Pith. sign in

REVIEW 3 major objections 5 minor 26 references

Human participation persists in AI tasks because some goals only become definite through interaction.

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

2026-08-01 07:05 UTC pith:MA2OENEG

load-bearing objection Useful taxonomy, genuine emergence distinction—but the necessity claim rests on an assumed psychological premise, and the formal model is a representation, not a proof. the 3 major comments →

arxiv 2607.21547 v1 pith:MA2OENEG submitted 2026-07-23 cs.AI cs.CLcs.ETcs.LGcs.MA

The Boundaries of Automation: A Theory of Persistent Human Participation

Fares Fourati , Hinrich Sch\"utze , Eyke H\"ullermeier , Iryna Gurevych This is my paper
classification cs.AI cs.CLcs.ETcs.LGcs.MA
keywords human–AI co-constructiontarget emergencelimits of automationhuman-in-the-loopconstructive preferencesinteraction dynamicsAI alignmentevaluation of AI systems
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 challenges the assumption that humans stay in the loop only because AI is still imperfect. It distinguishes three independent grounds for human participation: complementarity, normative or developmental value, and — the central argument — target emergence. In emergence-ground tasks, the evaluative target is not fixed in advance; it is revealed, refined, or constituted through interaction with candidate artifacts. If the argument holds, human involvement is a persistent design requirement for a class of tasks, not a temporary patch. The paper models this as a coupled dynamic system in which the target evolves with the interaction history.

Core claim

The paper's central claim is boxed: human–AI co-construction may remain necessary not only because AI has technical limitations, but because in some tasks the target becomes determinate only through interaction. A target — the criterion-bearing structure that fixes what counts as success — can change as artifacts, explanations, and comparisons are generated and evaluated. The interaction does not merely estimate a pre-existing target; it helps produce the evaluative conditions under which a target is endorsed. Even an oracle that correctly predicts the participant's eventual target cannot bypass the interaction unless it reproduces, in compressed form, the formative trajectory that would hav

What carries the argument

The load-bearing device is a history-dependent dynamical model in which the current target G_t is one component of a joint state (target, artifact, execution, evaluation), and the next target is given by an update process over the full interaction history. Complementary to this formal skeleton are conceptual distinctions: between artifact-level, execution-level, and target-level interaction, and among three forms of emergence — revelation, refinement, and constitution. These distinctions let the paper argue that interaction can change the standard of success, not just the artifact produced.

Load-bearing premise

The argument rests on the psychological premise that final endorsement of a target depends, at least partly, on the evaluative development produced by the interaction trajectory; if people can simply endorse a well-justified predicted target without that trajectory, the emergence ground loses its necessity.

What would settle it

Controlled experiment: start matched participants from the same initial target; one group iterates with an AI through proposals and revisions, the other receives the final target and artifact in one step with a compelling explanation. If the one-step group endorses at the same rate and with the same stability as the iterating group, the constitutive-trajectory claim is falsified. A complementary check: if two different proposal sequences produce no systematic difference in final target, path-dependence evidence for emergence is absent.

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

If this is right

  • Automation has a task-dependent boundary: tasks with stable, already-operative targets can be automated, while emergent-target tasks will retain human participation even as AI improves.
  • System design should aim to support target formation — comparison, reversibility, transparency, preservation of alternative trajectories — not merely optimize execution under a fixed objective.
  • Evaluation of AI systems should ask how well they help users clarify, inspect, negotiate, refine, and transform their targets, not just how faithfully they follow instructions.
  • Ethical governance of co-construction must distinguish influence (inherent in any suggestion or framing) from manipulation (steering while undermining the user's capacity to recognize and redirect the steering).
  • Alignment methods that assume static preferences will mis-specify objectives whenever target emergence is present.

Where Pith is reading between the lines

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

  • Editorial extension: if target emergence holds, preference-elicitation and alignment frameworks should be reconceived as governing a trajectory rather than recovering a fixed preference; the target is partly constituted by the elicitation process itself.
  • Editorial extension: a direct testable prediction follows — matched participants exposed to different sequences of AI proposals should converge to measurably different final targets in emergence-ground tasks, and these differences should persist beyond ordinary change of mind.
  • Editorial extension: the 'compressed persuasion' argument suggests that even a persuasive explanation is a form of co-construction as long as it alters the evaluative state; this could blur the line between assistance and manipulation in a way the paper does not fully develop.
  • Editorial extension: the stochastic argument in the appendix implies the barrier to bypassing co-construction is irreducible uncertainty in the target trajectory, not computational intractability; that suggests the limit is principled rather than contingent on model scale.

