pith. sign in

arxiv: 2606.12689 · v1 · pith:U2WG6CSTnew · submitted 2026-06-10 · 💻 cs.CL

Observable Patterns Are Not Explanations: A Causal-Geometric Analysis of Latent Reasoning Models

Darpan Aswal , Thomas Palmeira Ferraz , Yongxin Zhou , Maxime Peyrard This is my paper

Pith reviewed 2026-06-27 09:34 UTC · model grok-4.3

classification 💻 cs.CL
keywords latent reasoning modelsinterpretabilitycausal interventiongeometric analysischain of thoughtcontrol modelscoconutcodi
0
0 comments X

The pith

Observable patterns in latent states of reasoning models appear in controls and do not establish causal mechanisms.

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

The paper evaluates two latent reasoning models against controls that lack recurrence or curriculum training. It shows that patterns such as BFS-like frontiers and decodable arithmetic also occur in the controls and do not always change model behavior when altered. Causal interventions indicate that any effect of latent thoughts on outputs is graded rather than all-or-nothing, and this effect concentrates in low-rank geometric directions that become more structured as influence grows. A sympathetic reader would care because the result reframes how internal states should be interpreted in language models.

Core claim

Evaluating two LRMs against controls lacking the proposed recurrence or curriculum, we find these patterns also appear in the controls and do not always causally affect behavior. Causal interventions reveal that latent-thought utilization is not binary but graded, scaling with a thought's causal effect on model behavior. Geometric analyses reveal this effect concentrates in low-rank directions whose step-to-step geometry grows more structured as their behavioral influence increases. Latent thoughts should therefore be treated as hidden computation, not hidden explanation: decodability, attention, or static structure alone cannot establish mechanism.

What carries the argument

Matched control models combined with causal interventions that measure the behavioral effect of individual latent thoughts, followed by geometric analysis of their step-to-step structure.

If this is right

  • Decodability of latent states alone does not establish an internal reasoning mechanism.
  • Static geometric structure in latent thoughts does not by itself explain model outputs.
  • Interpretability claims for LRMs require explicit comparison to matched controls.
  • Causal tests are needed to determine whether any given latent thought affects behavior.

Where Pith is reading between the lines

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

  • The same logic may apply to interpretability claims based on attention patterns or activations in other transformer variants.
  • Extending the graded-causality measurement to new tasks could test whether low-rank concentration is a general property of useful latent computation.
  • Designing training regimes that explicitly control the rank and structure of latent directions could become a practical way to improve interpretability.

Load-bearing premise

The control models without recurrence or curriculum are otherwise comparable to the LRMs in capacity and training regime.

What would settle it

Demonstrating that the reported patterns are absent from controls or that every causal intervention on latent thoughts changes model behavior would falsify the central claim.

Figures

Figures reproduced from arXiv: 2606.12689 by Darpan Aswal, Maxime Peyrard, Thomas Palmeira Ferraz, Yongxin Zhou.

Figure 1
Figure 1. Figure 1: Probing for breadth-first-search (BFS) patterns on graph-hopping (top) and scratchpad-thinking on [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Per-layer IEKL on the residual stream under partner-prompt corruption across buckets {Pfull, Pmax, Pb, T1, . . . , TK, Tfull, Ab}. Prompt positions recover best across models and tasks. Thought posi￾tions Tt contribute near zero on graph-hopping; on arithmetic-reasoning, Tfull yields the strongest recovery for C, Cu, and CODI. Takeaway: Latent-thought influence is task-varying—when large, it can override c… view at source ↗
Figure 3
Figure 3. Figure 3: Gradient-subspace intervention flip rates (%) across amplification strength [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Geometric stability of the gradient-subspaces. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Attention mass distribution of the boundary token across prompt tokens, latent thoughts, and itself, [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: A logit-lens projection example for all models on both tasks (graph-hopping and arithmetic-reasoning) [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Per-layer IEKL under partner-prompt corruption across buckets {Pfull, Pmax, Pb, T1, . . . , TK, Tfull, Ab}, decomposed into attention (a) and MLP (b) outputs. Unlike the full residual stream ( [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: 2D PCA projection of the latent space struc [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Targeted interventions isolating and ablating specific variance components of the latent thoughts. [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
read the original abstract

