pith. sign in

arxiv: 2607.01077 · v1 · pith:VDZFECI5new · submitted 2026-07-01 · 💻 cs.CL · cs.LG

Message Passing Enables Efficient Reasoning

Xuecheng Liu , Daman Arora , Gokul Swamy , Andrea Zanette This is my paper

Pith reviewed 2026-07-02 12:42 UTC · model grok-4.3

classification 💻 cs.CL cs.LG
keywords message passinglanguage modelsreasoningsudokuchain of thoughtfork joinpreemption3-sat
0
0 comments X

The pith

Message passing between LLM threads cuts context size and solves 25x25 Sudoku puzzles that defeat standard chain-of-thought and fork-join methods.

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

The paper introduces Message Passing Language Models in which separate LLM threads exchange information through explicit send and receive operations rather than sharing entire histories. This design lowers communication overhead by sending only necessary messages and lets threads stop early when partial results from others make continuation unnecessary. Experiments show the approach needs asymptotically less context on Sudoku grids than either sequential chain-of-thought or parallel fork-join scaling. A single fine-tuned model solves 25 by 25 puzzles that remain hard for conventional methods and for frontier models without tools. The same protocol also improves branch pruning on 3-SAT instances and produces competitive results on long-context question answering when applied to prompted base models.

Core claim

MPLMs let LLM threads communicate pointwise via lightweight send and receive primitives. The resulting framework achieves reduced communication costs by avoiding redundant context sharing and supports preemption that terminates unpromising threads once sufficient information arrives from peers. On Sudoku the method requires asymptotically smaller context than serial CoT or parallel FJ; fine-tuning one model enables solution of 25x25 instances that standard approaches and untuned frontier models cannot handle. On 3-SAT preemption improves efficiency by discarding failing branches early. Appropriately prompted large models follow the protocol and match popular fork-join baselines on long-conte

What carries the argument

Send and receive primitives that let transient LLM threads exchange only the minimal messages needed for coordination instead of broadcasting full context.

If this is right

  • Threads share only the messages required for the next step instead of duplicating full histories.
  • Preemption lets a thread halt once peer messages render its branch unnecessary, cutting wasted computation on 3-SAT.
  • A single fine-tuned model reaches 25x25 Sudoku grids that remain out of reach for CoT, FJ, and frontier models without tools.
  • Prompted base models achieve competitive long-context QA accuracy while following the same send/receive rules.

Where Pith is reading between the lines

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

  • The protocol could be extended to multi-model ensembles where different specialized models exchange partial solutions.
  • Early termination might compound across many parallel threads, producing larger savings on problems with high branching factors.
  • Message passing might allow modular reuse of intermediate results across unrelated queries without reloading full context each time.

Load-bearing premise

Large pre-trained LLMs can be prompted or fine-tuned to follow the send/receive protocol reliably and to interpret incoming messages without introducing substantial errors or extra context bloat.

What would settle it

A controlled run in which fine-tuned MPLMs on 25x25 Sudoku produce solution rates no better than standard CoT on 9x9 grids would show the claimed scaling does not hold.

Figures

Figures reproduced from arXiv: 2607.01077 by Andrea Zanette, Daman Arora, Gokul Swamy, Xuecheng Liu.

Figure 1
Figure 1. Figure 1: Comparison of reasoning paradigms for scaling test-time compute. Top: Serial Chain-of-Thought (CoT), where a single model generates a monolithic reasoning trace sequentially from prompt to answer. Middle: Fork-Join parallelism, where multiple independent transient CoT threads are spawned in parallel and periodically synchronized via global join operations before continuing. Bottom: Message Passing Language… view at source ↗
Figure 2
Figure 2. Figure 2: Left: This figure describes the communication pattern for a 9 × 9 Sudoku grid. Consider the worker for the cell (4, 4) highlighted in green. This cell only sends and receives messages from its neighbors highlighted in blue, reducing context requirement. Right: This figure describes how preemption can improve efficiency in 3-SAT problems where a search branch that finds a solution can send a message its par… view at source ↗
Figure 3
Figure 3. Figure 3: Scaling of sequential tokens (left) and maximum context tokens (right) required to solve a Sudoku puzzle as a function of the problem size (N4 ). Curves are shown on a log-log axes with fitted power-law exponents α. Message Passing (MPLM) exhibits significantly lower scaling exponents than the Serial and Fork-Join (FJ) reasoning paradigms for both sequential tokens and maximum context required, indicating … view at source ↗
Figure 4
Figure 4. Figure 4: Left: Average inference latency across 100 test problems per variable count for MPLM, FJ, and Serial. Right: Latency and speedup on the maximum-speedup problem at each variable count, observed in the problems whose search tree is highly unbalanced. Note that Serial can’t be measured after 12 variables because it exhausts the context window. Partition Qwen3-30B-A3B Qwen3.6-35B-A3B MPLM Accuracy (↑) MPLM Lat… view at source ↗
Figure 5
Figure 5. Figure 5: Respawning at a regular frequency (10 iterations in our setting) effectively reduces maximum worker context. This mechanism has two key effects. First, it prevents unbounded growth of per-worker context: as shown in [PITH_FULL_IMAGE:figures/full_fig_p031_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: In the example above, we note the latency of individual workers in the settings with added noise (right) v.s. no noise (left). The y-axis denotes the ID of the worker and x-axis denotes time in seconds. The start and end of a bar denotes the time frame for which a worker was active during inference. N Synchronization Schemes Under Variable Latencies Across Workers As we discuss in Section B.1, there are mu… view at source ↗
Figure 7
Figure 7. Figure 7: An example Sudoku where wait-for-any can be faster than wait-for-all sychronization. Consider the Sudoku drawn here where there are only 3 cells to be filled denoted by Wa, Wb and Wc. Note that Wa and Wc can be resolved after 2 iterations. However, Wb can only be resolved after either Wa or Wc is resolved. Therefore, Wb is critically dependent on Wa and Wc but not both. Even if one of them finishes, Wb can… view at source ↗
Figure 8
Figure 8. Figure 8: In the example above, we show message passing can be used to selective route information across different parallel threads. This figure is an illustrative example from the LongBench-v2 dataset using the Qwen3-30B-A3B model. O Long Context Question Answering MPLMs decompose long-context reasoning into two stages: parallel local processing followed by targeted message routing. Given a long input document or … view at source ↗
read the original abstract

