pith. sign in

arxiv: 2606.04421 · v2 · pith:U5QIL2EFnew · submitted 2026-06-03 · 💻 cs.AI · cs.LG

Trivium: Temporal Regret as a First-Class Objective for Causal-Memory Controllers

Edward Y. Chang This is my paper

Pith reviewed 2026-06-28 06:34 UTC · model grok-4.3

classification 💻 cs.AI cs.LG
keywords temporal regretcausal memoryepistemic regretagentic systemscausal probinglong-horizon agentschange-point detection
0
0 comments X

The pith

Agents achieve O(log E) temporal regret by maintaining a persistent causal log and using budgeted probes instead of outcome-only optimization.

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

Current agentic systems and LLM pipelines optimize outcome reward alone, so the same causal miscalibrations can recur across episodes without logging when or why they persist. The paper treats temporal regret—the duration a faulty causal model is tolerated—as a first-class objective alongside outcome regret and epistemic regret over the working causal model. Modeling the agent as a stream of E episodes, it proves three conditional results: outcome-only learning cannot separate causal from spurious structure without interventions, a persistent causal log with budgeted probes yields logarithmic probe complexity and thus O(log E) temporal regret, and the bound becomes O(K log E) under K detectable change-points. Experiments on CausalBench-Seq show the method tracking the predicted logarithmic envelope while outcome-only baselines grow linearly.

Core claim

Modeling the agent as a stream of E episodes, under explicit causal-probing, persistence, and detectability assumptions, a persistent causal log with budgeted probes induces O(log E) temporal regret; the rate extends to O(K log E) under K detectable change-points.

What carries the argument

Temporal regret, the persistence duration of a miscalibrated causal model, tracked by the Trivium controller that jointly minimizes outcome, epistemic, and temporal regrets via a persistent causal log and budgeted probes.

If this is right

  • Outcome-only learning leaves temporal miscalibration persisting linearly even after outcome regret reaches zero.
  • With a persistent causal log and budgeted probes, total probe complexity stays logarithmic in the episode horizon.
  • Under K detectable change-points the temporal regret bound extends to O(K log E).
  • Trivium tracks the predicted logarithmic envelope on CausalBench-Seq while outcome-only baselines grow linearly.

Where Pith is reading between the lines

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

  • Revising an external causal model without retraining LLM weights could enable self-correction in deployed long-lived agents.
  • The same logarithmic scaling might extend to other persistent memory mechanisms in sequential decision systems.
  • Combining the three-regret controller with existing online learning methods could produce hybrid architectures for real-world streams.

Load-bearing premise

The environment allows explicit causal probing, causal structure persists across episodes, and change-points remain detectable.

What would settle it

Trivium exhibiting linear rather than logarithmic growth in temporal regret on CausalBench-Seq would falsify the O(log E) bound.

Figures

Figures reproduced from arXiv: 2606.04421 by Edward Y. Chang.

