pith. sign in

arxiv: 2606.18206 · v1 · pith:RNEZN74Unew · submitted 2026-06-16 · 💻 cs.AI

Fixed-Point Reasoners: Stable and Adaptive Deep Looped Transformers

Sajad Movahedi , Vera Milovanovi\'c , Shlomo Libo Feigin , Alexander Theus , Thomas Hofmann , Valentina Boeva , T. Konstantin Rusch , Antonio Orvieto This is my paper

Pith reviewed 2026-06-27 00:35 UTC · model grok-4.3

classification 💻 cs.AI
keywords looped transformersfixed-point convergenceadaptive computationreasoning benchmarkstransformer architecturessudokuarc-agicompositional reasoning
0
0 comments X

The pith

Fixed-point convergence acts as a stable halting mechanism in looped Transformers after pre-norm and residual scaling fix signal issues.

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

The paper proposes FPRM, a looped Transformer that treats the point where successive outputs become identical as the signal to stop computing. This replaces separate learned halting networks and lets the model use more iterations on harder inputs. Pre-norm layers plus residual scaling are introduced to keep signals stable across many loops, addressing the depth-related degradation that otherwise appears when halting is delayed. Experiments on Sudoku, Maze, state-tracking, and ARC-AGI show the model reaches correct solutions while automatically spending more steps on difficult cases.

Core claim

When pre-norm layers and residual scaling are added to a looped Transformer, the fixed point of the iteration becomes a reliable, end-to-end halting criterion that allows the model to adapt its effective depth to task difficulty and solve compositional reasoning problems.

What carries the argument

Fixed-point convergence used as the halting signal inside a looped Transformer equipped with pre-norm layers and residual scaling.

If this is right

  • The architecture can allocate variable compute per example without an auxiliary halting head.
  • Looped depth becomes determined by input content rather than a fixed hyperparameter.
  • The same model can handle both easy and hard instances of Sudoku, Maze, state tracking, and ARC-AGI by iterating until convergence.
  • Training remains end-to-end because the halting decision is a direct property of the forward pass.

Where Pith is reading between the lines

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

  • The approach may extend naturally to other sequence or grid tasks where solution quality improves with additional reasoning steps.
  • If convergence speed correlates with human-perceived difficulty, the iteration count could serve as an interpretable difficulty metric.
  • Removing the fixed-point assumption while keeping pre-norm and scaling might reveal whether the stability benefit is independent of the halting method.

Load-bearing premise

Pre-norm layers and residual scaling are sufficient to keep signals stable in deep loops so that convergence can be trusted as the stopping rule.

What would settle it

Run the model on the reported benchmarks and observe whether it remains stable, whether iteration count increases with problem difficulty, and whether accuracy matches or exceeds non-adaptive baselines.

Figures

Figures reproduced from arXiv: 2606.18206 by Alexander Theus, Antonio Orvieto, Sajad Movahedi, Shlomo Libo Feigin, Thomas Hofmann, T. Konstantin Rusch, Valentina Boeva, Vera Milovanovi\'c.