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

3 major / 5 minor

Summary. The paper challenges the assumption that human participation in AI systems is only a stopgap for insufficient capability. It distinguishes three grounds for persistent human–AI co-construction: technical/complementarity, normative/developmental, and emergence. The central thesis is that in some tasks the evaluative target is not fixed in advance but emerges through interaction, so human participation is constitutive of the target and may remain necessary even for highly capable AI. The paper develops a taxonomy of target emergence (revealed, refined, constituted targets), a dynamic model in which targets evolve as part of the interaction history, and an extended discussion of objections, especially the possibility that an AI oracle predicts the eventual target. It concludes with empirical, ethical, and design implications.

Significance. If the emergence-ground thesis could be established, it would be an important contribution to the automation debate: it identifies a class of tasks in which human participation is not a temporary patch but a persistent design requirement. The paper is philosophically careful: it uses modal language, distinguishes artifact-, execution-, and target-level interaction, and engages seriously with the strongest oracle-based objection. It also offers concrete empirical predictions (§4.1) and design principles. The formal and argumentative support, however, is weaker than the conclusion; the formal model is largely definitional and the key psychological premise is asserted rather than argued.

major comments (3)
  1. [§3.6, 'Convincing the human with a predicted target'] The response to the oracle objection rests on the premise that 'endorsement depends, at least in part, on the evaluative development produced by that trajectory.' This is the very claim the objection disputes. The next paragraph concedes that a system may present the predicted target with 'a compelling explanation' and that this is co-construction 'compressed into a single persuasive exchange.' But if one persuasive exchange can reproduce the formative effects, then the extended trajectory is not necessary; if it cannot, the paper must explain why explanation cannot substitute. As written, the argument either reduces to the truism that a human must be persuaded before accepting an output, or it asserts the trajectory-dependence it needs to prove. The formal model does not resolve this; it assumes relevant stochasticity rather than deriving it from target emergence.
  2. [Appendix D.1] The impossibility result is assumption-driven. The model sets G_{t+1}=G_t⊕ξ_t with ξ_t~Bernoulli(ε), ε>0, and then shows that exact trajectory reconstruction fails with probability approaching 1. But the irreducible unobserved disturbance ε is stipulated, not derived from any feature of emergence-ground tasks. Without ε>0 the update rule is deterministic and exact reconstruction is trivial. Moreover, the stronger oracle objection in §3.6 only requires predicting the eventual endorsed target, not reconstructing the whole realized trajectory; the exponentially decaying bound on Pr(Ĝ_{1:T}=G_{1:T}) does not address endpoint prediction. The appendix therefore cannot carry the load of the 'Modeling target evolution' objection.
  3. [§3.3, Eq. (1)] Equation (1), G_{t+1}=Φ_G(H_t), is a definitional existence statement: because Φ_G is unrestricted, any target trajectory can be represented in this form. The admission in Appendix D that the formulation 'makes no assumptions about functional form, computability, or observability' confirms that the model has no empirical content. Presenting this as a 'dynamic model' that 'marks the departure from fixed-objective optimization' overstates what the formalism establishes. The paper should either label Eq. (1) as purely representational, or add substantive constraints that distinguish target emergence from mere history-dependence.
minor comments (5)
  1. [Appendix C] Typo: 'explainswhymeaningful' should be 'explains why meaningful'.
  2. [Figure 1] The tandem-bicycle metaphor is decorative rather than informative; consider either integrating it with the argument or removing it.
  3. [§4.1] The empirical predictions are carefully hedged, but the paper could strengthen them by proposing specific experimental designs that would distinguish target emergence from latent-target articulation with random noise.
  4. [References] Some entries are incomplete or inconsistent (e.g., Fourati 2026, Pramod 2026); a final proofread for consistency would help.
  5. [§3.5] The definition of ∆_{t,k} uses 0≤k<t but the moving average notation is not explicit about boundary cases; clarify whether k is window length or number of past rounds.