Figure 1
Figure 1. Figure 1: Computational structure of temporal-regret minimization. Solid-bordered boxes are Trivium contributions; dashed-bordered boxes are load-bearing prior-work substrate (planning en￾vironment and CTL realization). Zone 2 (hatched) is a persistent transactional Causal Transaction Log. Information flows regret → log → learning → replan: Zone 1 emits three regret signals; Zone 2 logs one cross-agent entry per epi… view at source ↗
Figure 2
Figure 2. Figure 2: RQ2: cross-episode rate. CausalBench-Seq cumulative cross-episode temporal regret, 20 seeds, E=500. Left: per-episode trace with log-fit. Right: E ∈ {50, 100, 200, 300, 500} checkpoint values inside Theorems 3.4/3.5’s envelope. Trivium lies inside; RLVR grows linearly. RQ1: Does the epistemic signal separate from the outcome signal? This is the necessary con￾dition for any three-regret functional: if outco… view at source ↗
Figure 3
Figure 3. Figure 3: RQ3: drift-robust regret with K change-points. Left: cumulative regret under K ∈ {0, 1, 3, 5} change-points injected into CausalBench-Seq. Right: linear-in-K scaling at E=500 with slope 0.95 per change-point (R2=0.999), consistent with Theorem 3.8’s fixed-parameter O(K log E) specialization; the noCUSUM ablation overlaps in this regime because topology flips lie in the LRCP-detectable intersection (App. V.… view at source ↗
Figure 4
Figure 4. Figure 4: Exp A.1, RQ4. LRCP geometric-contraction trace on CausalBench-Seq, single disruption [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustrative example of temporal regret in a cross-episode control setting. Outcome-only [PITH_FULL_IMAGE:figures/full_fig_p039_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Trivium in relation to adjacent prior work. Trivium composes over a shared substrate [PITH_FULL_IMAGE:figures/full_fig_p039_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ablation A8 budget sweep (Be = α m0 log(e+1), 10 seeds, E=300). Left: cumulative RCE temp(e) fans out before commit and collapses to a common α-independent post-commit slope of 0.010 ± 0.001/ep. Right: commit time vs. 1/α, τ ≈ 1.29 + 2.87/α at R2 = 0.994, matching Corollary 3.6. Status. P2 corroborated quantitatively by the Ablation A8 budget sweep ( [PITH_FULL_IMAGE:figures/full_fig_p048_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Dispatch-coupling stand-in for the confounder-scaling envelope (Ablation A6; [PITH_FULL_IMAGE:figures/full_fig_p049_8.png] view at source ↗
Figure 3
Figure 3. Figure 3: Status. P4 corroborated by the CausalBench-Seq K-sweep (Ablation A7): regret grows linearly in K with slope 0.95/cp at R2 = 0.999, and per-change-point recovery time is constant at ≈ 5 episodes, matching Theorem 3.8’s O(K log E) rate and its O(log Epost) detection window. V.5 Ablation A1: Sandwich Tightness Across Horizons on CausalBench-Seq Setup. Stationary CausalBench-Seq (Appendix S), 15 seeds, Tep = 5… view at source ↗
Figure 9
Figure 9. Figure 9: Theorem 3.10 envelope calibration. The dashed diagonal is the predicted bound [PITH_FULL_IMAGE:figures/full_fig_p050_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Nonlinear-SCM stress test (App. V.14). (a) Trivium’s cumulative cross-episode tem￾poral regret across nonlinearity levels ν ∈ {0, 0.5, 1.0, 2.0} (10 seeds, E = 300). The log-shape is preserved across all four levels; the slope decreases as the tanh saturation cleans the interven￾tional signal. (b) Log-fit R2 versus nonlinearity for both controllers; Trivium remains above the qualitative-fit floor of 0.95 … view at source ↗
Figure 11
Figure 11. Figure 11: Job-shop-flavored stress test (App. V.15). [PITH_FULL_IMAGE:figures/full_fig_p055_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Plant-nonlinearity signature on real LLMs: per-model lift after strong audit on math [PITH_FULL_IMAGE:figures/full_fig_p059_12.png] view at source ↗
read the original abstract

Many current agentic systems and LLM pipelines correct mistakes by optimizing outcome reward. This addresses only the what of failure: when an outcome diverges from prediction, the why and when of the mismatch are not systematically logged, reviewed, or corrected, so the same error can recur episode after episode. We argue that this is a structural problem, not merely a model-capacity one. We propose long-horizon temporal regret as a first-class objective alongside outcome regret and epistemic regret over the working causal model. Temporal regret captures when failure persists: how long a miscalibrated causal model is tolerated before correction. Epistemic regret captures why failure persists: residual uncertainty or error in the working causal model. Together, the three regrets give a falsifiable account of what, why, and when a long-lived agent can fail. Modeling the agent as a stream of E episodes, we prove three conditional results under explicit causal-probing, persistence, and detectability assumptions. First, under observationally equivalent confounding, outcome-only learning cannot distinguish causal from spurious structure without an intervention channel, so temporal miscalibration can persist linearly even after outcome regret is driven to zero. Second, with a persistent causal log and budgeted probes, total probe complexity is logarithmic in the episode horizon, inducing O(log E) temporal regret. Third, under K detectable change-points, the rate extends to O(K log E). We instantiate Trivium and pre-register five falsifiable predictions. On CausalBench-Seq, Trivium follows the predicted logarithmic envelope while outcome-only baselines grow linearly. A pilot real-LLM stream provides preliminary external-validity evidence across one full E = 500 run and three E = 100 frontier-model pilots. Self-learning here means revising an external causal model, not retraining LLM weights.