Figure 1
Figure 1. Figure 1: Signal propagation and adap￾tivity, FPRM vs. TRM: Sudoku-Extreme performance as a function of compute across difficulty. Despite being non￾hierarchical, FPRM scales better, while correctly detecting the accuracy plateaus by using fixed-points for halting. Reasoning in neural networks has increasingly been framed as a problem of scaling test-time compute: a model should be able to spend more computation on … view at source ↗
Figure 2
Figure 2. Figure 2: The blessing and the curse of depth in Looped Transformers. Increasing the number of effective layers can unlock expressivity, but also creates a stability challenge: pre￾norm models without residual scaling can diverge in activation norm, while post-norm models may struggle to utilize the signal. SwiGLU + + Norm + No Stop? Yes Layer Attention Norm ShortConv [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Length generalization and adaptive compute as a function of sequence length. Shaded bands show 95% confidence intervals over seeds. The vertical dotted line marks the training length 32. The matched compute budget is 320 effective layers. State tracking. As shown in [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FPRM achieves (a) better accuracy, while (b) adapting more efficiently to the task difficulty. Difficulty is measured by the number of empty cells in the Sudoku grid. The max. compute budget is matched across models (4788 effective layers). From (b): effective layers are reported as medians with 25th–75th percentiles bands. The default behavior of TRM is without ACT at inference time (in black), which exha… view at source ↗
Figure 8
Figure 8. Figure 8: Decay rate and pa￾tience. Test accuracy and effec￾tive layer of FPRM with fixed￾point halting as a function of decay rate γ, for maximum￾patience P ∈ {5, 10}. 4.4.1 Boundedness of activation norms and trainability As noted in Section 3.1, the normalization scheme governs a trade-off between activation stability and signal propagation, which sharpens with depth. Post-norm keeps activations bounded but suffe… view at source ↗
Figure 9
Figure 9. Figure 9: The distribution of the residual scales in FPRM after training on the Sudoku￾Extreme dataset. Sudoku-Extreme. Compared to the previous experiment on the state-tracking task, here we run each model far beyond its trained depth, trying to detect the point where more compute no longer translates into improvements at test-time. We expect the performance of a model with fewer signal propagation issues to satura… view at source ↗
Figure 10
Figure 10. Figure 10: Landscape visualization for the setup proposed in Section [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Sudoku-Extreme dataset is imbal￾anced. The number of samples per difficulty level (number of empty cells). 101 102 103 104 Inference compute (effective layers) 0 20 40 60 80 100 Test accuracy (%) increasing difficulty 10−2 10−1 100 Residue norm [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: The optimal way to spend the fixed looping compute is to maximize deep supervision steps. The numbers next to markers are inner recurrence depths per each deep supervision step. The total depth of effective layers is approximately the same across all configurations of TRM and FPRM on the Sudoku-Extreme task. G Additional Experimental Details Weight initialization. It seems that initializing the weights us… view at source ↗
read the original abstract

Looped architectures provide an inductive bias toward learning step-by-step procedures for tasks that require compositional reasoning. The number of effective layers reached by looping determines the quality of the solution these models find. Like deep architectures, looped architectures are prone to a signal propagation problem induced by depth as the halting decision is postponed. In this paper, we address this signal propagation issue using pre-norm layers and residual scaling. Building on these architectural modifications, we propose FPRM, a Transformer-based Fixed-Point Reasoning Model that uses fixed-point convergence as an end-to-end halting mechanism in a looped architecture. We show that fixed-point halting allows FPRM to adapt its compute to task difficulty. FPRM is effective on common reasoning benchmarks, namely Sudoku, Maze, state-tracking, and ARC-AGI.

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 / 0 minor

Summary. The manuscript proposes Fixed-Point Reasoning Models (FPRM), a looped Transformer architecture that employs fixed-point convergence as an end-to-end halting mechanism. The authors argue that pre-norm layers and residual scaling address the signal propagation problem in deep looped architectures, enabling the model to adapt its computational effort to task difficulty. They claim effectiveness on reasoning benchmarks including Sudoku, Maze, state-tracking, and ARC-AGI.

Significance. If substantiated with quantitative evidence and analysis, the work could advance adaptive-depth models for compositional reasoning by providing a stability mechanism for looped transformers that uses convergence itself as the halting signal, rather than learned or fixed-depth alternatives.

major comments (2)
  1. Abstract: the claim of effectiveness on benchmarks is asserted after describing the architectural changes, but supplies no quantitative results, baselines, error bars, or experimental protocol; support for the central claim cannot be evaluated.
  2. Architecture section (description of FPRM and modifications): no convergence analysis, derivation, ablation studies, or empirical verification is provided showing that pre-norm layers and residual scaling resolve the depth-induced signal propagation problem sufficiently for fixed-point convergence to act as a reliable, stable halting criterion without divergence, vanishing gradients, or non-convergence. This assumption is load-bearing for attributing adaptive compute and benchmark performance to the fixed-point mechanism.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and agree that revisions are needed to strengthen the manuscript.

read point-by-point responses

  1. Referee: Abstract: the claim of effectiveness on benchmarks is asserted after describing the architectural changes, but supplies no quantitative results, baselines, error bars, or experimental protocol; support for the central claim cannot be evaluated.

    Authors: We agree that the abstract lacks supporting quantitative evidence. In the revised manuscript, we will incorporate key performance metrics (e.g., accuracy on Sudoku, Maze, state-tracking, and ARC-AGI), baseline comparisons, and error bars to substantiate the effectiveness claims. revision: yes

  2. Referee: Architecture section (description of FPRM and modifications): no convergence analysis, derivation, ablation studies, or empirical verification is provided showing that pre-norm layers and residual scaling resolve the depth-induced signal propagation problem sufficiently for fixed-point convergence to act as a reliable, stable halting criterion without divergence, vanishing gradients, or non-convergence. This assumption is load-bearing for attributing adaptive compute and benchmark performance to the fixed-point mechanism.