Latent reasoning models (LRMs) replace explicit chain-of-thought with continuous thoughts. Recent work treats observable latent-state patterns, such as BFS-like frontiers and decodable arithmetic computation, as evidence for internal reasoning mechanisms. Evaluating two LRMs (Coconut and CODI) against controls lacking the proposed recurrence or curriculum, we find these patterns also appear in the controls and do not always causally affect behavior. Causal interventions reveal that latent-thought utilization is not binary but graded, scaling with a thought's causal effect on model behavior. Geometric analyses reveal this effect concentrates in low-rank directions whose step-to-step geometry grows more structured as their behavioral influence increases. Latent thoughts should therefore be treated as hidden computation, not hidden explanation: decodability, attention, or static structure alone cannot establish mechanism. LRM interpretability thus requires matched controls and causal tests.

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

1 major / 0 minor

Summary. The manuscript claims that observable patterns in latent reasoning models (LRMs) such as Coconut and CODI—including BFS-like frontiers and decodable arithmetic computation—are not evidence of internal reasoning mechanisms. These patterns also appear in control models lacking recurrence or curriculum and do not consistently exert causal effects on behavior. Causal interventions demonstrate graded rather than binary utilization of latent thoughts, with behavioral influence scaling continuously; geometric analysis shows this influence concentrates in low-rank directions whose step-to-step structure increases with causal strength. The paper concludes that decodability, attention, or static structure alone cannot establish mechanism and that LRM interpretability requires matched controls plus causal tests, treating latent thoughts as hidden computation rather than hidden explanations.

Significance. If the results hold, the work is significant for shifting LRM interpretability away from purely observational analyses toward causal and geometric methods. The explicit use of matched controls, causal interventions, and low-rank geometric quantification supplies a concrete, falsifiable framework that could raise standards across neural interpretability research. Credit is due for grounding claims in intervention outcomes rather than fitted patterns and for the graded-effect finding, which directly challenges binary mechanism assumptions.

major comments (1)
  1. [Abstract] Abstract: the central claim—that patterns in controls undermine their mechanistic status inside LRMs—requires that controls match LRMs on capacity, depth, parameter count, training data, and optimization. The abstract states only that controls 'lack the proposed recurrence or curriculum'; without explicit confirmation of these other axes, the presence of identical patterns in controls does not falsify mechanism inside the LRMs. This matching premise is load-bearing for the graded-effect and geometry conclusions.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive review. We address the single major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim—that patterns in controls undermine their mechanistic status inside LRMs—requires that controls match LRMs on capacity, depth, parameter count, training data, and optimization. The abstract states only that controls 'lack the proposed recurrence or curriculum'; without explicit confirmation of these other axes, the presence of identical patterns in controls does not falsify mechanism inside the LRMs. This matching premise is load-bearing for the graded-effect and geometry conclusions.

    Authors: We agree the abstract is concise and does not enumerate every matching axis. The full manuscript (Methods and experimental setup) details that controls were constructed to match LRMs on capacity, depth, parameter count, training data, and optimization procedure; the sole controlled differences are the absence of recurrence and curriculum. This design isolates the contribution of the proposed mechanisms. To eliminate ambiguity we will revise the abstract to state explicitly that controls are matched on the listed axes in addition to lacking recurrence or curriculum. The graded-effect and geometry results rest on these matched controls as already described in the body of the paper. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on external controls and interventions

full rationale

The paper's argument proceeds by comparing observable patterns in LRMs against independently constructed control models (lacking recurrence/curriculum) and measuring causal effects via interventions on behavior. These steps are not self-definitional, as the controls are described as external baselines rather than derived from the LRMs' own definitions, and no equations or claims reduce a prediction to a fitted input by construction. No self-citations are invoked as load-bearing uniqueness theorems, and the abstract provides no evidence of ansatz smuggling or renaming. The control-comparability premise is a methodological assumption open to empirical challenge, not a definitional loop internal to the derivation chain.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no information on fitted parameters, background axioms, or new postulated entities; the work is presented as an empirical comparative study.