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 paper proposes temporal regret as a first-class objective alongside outcome and epistemic regret for long-horizon causal-memory controllers. Modeling the agent as a stream of E episodes, it derives three conditional results under explicit causal-probing, persistence, and detectability assumptions: outcome-only learning permits linear persistence of miscalibration; persistent causal logs with budgeted probes yield O(log E) temporal regret; and K detectable change-points extend this to O(K log E). Trivium is instantiated with five pre-registered falsifiable predictions; on CausalBench-Seq it follows the predicted logarithmic envelope while outcome-only baselines grow linearly, with preliminary evidence from real-LLM streams.

Significance. If the conditional derivations hold and the assumptions are satisfied in the tested regimes, the framework supplies a falsifiable account of what, why, and when failures persist, moving beyond pure outcome optimization. Explicit strengths include the pre-registered predictions, the parameter-free O(log E) bound derived from the stated assumptions, and the separation of temporal from outcome regret. These elements could guide memory-augmented agent design if the empirical results are shown to operate under the required conditions.

major comments (2)
  1. [Abstract] Abstract (paragraph beginning 'Modeling the agent as a stream of E episodes'): The claim that Trivium 'follows the predicted logarithmic envelope' on CausalBench-Seq is presented as empirical support for the O(log E) result, yet no verification is reported that the benchmark satisfies the causal-probing, persistence, or detectability assumptions (e.g., that change-points are K-detectable or that causal structure persists across episodes). This is load-bearing for the central empirical claim, as the bound applies only conditionally and the linear baseline growth is the only unconditionally supported prediction.
  2. [Abstract] Abstract (modeling paragraph): The second and third conditional results require that budgeted probes are available and sufficient and that detectability holds; without explicit checks or arguments that CausalBench-Seq meets these conditions, the headline O(log E) versus linear separation cannot be interpreted as a test of the derived bound rather than a post-hoc pattern match.
minor comments (1)
  1. The relation between temporal regret, outcome regret, and epistemic regret is introduced conceptually but would benefit from an early formal definition or equation linking the three quantities before the conditional proofs.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and for emphasizing the conditional character of the derived bounds. The comments correctly identify that the abstract presents the CausalBench-Seq results as support for the O(log E) claim without reporting explicit checks on the modeling assumptions. We address each point below and commit to revisions that qualify the claims and supply the missing arguments.

read point-by-point responses

  1. Referee: [Abstract] Abstract (paragraph beginning 'Modeling the agent as a stream of E episodes'): The claim that Trivium 'follows the predicted logarithmic envelope' on CausalBench-Seq is presented as empirical support for the O(log E) result, yet no verification is reported that the benchmark satisfies the causal-probing, persistence, or detectability assumptions (e.g., that change-points are K-detectable or that causal structure persists across episodes). This is load-bearing for the central empirical claim, as the bound applies only conditionally and the linear baseline growth is the only unconditionally supported prediction.

    Authors: We agree that the abstract does not report verification that CausalBench-Seq satisfies the causal-probing, persistence, and detectability assumptions, and that this makes the headline separation harder to interpret as a direct test of the conditional bound. The linear growth of outcome-only baselines is indeed the only unconditional prediction. In revision we will (i) qualify the abstract sentence to read that the observed envelope is consistent with the pre-registered predictions under the stated assumptions, and (ii) add an explicit subsection in the experimental section that argues, from the benchmark construction, that causal structure persists across episodes, that probes are budgeted and sufficient, and that the K change-points are detectable by the logging mechanism. revision_made = 'yes' revision: yes

  2. Referee: [Abstract] Abstract (modeling paragraph): The second and third conditional results require that budgeted probes are available and sufficient and that detectability holds; without explicit checks or arguments that CausalBench-Seq meets these conditions, the headline O(log E) versus linear separation cannot be interpreted as a test of the derived bound rather than a post-hoc pattern match.