    Authors: We acknowledge the absence of explicit convergence analysis, derivations, or targeted ablations in the current manuscript. The work relies on overall benchmark results to indicate stability. We will add a dedicated subsection with theoretical motivation for the modifications, a derivation of residual scaling, ablation studies isolating their impact on convergence, and empirical checks (e.g., gradient norms and iteration counts) to verify fixed-point behavior. revision: yes

Circularity Check

0 steps flagged

No circularity: architectural proposal rests on empirical validation, not self-referential derivation.

full rationale

The paper proposes pre-norm layers and residual scaling to stabilize looped transformers, then uses fixed-point convergence as a halting mechanism. No equations, derivations, or fitted parameters are presented that reduce by construction to the inputs. The central claims are supported by benchmark results on Sudoku, Maze, state-tracking, and ARC-AGI rather than any mathematical identity or self-citation chain. This is the common case of an empirical architecture paper with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Abstract introduces no explicit free parameters, mathematical axioms, or additional invented entities beyond naming the FPRM model itself.

invented entities (1)
  • FPRM no independent evidence
    purpose: Transformer-based Fixed-Point Reasoning Model that uses fixed-point convergence for halting
    Named and described as the central proposed architecture in the abstract.

pith-pipeline@v0.9.1-grok · 5690 in / 1170 out tokens · 50085 ms · 2026-06-27T00:35:49.571323+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

87 extracted references · 32 canonical work pages · 9 internal anchors

  1. [1]

    2026 , eprint=

    Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning , author=. 2026 , eprint=

    2026

  2. [2]

    Hierarchical Reasoning Model

    Guan Wang and Jin Li and Yuhao Sun and Xing Chen and Changling Liu and Yue Wu and Meng Lu and Sen Song and Yasin Abbasi. Hierarchical Reasoning Model , journal =. 2025 , url =. doi:10.48550/ARXIV.2506.21734 , eprinttype =. 2506.21734 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2506.21734 2025

  3. [3]

    CoRR , volume =

    Vardhan Palod and Karthik Valmeekam and Kaya Stechly and Subbarao Kambhampati , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2509.07339 , eprinttype =. 2509.07339 , timestamp =

    work page doi:10.48550/arxiv.2509.07339 2025

  4. [4]

    CoRR , volume =

    William Merrill and Ashish Sabharwal , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2503.03961 , eprinttype =. 2503.03961 , timestamp =

    work page doi:10.48550/arxiv.2503.03961 2025

  5. [5]

    2026 , eprint=

    Parcae: Scaling Laws For Stable Looped Language Models , author=. 2026 , eprint=

    2026

  6. [6]

    2022 , url =

    Learning Iterative Reasoning through Energy Minimization , booktitle =. 2022 , url =

    2022

  7. [7]

    Tenenbaum , editor =

    Yilun Du and Jiayuan Mao and Joshua B. Tenenbaum , editor =. Learning Iterative Reasoning through Energy Diffusion , booktitle =. 2024 , url =

    2024

  8. [8]

    Nowak and Dimitris Papailiopoulos , title =

    Liu Yang and Kangwook Lee and Robert D. Nowak and Dimitris Papailiopoulos , title =. The Twelfth International Conference on Learning Representations,. 2024 , url =

    2024

  9. [9]

    Training Large Language Models to Reason in a Continuous Latent Space

    Shibo Hao and Sainbayar Sukhbaatar and DiJia Su and Xian Li and Zhiting Hu and Jason Weston and Yuandong Tian , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2412.06769 , eprinttype =. 2412.06769 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.06769 2024

  10. [10]

    International Conference on Learning Representations , volume=

    Looped transformers for length generalization , author=. International Conference on Learning Representations , volume=

  11. [11]

    2024 , eprint=

    The Expressive Power of Transformers with Chain of Thought , author=. 2024 , eprint=

    2024

  12. [12]

    CoRR , volume =

    Hanlin Zhu and Shibo Hao and Zhiting Hu and Jiantao Jiao and Stuart Russell and Yuandong Tian , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2505.12514 , eprinttype =. 2505.12514 , timestamp =

    work page doi:10.48550/arxiv.2505.12514 2025

  13. [13]

