepseek-r1: In- centivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948,

    arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.