    Authors: We accept the referee's observation that the second and third results are conditional on budgeted probes being available and sufficient and on detectability holding, and that the manuscript provides no explicit argument that CausalBench-Seq satisfies these conditions. Consequently the reported separation cannot yet be read as a confirmatory test of the O(log E) or O(K log E) rates. In the revised version we will insert a short paragraph immediately after the benchmark description that (a) states the probe budget used, (b) confirms that the synthetic change-points are constructed to be K-detectable under the persistence assumption, and (c) notes that the observed logarithmic envelope is therefore only suggestive pending those checks. We will also tone down the abstract claim accordingly. revision_made = 'yes' revision: yes

Circularity Check

0 steps flagged

No significant circularity; results conditional on explicit assumptions with pre-registered predictions

full rationale

The paper explicitly derives the three conditional results (including O(log E) temporal regret) from the stated causal-probing, persistence, and detectability assumptions rather than by construction or fitting. It pre-registers five falsifiable predictions before reporting that Trivium follows the predicted logarithmic envelope on CausalBench-Seq. No self-citations, self-definitional steps, fitted inputs renamed as predictions, or ansatz smuggling appear in the text. The empirical claim is presented as a test of the pre-registered theoretical envelope, not a post-hoc match. The derivation chain is self-contained against the external benchmark under the listed assumptions.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 2 invented entities

The central claim rests on three domain assumptions explicitly named in the abstract and introduces two new regret quantities and the Trivium controller without independent external evidence for the new constructs.

axioms (3)
  • domain assumption causal-probing assumption
    Invoked to allow interventions that distinguish causal from spurious structure (abstract, modeling paragraph).
  • domain assumption persistence assumption
    Causal model and log persist across episodes (abstract, modeling paragraph).
  • domain assumption detectability assumption
    K change-points are detectable (abstract, third conditional result).
invented entities (2)
  • temporal regret no independent evidence
    purpose: Quantify duration of tolerated causal miscalibration
    New objective introduced to capture when failure persists.
  • Trivium controller no independent evidence
    purpose: Agent architecture that optimizes the three regrets
    The proposed system instantiated in the paper.

pith-pipeline@v0.9.1-grok · 5855 in / 1600 out tokens · 44915 ms · 2026-06-28T06:34:06.684636+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

88 extracted references · 7 canonical work pages · 3 internal anchors

  1. [1]

    Advances in Neural Information Processing Systems , volume=

    Causal bandits: Learning good interventions via causal inference , author=. Advances in Neural Information Processing Systems , volume=

  2. [2]

    Proceedings of the 34th International Conference on Machine Learning , volume=

    Identifying best interventions through online importance sampling , author=. Proceedings of the 34th International Conference on Machine Learning , volume=. 2017 , publisher=

    2017

  3. [3]

    Proceedings of the 35th International Conference on Machine Learning , series=

    Causal Bandits with Propagating Inference , author=. Proceedings of the 35th International Conference on Machine Learning , series=. 2018 , publisher=

    2018

  4. [4]

    Proceedings of the 23rd Annual Conference on Learning Theory , pages=

    Best arm identification in multi-armed bandits , author=. Proceedings of the 23rd Annual Conference on Learning Theory , pages=

  5. [5]

    Biometrika , volume=

    Continuous inspection schemes , author=. Biometrika , volume=. 1954 , doi=

    1954

  6. [6]

    The Annals of Mathematical Statistics , volume=

    Procedures for reacting to a change in distribution , author=. The Annals of Mathematical Statistics , volume=. 1971 , doi=

    1971

  7. [7]

    Statistica Sinica , volume=

    Sequential analysis: Some classical problems and new challenges , author=. Statistica Sinica , volume=

  8. [8]

    2009 , edition=

    Causality: Models, Reasoning, and Inference , author=. 2009 , edition=

    2009

  9. [9]

    2001 , publisher=

    Causation, Prediction, and Search , author=. 2001 , publisher=

    2001

  10. [10]

    Proceedings of the 18th National Conference on Artificial Intelligence (AAAI-02) , pages=

    A general identification condition for causal effects , author=. Proceedings of the 18th National Conference on Artificial Intelligence (AAAI-02) , pages=

  11. [11]

    Proceedings of the National Academy of Sciences , volume=

    Causal inference and the data-fusion problem , author=. Proceedings of the National Academy of Sciences , volume=. 2016 , doi=

    2016

  12. [12]

    Proceedings of the 2018 ACM SIGKDD Workshop on Causal Discovery , volume=

    Causal structure learning from multivariate time series in settings with unmeasured confounding , author=. Proceedings of the 2018 ACM SIGKDD Workshop on Causal Discovery , volume=. 2018 , publisher=

    2018

  13. [13]