    Stability and Generalization in Looped Transformers

    Asher Labovich , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2604.15259 , eprinttype =. 2604.15259 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.15259 2026

  14. [14]

    2016 , eprint=

    Exponential expressivity in deep neural networks through transient chaos , author=. 2016 , eprint=

    2016

  15. [15]

    CoRR , volume =

    Zirui Ren and Ziming Liu , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2601.10679 , eprinttype =. 2601.10679 , timestamp =

    work page doi:10.48550/arxiv.2601.10679 2026

  16. [16]

    2026 , eprint=

    LoopFormer: Elastic-Depth Looped Transformers for Latent Reasoning via Shortcut Modulation , author=. 2026 , eprint=

    2026

  17. [17]

    CoRR , volume =

    Ferdinand Kapl and Emmanouil Angelis and Kaitlin Maile and Johannes von Oswald and Stefan Bauer , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2602.16490 , eprinttype =. 2602.16490 , timestamp =

    work page doi:10.48550/arxiv.2602.16490 2026

  18. [18]

    CoRR , volume =

    Sajad Movahedi and Felix Sarnthein and Nicola Muca Cirone and Antonio Orvieto , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2503.10799 , eprinttype =. 2503.10799 , timestamp =

    work page doi:10.48550/arxiv.2503.10799 2025

  19. [19]

    CoRR , volume =

    Fran. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2412.04604 , eprinttype =. 2412.04604 , timestamp =

    work page doi:10.48550/arxiv.2412.04604 2024

  20. [20]

    7th International Conference on Learning Representations,

    Mostafa Dehghani and Stephan Gouws and Oriol Vinyals and Jakob Uszkoreit and Lukasz Kaiser , title =. 7th International Conference on Learning Representations,. 2019 , url =

    2019

  21. [21]

    Advances in Neural Information Processing Systems , volume=

    Scaling up test-time compute with latent reasoning: A recurrent depth approach , author=. Advances in Neural Information Processing Systems , volume=

  22. [22]

    Less is More: Recursive Reasoning with Tiny Networks

    Alexia Jolicoeur. Less is More: Recursive Reasoning with Tiny Networks , journal =. 2025 , url =. doi:10.48550/ARXIV.2510.04871 , eprinttype =. 2510.04871 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2510.04871 2025

  23. [23]

    The Illusion of State in State-Space Models , booktitle =

    William Merrill and Jackson Petty and Ashish Sabharwal , editor =. The Illusion of State in State-Space Models , booktitle =. 2024 , url =

    2024

  24. [24]

    2022 , eprint=

    Saturated Transformers are Constant-Depth Threshold Circuits , author=. 2022 , eprint=

    2022

  25. [25]

    2021 , eprint=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. 2021 , eprint=

    2021

  26. [26]

    2023 , eprint=

    Faith and Fate: Limits of Transformers on Compositionality , author=. 2023 , eprint=

    2023

  27. [27]

    Reddi , title =

    Nikunj Saunshi and Nishanth Dikkala and Zhiyuan Li and Sanjiv Kumar and Sashank J. Reddi , title =. The Thirteenth International Conference on Learning Representations,. 2025 , url =

    2025

  28. [28]

    CoRR , volume =

    Alex Graves , title =. CoRR , volume =. 2016 , url =. 1603.08983 , timestamp =

    Pith/arXiv arXiv 2016

  29. [29]

    LoopViT: Scaling Visual

    Wen. LoopViT: Scaling Visual. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2602.02156 , eprinttype =. 2602.02156 , timestamp =

    work page doi:10.48550/arxiv.2602.02156 2026

  30. [30]

    Gomez and Lukasz Kaiser and Illia Polosukhin , editor =

    Ashish Vaswani and Noam Shazeer and Niki Parmar and Jakob Uszkoreit and Llion Jones and Aidan N. Gomez and Lukasz Kaiser and Illia Polosukhin , editor =. Attention is All you Need , booktitle =. 2017 , url =

    2017

  31. [31]

    Neural GPUs Learn Algorithms , booktitle =

    Lukasz Kaiser and Ilya Sutskever , editor =. Neural GPUs Learn Algorithms , booktitle =. 2016 , url =

    2016

  32. [32]

    CoRR , volume =

    Andrea Banino and Jan Balaguer and Charles Blundell , title =. CoRR , volume =. 2021 , url =. 2107.05407 , timestamp =

    arXiv 2021

  33. [33]