Circularity Check

2 steps flagged

The necessity claim is defended by classifying any successful explanation as co-construction, and the formal impossibility result assumes the irreducible noise it then concludes is the obstacle.

specific steps
  1. self definitional [Section 3.6, 'Convincing the human with a predicted target']
    "However, if the explanation changes the participant’s evaluative state in ways that shape which target the participant ultimately accepts, then the explanation itself functions as part of the target-formation process. ... In such a case, co-construction has not disappeared; it has been compressed into a single persuasive exchange."

    The paper's central thesis is that co-construction is necessary because targets become determinate only through interaction. To answer the objection that an AI could predict the target and present it with a compelling explanation, the paper defines any successful explanation that shapes acceptance as part of the target-formation process, hence as co-construction. This makes 'co-construction cannot be bypassed' true by definition: any exchange that changes the target counts as co-construction. The contested premise—that endorsement depends on the trajectory of comparisons and revisions—is reasserted rather than derived. If a one-shot explanation can produce the same endorsement, the necessity claim reduces to the trivial point that a human must be persuaded; if it cannot, the paper has mere

  2. self definitional [Appendix D.1, 'On Simulating the Whole Trajectory from Scratch' (equations D.1–D.2)]
    "Suppose the target evolves according to Gt+1 = Gt ⊕ ξt, where ⊕ denotes exclusive disjunction and ξt ∼ Bernoulli(ε), 0 < ε < 1/2, represents an unobserved disturbance arising from latent cognitive states, attention, interpretation, learning, or other stochastic influences. ... The obstacle is therefore not computational complexity, but irreducible uncertainty in target emergence itself."

    The model assumes irreducible stochastic disturbances ε > 0 in the target-update rule, then derives Pr(Ĝ1:T = G1:T | G0) ≤ (1−ε)^T. The subsequent conclusion that the obstacle is 'irreducible uncertainty in target emergence itself' restates the assumption: the uncertainty was inserted into the update equation, not derived from any feature of emergence tasks. Thus the formal impossibility of exact trajectory reconstruction is contained in the model's assumptions by construction. Moreover, the stated objection only requires endpoint prediction—which the paper grants via an oracle—so this trajectory-reconstruction result does not address the actual bypass argument without the extra, unproved assumption that the formative trajectory is indispensable.

full rationale

The paper is not a pure renaming exercise: it draws on substantial external literature (Simon, Slovic, Roy, Rittel and Webber, Maher and Poon, Suchman) and Section 4.1 offers empirically testable predictions such as path dependence and reordering of preferences. There is no load-bearing self-citation chain; self-citations such as Dutta et al. (2025) are used for related work, not as the basis of the central claim. However, two circular moves prevent a low score. First, the response to the oracle objection defines co-construction functionally so that any successful persuasive explanation counts as compressed co-construction; this makes the necessity claim partly true by definition and presupposes the very trajectory-dependence of endorsement that the objection challenges. Second, Appendix D.1 builds irreducible noise ε > 0 into the target update and then concludes that irreducible uncertainty is the obstacle; the formality restates its own input. The paper explicitly limits D.1 to an existence claim, so the formal model alone does not prove that real emergence tasks contain such noise, but Section 3.6 leans on it when answering the 'modeling target evolution' objection. Because the central claim has substantial independent conceptual content yet several key steps reduce to definitional moves, a score of 6 is appropriate rather than 0–2 or 8–10.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 2 invented entities

The paper's central argument rests on domain assumptions about human target formation and one ad hoc stochastic model; it introduces no physical entities. The main free 'parameter' is the irreducible uncertainty epsilon in the illustrative impossibility model. Overall, the formal contribution is largely representational; the substantive burden is carried by the domain assumptions listed above.