    Advances in Neural Information Processing Systems , volume=

    Confounding-robust policy improvement , author=. Advances in Neural Information Processing Systems , volume=

  14. [14]

    Advances in Neural Information Processing Systems , volume=

    Off-policy policy evaluation for sequential decisions under unobserved confounding , author=. Advances in Neural Information Processing Systems , volume=

  15. [15]

    Offline Reinforcement Learning: Tutorial, Review, and Perspectives on Open Problems

    Offline reinforcement learning: Tutorial, review, and perspectives on open problems , author=. arXiv preprint arXiv:2005.01643 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 2005

  16. [16]

    Proceedings of the 1987 ACM SIGMOD International Conference on Management of Data , pages=

    Sagas , author=. Proceedings of the 1987 ACM SIGMOD International Conference on Management of Data , pages=. 1987 , doi=

    1987

  17. [17]

    2005 , note=

    Event sourcing , author=. 2005 , note=

    2005

  18. [18]

    Distributed Computing , volume=

    Causal memory: Definitions, implementation, and programming , author=. Distributed Computing , volume=

  19. [19]

    Garivier, Aur. The. Proceedings of the 24th Annual Conference on Learning Theory , volume=. 2011 , publisher=

    2011

  20. [20]

    Advances in Neural Information Processing Systems , volume=

    Hindsight experience replay , author=. Advances in Neural Information Processing Systems , volume=

  21. [21]

    [Anonymized venue] , year=

    [Anonymized prior work, A: transactional substrate for multi-agent LLM planning] , author=. [Anonymized venue] , year=

  22. [22]

    [Anonymized venue] , year=

    [Anonymized prior work, B: planning benchmark for LLMs and multi-agent systems] , author=. [Anonymized venue] , year=

  23. [23]

    2013 , publisher=

    Concentration Inequalities: A Nonasymptotic Theory of Independence , author=. 2013 , publisher=

    2013

  24. [24]

    2018 , publisher=

    High-Dimensional Probability: An Introduction with Applications in Data Science , author=. 2018 , publisher=

    2018

  25. [25]

    Proceedings of the 36th Conference on Uncertainty in Artificial Intelligence , volume=

    Regret analysis of bandit problems with causal background knowledge , author=. Proceedings of the 36th Conference on Uncertainty in Artificial Intelligence , volume=. 2020 , publisher=

    2020

  26. [26]

    Journal of Machine Learning Research , volume=

    Causal bandits for linear structural equation models , author=. Journal of Machine Learning Research , volume=

  27. [27]

    Advances in Neural Information Processing Systems , volume=

    Linear causal bandits: Unknown graph and soft interventions , author=. Advances in Neural Information Processing Systems , volume=. 2024 , note=

    2024

  28. [28]

    Advances in Neural Information Processing Systems , volume=

    Partial structure discovery is sufficient for no-regret learning in causal bandits , author=. Advances in Neural Information Processing Systems , volume=. 2024 , note=

    2024

  29. [29]

    and Ibeling, Duligur and Icard, Thomas , booktitle=

    Bareinboim, Elias and Correa, Juan D. and Ibeling, Duligur and Icard, Thomas , booktitle=. On. 2022 , publisher=

    2022

  30. [30]

    Philosophy of Science , volume=

    Interventions and causal inference , author=. Philosophy of Science , volume=. 2007 , doi=

    2007

  31. [31]

    Advances in Neural Information Processing Systems , volume=

    Jin, Zhijing and Chen, Yuen and Leeb, Felix and Gresele, Luigi and Kamal, Ojasv and Lyu, Zhiheng and Blin, Kevin and Gonzalez Adauto, Fernando and Kleiman-Weiner, Max and Sachan, Mrinmaya and Sch. Advances in Neural Information Processing Systems , volume=

  32. [32]

    Advances in Neural Information Processing Systems , volume=

    Unveiling Causal Reasoning in Large Language Models: Reality or Mirage? , author=. Advances in Neural Information Processing Systems , volume=

  33. [33]

    Transactions on Machine Learning Research , year=

    Causal parrots: Large language models may talk causality but are not causal , author=. Transactions on Machine Learning Research , year=

  34. [34]

    Transactions on Machine Learning Research , year=

    Causal reasoning and large language models: Opening a new frontier for causality , author=. Transactions on Machine Learning Research , year=