    2022 , eprint=

    DeepNet: Scaling Transformers to 1,000 Layers , author=. 2022 , eprint=

    2022

  34. [34]

    The Lipschitz Constant of Self-Attention , booktitle =

    Hyunjik Kim and George Papamakarios and Andriy Mnih , editor =. The Lipschitz Constant of Self-Attention , booktitle =. 2021 , url =

    2021

  35. [35]

    Tom B. Brown and Benjamin Mann and Nick Ryder and Melanie Subbiah and Jared Kaplan and Prafulla Dhariwal and Arvind Neelakantan and Pranav Shyam and Girish Sastry and Amanda Askell and Sandhini Agarwal and Ariel Herbert. Language Models are Few-Shot Learners , booktitle =. 2020 , url =

    2020

  36. [36]

    2024 , month = sep, howpublished =

    Learning to Reason with. 2024 , month = sep, howpublished =

    2024

  37. [37]

    Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

    Charlie Snell and Jaehoon Lee and Kelvin Xu and Aviral Kumar , title =. CoRR , volume =. 2024 , url =. doi:10.48550/ARXIV.2408.03314 , eprinttype =. 2408.03314 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2408.03314 2024

  38. [38]

    Chi and Quoc V

    Jason Wei and Xuezhi Wang and Dale Schuurmans and Maarten Bosma and Brian Ichter and Fei Xia and Ed H. Chi and Quoc V. Le and Denny Zhou , editor =. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , booktitle =. 2022 , url =

    2022

  39. [39]

    Daya Guo and Dejian Yang and Haowei Zhang and others , title =. Nat. , volume =. 2025 , url =. doi:10.1038/S41586-025-09422-Z , timestamp =

    work page doi:10.1038/s41586-025-09422-z 2025

  40. [40]

    Can You Learn an Algorithm? Generalizing from Easy to Hard Problems with Recurrent Networks , booktitle =

    Avi Schwarzschild and Eitan Borgnia and Arjun Gupta and Furong Huang and Uzi Vishkin and Micah Goldblum and Tom Goldstein , editor =. Can You Learn an Algorithm? Generalizing from Easy to Hard Problems with Recurrent Networks , booktitle =. 2021 , url =

    2021

  41. [41]

    Attention is not all you need: pure attention loses rank doubly exponentially with depth , booktitle =

    Yihe Dong and Jean. Attention is not all you need: pure attention loses rank doubly exponentially with depth , booktitle =. 2021 , url =

    2021

  42. [42]

    Simard and Paolo Frasconi , title =

    Yoshua Bengio and Patrice Y. Simard and Paolo Frasconi , title =. 1994 , url =. doi:10.1109/72.279181 , timestamp =

    work page doi:10.1109/72.279181 1994

  43. [43]

    On the difficulty of training recurrent neural networks , booktitle =

    Razvan Pascanu and Tom. On the difficulty of training recurrent neural networks , booktitle =. 2013 , url =

    2013

  44. [44]

    On Layer Normalization in the Transformer Architecture , booktitle =

    Ruibin Xiong and Yunchang Yang and Di He and Kai Zheng and Shuxin Zheng and Chen Xing and Huishuai Zhang and Yanyan Lan and Liwei Wang and Tie. On Layer Normalization in the Transformer Architecture , booktitle =. 2020 , url =

    2020

  45. [45]

    CoRR , volume =

    Alex Graves and Greg Wayne and Ivo Danihelka , title =. CoRR , volume =. 2014 , url =. 1410.5401 , timestamp =

    Pith/arXiv arXiv 2014

  46. [46]

    End-to-end Algorithm Synthesis with Recurrent Networks: Extrapolation without Overthinking , booktitle =

    Arpit Bansal and Avi Schwarzschild and Eitan Borgnia and Zeyad Emam and Furong Huang and Micah Goldblum and Tom Goldstein , editor =. End-to-end Algorithm Synthesis with Recurrent Networks: Extrapolation without Overthinking , booktitle =. 2022 , url =

    2022

  47. [47]

    Signal Propagation in Transformers: Theoretical Perspectives and the Role of Rank Collapse , booktitle =

    Lorenzo Noci and Sotiris Anagnostidis and Luca Biggio and Antonio Orvieto and Sidak Pal Singh and Aur. Signal Propagation in Transformers: Theoretical Perspectives and the Role of Rank Collapse , booktitle =. 2022 , url =

    2022

  48. [48]