free parameters (2)
  • epsilon (irreducible one-step disturbance probability) = 0 < epsilon < 1/2 (illustrated with 0.05)
    Chosen by hand in Appendix D.1 to make exact trajectory reconstruction impossible; no empirical estimate. It encodes the conclusion that target evolution is irreducibly unpredictable.
  • D (target dissimilarity measure) = unspecified abstract function
    Introduced in Section 3.5 to define Delta_t; no empirical or theoretical calibration, but used only illustratively.
axioms (4)
  • domain assumption Human-driven targets are grounded in the participant's ongoing evaluative activity (Definition 1).
    Defines the object of study and excludes cases where targets are externally fixed or machine-grounded, so the necessity claim is limited to this class.
  • domain assumption There exist tasks in which the evaluative target is not fully determined before interaction and is changed by it (Section 3.1).
    Existence claim used to motivate the emergence ground; supported only by examples and citations, not by a systematic empirical survey.
  • domain assumption Final endorsement of a target depends at least partly on the evaluative development produced by the interaction trajectory (Section 3.6, coffee-shop example).
    This is the load-bearing psychological premise; if false, a perfect predictor could bypass co-construction by presenting the predicted target and artifacts.
  • ad hoc to paper Irreducible unobserved disturbances xi_t with epsilon > 0 affect target transitions (Appendix D.1).
    Imported solely to demonstrate that exact advance reconstruction can fail; not derived from target emergence or from empirical data.
invented entities (2)
  • Target (G_t) as a criterion-bearing evaluative structure no independent evidence
    purpose: Umbrella construct to speak uniformly about objectives, preferences, rewards, and instructions, and to allow target emergence to be stated representation-independently.
    Explicitly introduced as a term of art (Appendix A); no measurement protocol or falsifiable handle is provided, so it is an analytical entity rather than an independently evidenced one.
  • Evaluative state S_t and its projections (S^G_t, S^X_t, S^E_t, S^X|G,E_t, S^E|G,X_t) no independent evidence
    purpose: Analytical decomposition distinguishing endorsement of target, artifact, and execution; used to define convergence, exploration, stable dissatisfaction, and branching.
    Functional rather than psychological decomposition (Appendix D); no empirical validation that these components map onto measurable mental states.

pith-pipeline@v1.3.0-alltime-deepseek · 21428 in / 11233 out tokens · 121474 ms · 2026-08-01T07:05:43.917755+00:00 · methodology

0 comments
read the original abstract

The rapid progress of AI has intensified the long-standing pursuit of automation: replacing human participation with algorithms wherever possible. Implicit in this pursuit is the assumption that humans remain in the loop only because current AI systems are not yet sufficiently capable. This paper challenges that assumption. Rather than asking how far automation can extend, we ask where its conceptual limits lie and argue that human participation may persist even with highly capable AI systems for three distinct reasons. Technical or complementarity grounds arise when humans contribute capabilities or perspectives unavailable to AI. Normative or developmental grounds arise when participation itself is valuable for human agency or learning. Most importantly, emergence grounds arise from target emergence: in some activities, the target is not fully specified in advance but instead emerges through the interaction itself. In these cases, human participation is not merely a means of improving execution but is constitutive of the target being produced. Human--AI co-construction, understood as the joint production of outcomes by humans and AI systems, is therefore not simply a temporary response to imperfect AI, but a persistent feature of activities whose objectives emerge through participation. This perspective has important implications for the limits of automation and for the design, evaluation, and ethics of future AI systems.

Figures

Figures reproduced from arXiv: 2607.21547 by Eyke H\"ullermeier, Fares Fourati, Hinrich Sch\"utze, Iryna Gurevych.

Figure 1
Figure 1. Figure 1: A motivating metaphor for target emergence in human–AI co-construction. The tandem bicycle symbolizes joint participation in an exploratory journey whose destination is not necessarily fully specified in advance. The AI’s gesture represents the introduction of new possibilities and directions that may reveal, refine, or transform the target through interaction. et al., 2020], longer context windows [Bai et… view at source ↗
Figure 2
Figure 2. Figure 2: Levels of human–AI interaction. Artifact-level interaction modifies the output, executional interaction modifies how it is produced, and target-level interaction modifies what counts as success. Horizontal arrows denote changes within a level; vertical arrows indicate that changes at one level may affect another. Schön, 2017, De Peuter et al., 2023, Huellermeier and Słowiński, 2024, Kanwal and Tran, 2026].… 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

26 extracted references · 1 canonical work pages

  1. [3]

    Language models are few-shot learners.Advances in Neural Information Processing Systems, 33:1877–1901,

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, 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,

    1901

  2. [5]

    Posix: A prompt sensitivity index for large language models

    Anwoy Chatterjee, HSVNS Kowndinya Renduchintala, Sumit Bhatia, and Tanmoy Chakraborty. Posix: A prompt sensitivity index for large language models. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 14550–14565,

    2024

  3. [8]