  35. [35]

    2024 , eprint=

    Causality for Large Language Models , author=. 2024 , eprint=

    2024

  36. [36]

    2024 , eprint=

    A Critical Review of Causal Reasoning Benchmarks for Large Language Models , author=. 2024 , eprint=

    2024

  37. [37]

    Llms can't plan, but can help planning in llm-modulo frameworks, 2024

    Kambhampati, Subbarao and Valmeekam, Karthik and Guan, Lin and Stechly, Kaya and Verma, Mudit and Bhambri, Siddhant and Saldyt, Lucas and Murthy, Anil , year=. 2402.01817 , archivePrefix=

    work page arXiv

  38. [38]

    Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=

    Probing neural network comprehension of natural language arguments , author=. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=. 2019 , publisher=

    2019

  39. [39]

    2001 , publisher=

    Knowledge in Action: Logical Foundations for Specifying and Implementing Dynamical Systems , author=. 2001 , publisher=

    2001

  40. [40]

    2025 , doi=

    Ban, Taiyu and Chen, Lyuzhou and Lyu, Derui and Wang, Xiangyu and Zhu, Qinrui and Chen, Huanhuan , journal=. 2025 , doi=

    2025

  41. [41]

    Nature Machine Intelligence , volume=

    Shortcut learning in deep neural networks , author=. Nature Machine Intelligence , volume=. 2020 , doi=

    2020

  42. [42]

    Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=

    Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference , author=. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=. 2019 , publisher=

    2019

  43. [43]

    Advances in Neural Information Processing Systems , volume=

    The pitfalls of simplicity bias in neural networks , author=. Advances in Neural Information Processing Systems , volume=

  44. [44]

    Journal of Machine Learning Research , volume=

    Underspecification presents challenges for credibility in modern machine learning , author=. Journal of Machine Learning Research , volume=

  45. [45]

    The Journal of Symbolic Logic , volume=

    On the logic of theory change: Partial meet contraction and revision functions , author=. The Journal of Symbolic Logic , volume=. 1985 , doi=

    1985

  46. [46]

    1999 , publisher=

    A Textbook of Belief Dynamics: Theory Change and Database Updating , author=. 1999 , publisher=

    1999

  47. [47]

    Proceedings of the 37th International Conference on Machine Learning , volume=

    Designing optimal dynamic treatment regimes: A causal reinforcement learning approach , author=. Proceedings of the 37th International Conference on Machine Learning , volume=. 2020 , publisher=

    2020

  48. [48]

    Machine Learning , volume=

    Learning to predict by the methods of temporal differences , author=. Machine Learning , volume=. 1988 , doi=

    1988

  49. [49]

    Watkins, Christopher J. C. H. and Dayan, Peter , journal=. Technical Note:. 1992 , doi=

    1992

  50. [50]

    Biometrika , volume=

    On the likelihood that one unknown probability exceeds another in view of the evidence of two samples , author=. Biometrika , volume=. 1933 , doi=

    1933

  51. [51]

    Advances in Applied Mathematics , volume=

    Asymptotically efficient adaptive allocation rules , author=. Advances in Applied Mathematics , volume=. 1985 , doi=

    1985

  52. [52]

    Journal of the Royal Statistical Society

    Present position and potential developments: Some personal views: Statistical theory: The prequential approach , author=. Journal of the Royal Statistical Society. Series A (General) , volume=. 1984 , doi=

    1984

  53. [53]

    The Twelfth International Conference on Learning Representations , year=

    Can large language models infer causation from correlation? , author=. The Twelfth International Conference on Learning Representations , year=

  54. [54]

    Proceedings of the IEEE , volume=

    Toward causal representation learning , author=. Proceedings of the IEEE , volume=. 2021 , doi=

    2021

  55. [55]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    Exploring defeasibility in causal reasoning , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=. 2024 , publisher=

    2024

  56. [56]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Uncertainty in Causality: A New Frontier , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2025 , publisher=

    2025

  57. [57]

    2025 , publisher=

    Yu, Longxuan and Chen, Delin and Xiong, Siheng and Wu, Qingyang and Li, Dawei and Chen, Zhikai and Liu, Xiaoze and Pan, Liangming , booktitle=. 2025 , publisher=. doi:10.18653/v1/2025.naacl-long.622 , note=

    work page doi:10.18653/v1/2025.naacl-long.622 2025

  58. [58]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining , year=

    Causal discovery through synergizing large language model and data-driven reasoning , author=. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining , year=