    CoRR , volume =

    Wenfang Sun and Xinyuan Song and Pengxiang Li and Lu Yin and Yefeng Zheng and Shiwei Liu , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2502.05795 , eprinttype =. 2502.05795 , timestamp =

    work page doi:10.48550/arxiv.2502.05795 2025

  49. [49]

    Zico Kolter and Vladlen Koltun , editor =

    Shaojie Bai and J. Zico Kolter and Vladlen Koltun , editor =. Deep Equilibrium Models , booktitle =. 2019 , url =

    2019

  50. [50]

    Mathematics of Computation , year=

    A Class of Methods for Solving Nonlinear Simultaneous Equations , author=. Mathematics of Computation , year=

  51. [51]

    2026 , eprint=

    PonderLM: Pretraining Language Models to Ponder in Continuous Space , author=. 2026 , eprint=

    2026

  52. [52]

    Advances in neural information processing systems , volume=

    Implicit generation and modeling with energy based models , author=. Advances in neural information processing systems , volume=

  53. [53]

    International conference on machine learning , pages=

    Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  54. [54]

    2026 , url=

    Tao Zhang and Jia-Shu Pan and Ruiqi Feng and Tailin Wu , booktitle=. 2026 , url=

    2026

  55. [55]

    CoRR , volume =

    Alexi Gladstone and Ganesh Nanduru and Md Mofijul Islam and Peixuan Han and Hyeonjeong Ha and Aman Chadha and Yilun Du and Heng Ji and Jundong Li and Tariq Iqbal , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2507.02092 , eprinttype =. 2507.02092 , timestamp =

    work page doi:10.48550/arxiv.2507.02092 2025

  56. [56]

    2021 , note=

    Should EBMs Model the Energy or the Score? , author=. 2021 , note=

    2021

  57. [57]

    Structured Prediction Energy Networks , booktitle =

    David Belanger and Andrew McCallum , editor =. Structured Prediction Energy Networks , booktitle =. 2016 , url =

    2016

  58. [58]

    End-to-End Learning for Structured Prediction Energy Networks , booktitle =

    David Belanger and Bishan Yang and Andrew McCallum , editor =. End-to-End Learning for Structured Prediction Energy Networks , booktitle =. 2017 , url =

    2017

  59. [59]

    Predicting structured data , volume=

    A tutorial on energy-based learning , author=. Predicting structured data , volume=

  60. [60]

    Score-Based Generative Modeling through Stochastic Differential Equations , booktitle =

    Yang Song and Jascha Sohl. Score-Based Generative Modeling through Stochastic Differential Equations , booktitle =. 2021 , url =

    2021

  61. [61]

    Denoising Diffusion Probabilistic Models , booktitle =

    Jonathan Ho and Ajay Jain and Pieter Abbeel , editor =. Denoising Diffusion Probabilistic Models , booktitle =. 2020 , url =

    2020

  62. [62]

    Poggio , title =

    Renee Ge and Qianli Liao and Tomaso A. Poggio , title =. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2510.00355 , eprinttype =. 2510.00355 , timestamp =

    work page doi:10.48550/arxiv.2510.00355 2025

  63. [63]

    CoRR , volume =

    Shixiang Song and He Li and Zitong Wang and Boyi Zeng and Feichen Song and Yixuan Wang and Zhiqin John Xu and Ziwei He and Zhouhan Lin , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2603.01914 , eprinttype =. 2603.01914 , timestamp =

    work page doi:10.48550/arxiv.2603.01914 2026

  64. [64]

    Shaojie Bai, J

    Sangmin Bae and Yujin Kim and Reza Bayat and Sungnyun Kim and Jiyoun Ha and Tal Schuster and Adam Fisch and Hrayr Harutyunyan and Ziwei Ji and Aaron C. Courville and Se. Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation , journal =. 2025 , url =. doi:10.48550/ARXIV.2507.10524 , eprinttype =. 2507.10524 , timestamp =

    work page doi:10.48550/arxiv.2507.10524 2025

  65. [65]

    Solve the Loop: Attractor Models for Language and Reasoning

    Jacob Fein. Solve the Loop: Attractor Models for Language and Reasoning , journal =. 2026 , url =. doi:10.48550/ARXIV.2605.12466 , eprinttype =. 2605.12466 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2605.12466 2026

  66. [66]