    A coherence-based measure of agi.arXiv preprint arXiv:2510.20784,

    Fares Fourati. A coherence-based measure of agi.arXiv preprint arXiv:2510.20784,

    arXiv

  4. [9]

    DeepSeek-R1: Incen- tivizing reasoning capability in LLMs via reinforcement learning.arXiv preprint arXiv:2501.12948,

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. DeepSeek-R1: Incen- tivizing reasoning capability in LLMs via reinforcement learning.arXiv preprint arXiv:2501.12948,

    Pith/arXiv arXiv

  5. [11]

    Why do we do astrophysics?arXiv preprint arXiv:2602.10181,

    David W Hogg. Why do we do astrophysics?arXiv preprint arXiv:2602.10181,

    arXiv

  6. [13]

    URLhttps://openreview.net/forum?id= f7OkIurx4b

    ISSN 2835-8856. URLhttps://openreview.net/forum?id= f7OkIurx4b. Survey Certification. Salma Kharrat, FaresFourati, and Marco Canini. Acing: Actor-critic for instruction learning in black-box LLMs. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 19086–19113,

    2025

  7. [15]

    Eliciting human preferences with language models

    Belinda Li, Alex Tamkin, Noah Goodman, and Jacob Andreas. Eliciting human preferences with language models. InInternational Conference on Learning Representations, volume 2025, pages 80984–81013,

    2025

  8. [18]

    ISBN 979-8-89176-390-6

    Association for Computational Linguistics. ISBN 979-8-89176-390-6. doi: 10.18653/v1/2026. acl-long.567. URLhttps://aclanthology.org/2026.acl-long.567/. Dhruv Sahnan, Subhabrata Dutta, Tanmoy Chakraborty, Preslav Nakov, and Iryna Gurevych. Co-factchecker: A framework for human-AI collaborative claim 30 verification using large reasoning models.arXiv prepri...

    Pith/arXiv arXiv 2026

  9. [23]

    32 Yilun Zhu, Joel Ruben Antony Moniz, Shruti Bhargava, Jiarui Lu, Dhivya Piravipe- rumal, Site Li, Yuan Zhang, Hong Yu, and Bo-Hsiang Tseng

    URLhttps://proceedings.mlr.press/v235/zhou24t.html. 32 Yilun Zhu, Joel Ruben Antony Moniz, Shruti Bhargava, Jiarui Lu, Dhivya Piravipe- rumal, Site Li, Yuan Zhang, Hong Yu, and Bo-Hsiang Tseng. Can large language models understand context? InFindings of the Association for Computational Linguistics: EACL 2024, pages 2004–2018,

    2024

  10. [24]

    A call for collaborative intelligence: Why human-agent systems should precede AI autonomy.arXiv preprint arXiv:2506.09420,

    Henry Peng Zou, Wei-Chieh Huang, Yaozu Wu, Chunyu Miao, Dongyuan Li, Aiwei Liu, Yue Zhou, Yankai Chen, Weizhi Zhang, Yangning Li, et al. A call for collaborative intelligence: Why human-agent systems should precede AI autonomy.arXiv preprint arXiv:2506.09420,

    Pith/arXiv arXiv

  11. [25]

    Henry Peng Zou, Wei-Chieh Huang, Yaozu Wu, Jizhou Guo, Yankai Chen, Chunyu Miao, Hoang H Nguyen, Yue Zhou, Weizhi Zhang, Liancheng Fang, Hanrong Zhang, Fangxin Wang, Pengfei Zhang, Langzhou He, Yangning Li, Dongyuan Li, Renhe Jiang, and Philip S. Yu. LLM-based human-agent collaboration and interaction systems: A survey. InFindings of the Association for C...