pith-pipeline@v0.9.1-grok · 5686 in / 1100 out tokens · 25090 ms · 2026-06-27T09:34:22.262508+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

67 extracted references · 9 canonical work pages · 1 internal anchor

  1. [1]

    2025 , url=

    Shibo Hao and Sainbayar Sukhbaatar and DiJia Su and Xian Li and Zhiting Hu and Jason E Weston and Yuandong Tian , booktitle=. 2025 , url=

    2025

  2. [2]

    Probing the Probing Paradigm: Does Probing Accuracy Entail Task Relevance?

    Ravichander, Abhilasha and Belinkov, Yonatan and Hovy, Eduard , editor =. Probing the Probing Paradigm: Does Probing Accuracy Entail Task Relevance?. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume. 2021. doi:10.18653/v1/2021.eacl-main.295

    work page doi:10.18653/v1/2021.eacl-main.295 2021

  3. [3]

    and Liang, P

    Hewitt, John and Liang, Percy. Designing and Interpreting Probes with Control Tasks. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019. doi:10.18653/v1/D19-1275

    work page doi:10.18653/v1/d19-1275 2019

  4. [4]

    Forty-second International Conference on Machine Learning , year=

    Are Sparse Autoencoders Useful? A Case Study in Sparse Probing , author=. Forty-second International Conference on Machine Learning , year=

  5. [5]

    Computer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXIII , pages =

    Teney, Damien and Peyrard, Maxime and Abbasnejad, Ehsan , title =. Computer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XXIII , pages =. 2022 , isbn =. doi:10.1007/978-3-031-20050-2_27 , abstract =

    work page doi:10.1007/978-3-031-20050-2_27 2022

  6. [7]

    Probing for the Usage of Grammatical Number

    Lasri, Karim and Pimentel, Tiago and Lenci, Alessandro and Poibeau, Thierry and Cotterell, Ryan. Probing for the Usage of Grammatical Number. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022. doi:10.18653/v1/2022.acl-long.603

    work page doi:10.18653/v1/2022.acl-long.603 2022

  7. [8]

    arXiv preprint arXiv:2110.14168 , year=

    Training verifiers to solve math word problems , author=. arXiv preprint arXiv:2110.14168 , year=

    Pith/arXiv arXiv

  8. [9]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Codi: Compressing chain-of-thought into continuous space via self-distillation , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  9. [10]

    The Twelfth International Conference on Learning Representations , year=

    Think before you speak: Training Language Models With Pause Tokens , author=. The Twelfth International Conference on Learning Representations , year=

  10. [11]

    2019 , url=

    Radford, Alec and Wu, Jeff and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya , journal=. 2019 , url=

    2019

  11. [12]

    Amnesic probing: Behavioral explanation with amnesic counterfactuals.Transactions of the Association for Computational Linguistics, 9:160–175, 2021

    Elazar, Yanai and Ravfogel, Shauli and Jacovi, Alon and Goldberg, Yoav. Transactions of the Association for Computational Linguistics. 2021. doi:10.1162/tacl_a_00359

    work page doi:10.1162/tacl_a_00359 2021

  12. [13]

    Neuroimage , volume=

    Neural correlates of interspecies perspective taking in the post-mortem Atlantic Salmon: an argument for multiple comparisons correction , author=. Neuroimage , volume=. 2009 , publisher=

    2009

  13. [14]

    arXiv preprint arXiv:2512.18792 , year=

    The Dead Salmons of AI Interpretability , author=. arXiv preprint arXiv:2512.18792 , year=

    arXiv

  14. [15]