  59. [59]

    Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence , pages=

    Large Language Models for Causal Discovery: Current Landscape and Future Directions , author=. Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence , pages=. 2025 , note=

    2025

  60. [60]

    Proceedings of the 42nd International Conference on Machine Learning , series=

    Spurious correlations in high dimensional regression: The roles of regularization, simplicity bias, and over-parameterization , author=. Proceedings of the 42nd International Conference on Machine Learning , series=

  61. [61]

    2025 , eprint=

    Epidemiology of Large Language Models: A Benchmark for Observational Distribution Knowledge , author=. 2025 , eprint=

    2025

  62. [62]

    arXiv preprint arXiv:2510.04721 , year=

    Petrov, Ivo and Dekoninck, Jasper and Vechev, Martin , year=. 2510.04721 , archivePrefix=

    work page arXiv

  63. [63]

    The Twelfth International Conference on Learning Representations , year=

    Towards understanding sycophancy in language models , author=. The Twelfth International Conference on Learning Representations , year=

  64. [64]

    Advances in Neural Information Processing Systems , volume=

    Language models don't always say what they think: Unfaithful explanations in chain-of-thought prompting , author=. Advances in Neural Information Processing Systems , volume=. 2023 , note=

    2023

  65. [65]

    The reversal curse:

    Berglund, Lukas and Tong, Meg and Kaufmann, Max and Balesni, Mikita and Stickland, Asa Cooper and Korbak, Tomasz and Evans, Owain , booktitle=. The reversal curse:. 2024 , note=

    2024

  66. [66]

    2022 , howpublished=

    A Path Towards Autonomous Machine Intelligence , author=. 2022 , howpublished=

    2022

  67. [67]

    Proceedings of the 12th International Conference on Learning Representations (ICLR) , year=

    Large Language Models Cannot Self-Correct Reasoning Yet , author=. Proceedings of the 12th International Conference on Learning Representations (ICLR) , year=

  68. [68]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Reflexion: Language Agents with Verbal Reinforcement Learning , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  69. [69]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Self-Refine: Iterative Refinement with Self-Feedback , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  70. [70]

    , booktitle=

    Zelikman, Eric and Wu, Yuhuai and Mu, Jesse and Goodman, Noah D. , booktitle=

  71. [71]

    Proceedings of the 41st International Conference on Machine Learning (ICML) , year=

    Self-Rewarding Language Models , author=. Proceedings of the 41st International Conference on Machine Learning (ICML) , year=

  72. [72]

    Proceedings of the 11th International Conference on Learning Representations (ICLR) , year=

    Self-Consistency Improves Chain of Thought Reasoning in Language Models , author=. Proceedings of the 11th International Conference on Learning Representations (ICLR) , year=

  73. [73]

    2026 , note=

    A blind auditing protocol for LLM reasoning , author=. 2026 , note=

    2026

  74. [74]

    Training Verifiers to Solve Math Word Problems

    Training Verifiers to Solve Math Word Problems , author=. arXiv preprint arXiv:2110.14168 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  75. [75]

    Proceedings of the 12th International Conference on Learning Representations (ICLR) , year=

    Towards Understanding Sycophancy in Language Models , author=. Proceedings of the 12th International Conference on Learning Representations (ICLR) , year=

  76. [76]

    Inverse Scaling: When Bigger Isn't Better

    Inverse Scaling: When Bigger Isn't Better , author=. arXiv preprint arXiv:2306.09479 , year=

    work page Pith review arXiv

  77. [77]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    TruthfulQA: Measuring How Models Mimic Human Falsehoods , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  78. [78]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  79. [79]

    2000 , publisher=

    Causation, Prediction, and Search , author=. 2000 , publisher=

    2000

  80. [80]

    Position:

    Kambhampati, Subbarao and Valmeekam, Karthik and Guan, Lin and Stechly, Kaya and Verma, Mudit and Bhambri, Siddhant and Saldyt, Lucas and Murthy, Anil , booktitle=. Position:. 2024 , publisher=

    2024

Showing first 80 references.