    2026

  12. [26]

    Modify the classifier so that it reports the predicted probability as well as the predicted class

    Association for Computational Linguistics. ISBN 979-8-89176-395-1. doi: 10.18653/v1/2026.findings-acl.1811. URL https://aclanthology.org/ 2026.findings-acl.1811/. 33 A Terminological Distinctions Table 2: Targets and related evaluative notions. Notion Primary focus Preference Expresses which available alternative is favored over another. Utility function ...

    work page doi:10.18653/v1/2026.findings-acl.1811 2026

  13. [1955]

    Herbert A

    doi: 10.2307/1884852. Herbert A. Simon. Rational choice and the structure of the environment.Psycho- logical Review, 63(2):129–138,

    doi:10.2307/1884852

  14. [1956]

    Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al

    doi: 10.1037/h0042769. Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al. OpenAI GPT-5 system card.arXiv preprint arXiv:2601.03267,

    Pith/arXiv arXiv

  15. [1962]

    A definition of agi.arXiv preprint arXiv:2510.18212,

    Dan Hendrycks, Dawn Song, Christian Szegedy, Honglak Lee, Yarin Gal, Erik Brynjolfsson, Sharon Li, Andy Zou, Lionel Levine, Bo Han, et al. A definition of agi.arXiv preprint arXiv:2510.18212,

    arXiv

  16. [1965]

    ARC-AGI-3: A new challenge for frontier agentic intelligence

    ARC Foundation. ARC-AGI-3: A new challenge for frontier agentic intelligence. arXiv preprint arXiv:2603.24621,

    Pith/arXiv arXiv

  17. [1993]

    Expert preference- based evaluation of automated related work generation.arXiv preprint arXiv:2508.07955,

    Furkan Şahinuç, Subhabrata Dutta, and Iryna Gurevych. Expert preference- based evaluation of automated related work generation.arXiv preprint arXiv:2508.07955,

    Pith/arXiv arXiv

  18. [1995]

    Scaling LLM test-time compute optimally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314,

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. Scaling LLM test-time compute optimally can be more effective than scaling model parameters.arXiv preprint arXiv:2408.03314,

    Pith/arXiv arXiv

  19. [1998]

    Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,

    31 Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,

    Pith/arXiv arXiv

  20. [2014]

    ToolLLM: Facilitating large language models to master 16000+ real-world apis

    Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, et al. ToolLLM: Facilitating large language models to master 16000+ real-world apis. InInternational Conference on Learning Representations, volume 2024, pages 9695–9717,

    2024

  21. [2016]

    Constructive alignment: Governing preference dynamics in human-AI interaction.arXiv preprint arXiv:2607.00001,

    Max Kanwal and Caryn Tran. Constructive alignment: Governing preference dynamics in human-AI interaction.arXiv preprint arXiv:2607.00001,

    Pith/arXiv arXiv

  22. [2020]

    Sequor: A multi-turn benchmark for realistic constraint following.arXiv preprint arXiv:2605.06353,

    Beatriz Canaverde, Duarte M Alves, José Pombal, Giuseppe Attanasio, and André FT Martins. Sequor: A multi-turn benchmark for realistic constraint following.arXiv preprint arXiv:2605.06353,

    Pith/arXiv arXiv

  23. [2022]

    28 Thomas Kwa, Ben West, Joel Becker, Amy Deng, Katharyn Garcia, Max Hasin, Sami Jawhar, Megan Kinniment, Nate Rush, Sydney Von Arx, et al

    doi: 10.1109/ ICOA55659.2022.9934291. 28 Thomas Kwa, Ben West, Joel Becker, Amy Deng, Katharyn Garcia, Max Hasin, Sami Jawhar, Megan Kinniment, Nate Rush, Sydney Von Arx, et al. Measuring AI ability to complete long software tasks.Advances in Neural Information Processing Systems, 38:92213–92266,

    arXiv 2022

  24. [2024]

    International AI safety report 2026.arXiv preprint arXiv:2602.21012,

    Yoshua Bengio, Stephen Clare, Carina Prunkl, Maksym Andriushchenko, Ben Bucknall, Malcolm Murray, Rishi Bommasani, Stephen Casper, Tom Davidson, Raymond Douglas, et al. International AI safety report 2026.arXiv preprint arXiv:2602.21012,

    arXiv 2026

  25. [2025]

    Logic: The theory of inquiry, 1938.The Later Works of John Dewey, 12,

    John Dewey. Logic: The theory of inquiry, 1938.The Later Works of John Dewey, 12,

    1938

  26. [2026]

    a is b” fail to learn “b is a

    Lukas Berglund, Meg Tong, Maximilian Kaufmann, Mikita Balesni, Asa Stickland, Tomek Korbak, and Owain Evans. The reversal curse: LLMs trained on “a is b” fail to learn “b is a”. InInternational Conference on Learning Representations, volume 2024, pages 18623–18642,

    2024