    International conference on machine learning , pages=

    Similarity of neural network representations revisited , author=. International conference on machine learning , pages=. 2019 , organization=

    2019

  15. [16]

    2023 , eprint=

    Implicit Chain of Thought Reasoning via Knowledge Distillation , author=. 2023 , eprint=

    2023

  16. [17]

    2024 , eprint=

    From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step , author=. 2024 , eprint=

    2024

  17. [18]

    Proceedings of the 40th International Conference on Machine Learning , pages =

    Looped Transformers as Programmable Computers , author =. Proceedings of the 40th International Conference on Machine Learning , pages =. 2023 , editor =

    2023

  18. [19]

    International Conference on Learning Representations , year=

    Universal Transformers , author=. International Conference on Learning Representations , year=

  19. [20]

    2025 , url=

    Cywinski, Bartosz and Bussmann, Bart and Conmy, Arthur and Engels, Joshua and Nanda, Neel and Rajamanoharan, Senthooran , journal=. 2025 , url=

    2025

  20. [21]

    Mechanistic Interpretability Workshop at NeurIPS 2025 , year=

    Brad Peters and Sayam Goyal and Mar. Mechanistic Interpretability Workshop at NeurIPS 2025 , year=

    2025

  21. [22]

    2026 , url=

    Yingqian Cui and Zhenwei Dai and Bing He and Zhan Shi and Hui Liu and Rui Sun and Zhiji Liu and Yue Xing and Jiliang Tang and Benoit Dumoulin , booktitle=. 2026 , url=

    2026

  22. [23]

    2512.21711 , archivePrefix=

    Yuyi Zhang and Boyu Tang and Tianjie Ju and Sufeng Duan and Gongshen Liu , year=. 2512.21711 , archivePrefix=

    arXiv

  23. [24]

    2026 , eprint=

    The Illusion of Superposition? A Principled Analysis of Latent Thinking in Language Models , author=. 2026 , eprint=

    2026

  24. [25]

    Reasoning by Superposition: A Theoretical Perspective on Chain of Continuous Thought , url =

    Zhu, Hanlin and Hao, Shibo and Hu, Zhiting and Jiao, Jiantao and Russell, Stuart J and Tian, Yuandong , booktitle =. Reasoning by Superposition: A Theoretical Perspective on Chain of Continuous Thought , url =

  25. [26]

    2026 , eprint=

    Do Latent-CoT Models Think Step-by-Step? A Mechanistic Study on Sequential Reasoning Tasks , author=. 2026 , eprint=

    2026

  26. [27]

    arXiv preprint arXiv:2504.19678 , year=

    From llm reasoning to autonomous ai agents: A comprehensive review , author=. arXiv preprint arXiv:2504.19678 , year=

    Pith/arXiv arXiv

  27. [28]

    Advances in neural information processing systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

  28. [29]

    arXiv preprint arXiv:2505.15778 , year=

    Soft thinking: Unlocking the reasoning potential of llms in continuous concept space , author=. arXiv preprint arXiv:2505.15778 , year=

    arXiv

  29. [30]

    arXiv preprint arXiv:2505.05410 , year=

    Reasoning models don't always say what they think , author=. arXiv preprint arXiv:2505.05410 , year=

    Pith/arXiv arXiv

  30. [31]

    Proceedings of the 2023 ACM conference on fairness, accountability, and transparency , pages=

    Harms from increasingly agentic algorithmic systems , author=. Proceedings of the 2023 ACM conference on fairness, accountability, and transparency , pages=

    2023

  31. [32]

    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=

  32. [33]

    arXiv preprint arXiv:2404.14082 , year=

    Mechanistic interpretability for AI safety--a review , author=. arXiv preprint arXiv:2404.14082 , year=

    Pith/arXiv arXiv

  33. [34]

    arXiv preprint arXiv:2602.22441 , year=

    How Do Latent Reasoning Methods Perform Under Weak and Strong Supervision? , author=. arXiv preprint arXiv:2602.22441 , year=

    arXiv

  34. [35]