While inference-time scaling has improved the reasoning abilities of large language models (LLMs), the need to generate long chains-of-thought (CoTs) is a computational bottleneck. Thus, in contrast to sequential scaling methods like CoT, recent parallel scaling techniques instead use fork and join (FJ) primitives to divide work across multiple LLM threads. However, in the fork-join paradigm, threads are typically transient and do not communicate pointwise with one another which limits scalability. To tackle this, we introduce Message Passing Language Models (MPLMs), a framework for LLM reasoning in which threads communicate directly via lightweight send and receive primitives. MPLMs enable efficient scaling through two key mechanisms: (1) reduced communication costs, achieved by avoiding redundant context sharing, and (2) preemption, which allows threads to terminate early based on partial information from their peers. We demonstrate the promise of MPLMs on 3 classes of tasks. First, on Sudoku puzzles, we show that MPLMs require an asymptotically smaller context than both serial CoT and parallel FJ. We then fine-tune a single model to solve 25 x 25 puzzles that remain challenging for standard CoT and FJ approaches, as well as frontier reasoning models without tools. Second, on 3-SAT puzzles, the capability of preemption allows termination of unpromising branches, which results in improved efficiency. Finally, we show that appropriately prompted large pre-trained models follow the MPLM protocol, achieving competitive results on long-context question answering relative to popular fork-join approaches.

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

3 major / 0 minor

Summary. The paper introduces Message Passing Language Models (MPLMs), a framework in which LLM threads communicate directly via lightweight send/receive primitives rather than relying on transient fork-join or serial CoT. It claims two efficiency mechanisms—reduced redundant context sharing and preemption based on partial peer information—and reports that MPLMs achieve asymptotically smaller context on Sudoku puzzles, that a single fine-tuned model solves 25x25 Sudoku instances that defeat standard CoT, FJ, and frontier models, that preemption improves efficiency on 3-SAT, and that prompted pre-trained models follow the protocol competitively on long-context QA.

Significance. If the empirical claims hold with verifiable context-length scaling and protocol adherence, the work would offer a concrete alternative to existing parallel scaling methods by enabling pointwise thread communication and early termination, potentially lowering inference cost for structured reasoning tasks. The absence of quantitative tables, error bars, or scaling plots in the available text, however, leaves the magnitude of any advantage unassessable.

major comments (3)
  1. [Abstract] Abstract: the central claim that MPLMs require an asymptotically smaller context than serial CoT and parallel FJ on Sudoku is stated without any reported token counts, scaling curves, or per-puzzle-size measurements; without these data the asymptotic reduction cannot be evaluated.
  2. [Abstract] Abstract: the assertion that a single fine-tuned model solves 25x25 Sudoku puzzles that remain challenging for CoT, FJ, and frontier models lacks success rates, baseline comparisons, or any description of how message interpretation errors or context bloat were measured or controlled.
  3. [Abstract] Abstract: the preemption mechanism on 3-SAT is described only qualitatively; no termination rates, branch-pruning statistics, or efficiency metrics relative to non-preemptive FJ are supplied, leaving the efficiency gain unsubstantiated.

Simulated Author's Rebuttal

3 responses · 0 unresolved

Thank you for the constructive feedback. We agree the abstract would benefit from quantitative details supporting the claims and will revise it to include key metrics from the experiments. We address each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that MPLMs require an asymptotically smaller context than serial CoT and parallel FJ on Sudoku is stated without any reported token counts, scaling curves, or per-puzzle-size measurements; without these data the asymptotic reduction cannot be evaluated.