    Zico Kolter , title =

    Zhengyang Geng and J. Zico Kolter , title =. CoRR , volume =. 2023 , url =. doi:10.48550/ARXIV.2310.18605 , eprinttype =. 2310.18605 , timestamp =

    work page doi:10.48550/arxiv.2310.18605 2023

  67. [67]

    Donald G. M. Anderson , title =. J. 1965 , url =. doi:10.1145/321296.321305 , timestamp =

    work page doi:10.1145/321296.321305 1965

  68. [68]

    Marcelo O. R. Prates and Lu. Problem Solving at the Edge of Chaos: Entropy, Puzzles and the Sudoku Freezing Transition , booktitle =. 2018 , url =. doi:10.1109/ICTAI.2018.00109 , timestamp =

    work page doi:10.1109/ictai.2018.00109 2018

  69. [69]

    Zico Kolter and Roger B

    Cem Anil and Ashwini Pokle and Kaiqu Liang and Johannes Treutlein and Yuhuai Wu and Shaojie Bai and J. Zico Kolter and Roger B. Grosse , editor =. Path Independent Equilibrium Models Can Better Exploit Test-Time Computation , booktitle =. 2022 , url =

    2022

  70. [70]

    On Training Implicit Models , booktitle =

    Zhengyang Geng and Xin. On Training Implicit Models , booktitle =. 2021 , url =

    2021

  71. [71]

    Workshop on Latent

    Recursive Reasoning as Attractor Landscape Search: Mechanistic Dynamics of the Tiny Recursive Model , author=. Workshop on Latent. 2026 , url=

    2026

  72. [72]

    Zico Kolter , editor =

    Shaojie Bai and Vladlen Koltun and J. Zico Kolter , editor =. Stabilizing Equilibrium Models by Jacobian Regularization , booktitle =. 2021 , url =

    2021

  73. [73]

    Looped Transformers as Programmable Computers , booktitle =

    Angeliki Giannou and Shashank Rajput and Jy. Looped Transformers as Programmable Computers , booktitle =. 2023 , url =

    2023

  74. [74]

    Loop, Think, & Generalize: Implicit Reasoning in Recurrent-Depth Transformers

    Harsh Kohli and Srinivasan Parthasarathy and Huan Sun and Yuekun Yao , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2604.07822 , eprinttype =. 2604.07822 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.07822 2026

  75. [75]

    Hugh Blayney and Alvaro Arroyo and Johan S. Obando. A Mechanistic Analysis of Looped Reasoning Language Models , journal =. 2026 , url =. doi:10.48550/ARXIV.2604.11791 , eprinttype =. 2604.11791 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.11791 2026

  76. [76]

    Osher and Wotao Yin , title =

    Samy Wu Fung and Howard Heaton and Qiuwei Li and Daniel McKenzie and Stanley J. Osher and Wotao Yin , title =. Thirty-Sixth. 2022 , url =. doi:10.1609/AAAI.V36I6.20619 , timestamp =

    work page doi:10.1609/aaai.v36i6.20619 2022

  77. [77]

    Peri-LN: Revisiting Normalization Layer in the Transformer Architecture , booktitle =

    Jeonghoon Kim and Byeongchan Lee and Cheonbok Park and Yeontaek Oh and Beomjun Kim and Taehwan Yoo and Seongjin Shin and Dongyoon Han and Jinwoo Shin and Kang Min Yoo , editor =. Peri-LN: Revisiting Normalization Layer in the Transformer Architecture , booktitle =. 2025 , url =

    2025

  78. [78]

    Smith and Albert Gu and Anushan Fernando and

    Antonio Orvieto and Samuel L. Smith and Albert Gu and Anushan Fernando and. Resurrecting Recurrent Neural Networks for Long Sequences , booktitle =. 2023 , url =

    2023

  79. [79]

    Divya Jyoti Bajpai and Manjesh Kumar Hanawal

    Ahmadreza Jeddi and Marco Ciccone and Babak Taati , title =. CoRR , volume =. 2026 , url =. doi:10.48550/ARXIV.2602.11451 , eprinttype =. 2602.11451 , timestamp =

    work page doi:10.48550/arxiv.2602.11451 2026

  80. [80]

    Scaling Latent Reasoning via Looped Language Models

    Rui. Scaling Latent Reasoning via Looped Language Models , journal =. 2025 , url =. doi:10.48550/ARXIV.2510.25741 , eprinttype =. 2510.25741 , timestamp =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2510.25741 2025

Showing first 80 references.