    2026 , url=

    Xilin Wei and Xiaoran Liu and Yuhang Zang and Xiaoyi Dong and Yuhang Cao and Jiaqi Wang and Xipeng Qiu and Dahua Lin , booktitle=. 2026 , url=

    2026

  35. [36]

    2026 , url=

    Connor Dilgren and Sarah Wiegreffe , booktitle=. 2026 , url=

    2026

  36. [37]

    2026 , url=

    Zihan Wang and Yijun Dong and Qi Lei , booktitle=. 2026 , url=

    2026

  37. [38]

    arXiv preprint arXiv:2209.10652 , year=

    Toy models of superposition , author=. arXiv preprint arXiv:2209.10652 , year=

    Pith/arXiv arXiv

  38. [39]

    arXiv preprint arXiv:2604.08764 , year=

    Revisiting Anisotropy in Language Transformers: The Geometry of Learning Dynamics , author=. arXiv preprint arXiv:2604.08764 , year=

    Pith/arXiv arXiv

  39. [40]

    arXiv preprint arXiv:1812.04754 , year=

    Gradient descent happens in a tiny subspace , author=. arXiv preprint arXiv:1812.04754 , year=

    Pith/arXiv arXiv

  40. [41]

    arXiv preprint arXiv:2311.03658 , year=

    The linear representation hypothesis and the geometry of large language models , author=. arXiv preprint arXiv:2311.03658 , year=

    Pith/arXiv arXiv

  41. [42]

    arXiv preprint arXiv:2310.02207 , year=

    Language models represent space and time , author=. arXiv preprint arXiv:2310.02207 , year=

    arXiv

  42. [43]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Dissecting recall of factual associations in auto-regressive language models , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  43. [44]

    arXiv preprint arXiv:2505.11423 , year=

    When thinking fails: The pitfalls of reasoning for instruction-following in llms , author=. arXiv preprint arXiv:2505.11423 , year=

    arXiv

  44. [45]

    Faithful chain-of-thought reasoning , author=. Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  45. [46]

    2507.11473 , archivePrefix=

    Tomek Korbak and Mikita Balesni and Elizabeth Barnes and Yoshua Bengio and Joe Benton and Joseph Bloom and Mark Chen and Alan Cooney and Allan Dafoe and Anca Dragan and Scott Emmons and Owain Evans and David Farhi and Ryan Greenblatt and Dan Hendrycks and Marius Hobbhahn and Evan Hubinger and Geoffrey Irving and Erik Jenner and Daniel Kokotajlo and Victor...

    Pith/arXiv arXiv

  46. [47]

    Journal of Machine Learning Research , year =

    Atticus Geiger and Duligur Ibeling and Amir Zur and Maheep Chaudhary and Sonakshi Chauhan and Jing Huang and Aryaman Arora and Zhengxuan Wu and Noah Goodman and Christopher Potts and Thomas Icard , title =. Journal of Machine Learning Research , year =

  47. [48]

    Belinkov

    Belinkov, Yonatan. Probing Classifiers: Promises, Shortcomings, and Advances. Computational Linguistics. 2022. doi:10.1162/coli_a_00422

    work page internal anchor Pith review doi:10.1162/coli_a_00422 2022

  48. [49]

    2023 , eprint=

    Eliciting Latent Predictions from Transformers with the Tuned Lens , author=. 2023 , eprint=

    2023

  49. [50]

    A ttention is not E xplanation

    Jain, Sarthak and Wallace, Byron C. A ttention is not E xplanation. Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019. doi:10.18653/v1/N19-1357

    work page doi:10.18653/v1/n19-1357 2019

  50. [51]

    Interpretability in the Wild: a Circuit for Indirect Object Identification in

    Kevin Ro Wang and Alexandre Variengien and Arthur Conmy and Buck Shlegeris and Jacob Steinhardt , booktitle=. Interpretability in the Wild: a Circuit for Indirect Object Identification in. 2023 , url=

    2023

  51. [52]

    2022 , journal=

    Causal scrubbing, a method for rigorously testing interpretability hypotheses , author=. 2022 , journal=

    2022

  52. [53]

    2024 , eprint=

    How to use and interpret activation patching , author=. 2024 , eprint=

    2024

  53. [54]

    The Thirteenth International Conference on Learning Representations , year=

    Everything, Everywhere, All at Once: Is Mechanistic Interpretability Identifiable? , author=. The Thirteenth International Conference on Learning Representations , year=

  54. [55]

    arXiv preprint arXiv:2503.08679 , year=

    Chain-of-thought reasoning in the wild is not always faithful , author=. arXiv preprint arXiv:2503.08679 , year=

    Pith/arXiv arXiv

  55. [56]

    arXiv preprint arXiv:2603.05488 , year=

    Reasoning theater: Disentangling model beliefs from chain-of-thought , author=. arXiv preprint arXiv:2603.05488 , year=

    Pith/arXiv arXiv

  56. [57]

    arXiv preprint arXiv:2510.24941 , year=

    Can Aha Moments Be Fake? Identifying True and Decorative Thinking Steps in Chain-of-Thought , author=. arXiv preprint arXiv:2510.24941 , year=

    Pith/arXiv arXiv

  57. [58]

    Advances in neural information processing systems , volume=

    Locating and editing factual associations in gpt , author=. Advances in neural information processing systems , volume=

  58. [59]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Towards interpretable sequence continuation: Analyzing shared circuits in large language models , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  59. [60]

    2023 , eprint=

    The Hydra Effect: Emergent Self-repair in Language Model Computations , author=. 2023 , eprint=

    2023

  60. [61]

    2026 , eprint=

    What Really Controls Temporal Reasoning in Large Language Models: Tokenisation or Representation of Time? , author=. 2026 , eprint=

    2026

  61. [62]

    Date Fragments: A Hidden Bottleneck of Tokenization for Temporal Reasoning

    Bhatia, Gagan and Peyrard, Maxime and Zhao, Wei. Date Fragments: A Hidden Bottleneck of Tokenization for Temporal Reasoning. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.emnlp-main.159

    work page doi:10.18653/v1/2025.emnlp-main.159 2025

  62. [63]

    A Glitch in the Matrix? Locating and Detecting Language Model Grounding with Fakepedia

    Monea, Giovanni and Peyrard, Maxime and Josifoski, Martin and Chaudhary, Vishrav and Eisner, Jason and Kiciman, Emre and Palangi, Hamid and Patra, Barun and West, Robert. A Glitch in the Matrix? Locating and Detecting Language Model Grounding with Fakepedia. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1:...

    work page doi:10.18653/v1/2024.acl-long.369 2024

  63. [64]

    Bogdan and Uzay Macar and Neel Nanda and Arthur Conmy , booktitle=

    Paul C. Bogdan and Uzay Macar and Neel Nanda and Arthur Conmy , booktitle=. 2025 , url=

    2025

  64. [65]

    The Fourteenth International Conference on Learning Representations , year=

    The Geometry of Reasoning: Flowing Logics in Representation Space , author=. The Fourteenth International Conference on Learning Representations , year=

  65. [66]

    2024 , month = sep, url =

    2024

  66. [67]

    2407.21783 , archivePrefix=

    Grattafiori, Aaron and Dubey, Abhimanyu and Jauhri, Abhinav and Pandey, Abhinav and Kadian, Abhishek and Al-Dahle, Ahmad and Letman, Aiesha and Mathur, Akhil and Schelten, Alan and Vaughan, Alex and others , year=. 2407.21783 , archivePrefix=

    Pith/arXiv arXiv

  67. [68]

    Transformer Circuits Thread , volume=

    A mathematical framework for transformer circuits , author=. Transformer Circuits Thread , volume=