    Authors: We acknowledge that the abstract states the asymptotic claim without supporting numbers. The full manuscript contains scaling curves, per-puzzle-size token counts, and context-length measurements in the Sudoku experiments demonstrating the reduction relative to CoT and FJ. We will revise the abstract to report these key quantitative results. revision: yes

  2. Referee: [Abstract] Abstract: the assertion that a single fine-tuned model solves 25x25 Sudoku puzzles that remain challenging for CoT, FJ, and frontier models lacks success rates, baseline comparisons, or any description of how message interpretation errors or context bloat were measured or controlled.

    Authors: The manuscript reports success rates, baseline comparisons, and controls for message errors and context in the experimental results. The abstract summarizes without these specifics. We will update the abstract to include the success rates for 25x25 puzzles along with brief mention of the evaluation controls. revision: yes

  3. Referee: [Abstract] Abstract: the preemption mechanism on 3-SAT is described only qualitatively; no termination rates, branch-pruning statistics, or efficiency metrics relative to non-preemptive FJ are supplied, leaving the efficiency gain unsubstantiated.

    Authors: We agree the abstract presents preemption qualitatively. The full paper supplies termination rates, branch-pruning statistics, and efficiency comparisons to non-preemptive FJ. We will incorporate these metrics into the revised abstract. revision: yes

Circularity Check

0 steps flagged

No circularity; new protocol introduced and evaluated empirically

full rationale

The paper presents MPLMs as a new framework using send/receive primitives, with claims about context reduction and puzzle-solving performance supported by direct experiments on Sudoku, 3-SAT, and QA tasks rather than any mathematical derivation. No equations, predictions, or central results reduce to fitted parameters, self-citations, or prior ansatzes by construction. The protocol is defined independently and tested against external benchmarks (CoT, FJ, frontier models), making the work self-contained without load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the domain assumption that LLMs can be made to follow a new communication protocol; no free parameters or invented physical entities are introduced.

axioms (1)
  • domain assumption Large pre-trained LLMs can be prompted or fine-tuned to follow the MPLM send/receive protocol without substantial errors
    Invoked when the abstract states that appropriately prompted models achieve competitive results and that a fine-tuned model solves 25x25 puzzles.
invented entities (1)
  • Message Passing Language Model (MPLM) no independent evidence
    purpose: Framework enabling direct thread-to-thread communication in LLM reasoning
    New named construct introduced to organize the send/receive primitives and preemption mechanism.

pith-pipeline@v0.9.1-grok · 5807 in / 1337 out tokens · 22396 ms · 2026-07-02T12:42:03.464965+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

56 extracted references · 56 canonical work pages · 1 internal anchor

  1. [1]

    Scaling Learning Algorithms Towards

    Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards

    work page

  2. [2]

    and Osindero, Simon and Teh, Yee Whye , journal =

    Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume =

    work page

  3. [3]

    2016 , publisher=

    Deep learning , author=. 2016 , publisher=

    work page 2016

  4. [4]

    Guo, Daya and Yang, Dejian and Zhang, Haowei and Song, Junxiao and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Zhang, Ruoyu and Ma, Shirong and Bi, Xiao and Zhang, Xiaokang and Yu, Xingkai and Wu, Yu and Wu, Z. F. and Gou, Zhibin and others , year=. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning , volume=. Nature , publisher=....

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

  5. [5]

    2024 , eprint=

    OpenAI o1 System Card , author=. 2024 , eprint=

    work page 2024

  6. [6]

    Second Conference on Language Modeling , year=

    Learning Adaptive Parallel Reasoning with Language Models , author=. Second Conference on Language Modeling , year=

    work page

  7. [7]

    2025 , eprint=

    Parallel-R1: Towards Parallel Thinking via Reinforcement Learning , author=. 2025 , eprint=

    work page 2025

  8. [8]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Multiverse: Your Language Models Secretly Decide How to Parallelize and Merge Generation , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    work page

  9. [9]

    Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author=. Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

    work page

  10. [10]

    Gonzalez and Clark Barrett and Ying Sheng , booktitle=

    Lianmin Zheng and Liangsheng Yin and Zhiqiang Xie and Chuyue Sun and Jeff Huang and Cody Hao Yu and Shiyi Cao and Christos Kozyrakis and Ion Stoica and Joseph E. Gonzalez and Clark Barrett and Ying Sheng , booktitle=. 2024 , url=

    work page 2024

  11. [11]

    2025 , eprint=

    Qwen3 Technical Report , author=. 2025 , eprint=

    work page 2025

  12. [12]

    The Eleventh International Conference on Learning Representations , year=

    Self-Consistency Improves Chain of Thought Reasoning in Language Models , author=. The Eleventh International Conference on Learning Representations , year=

    work page

  13. [13]

    2026 , eprint=

    Recursive Language Models , author=. 2026 , eprint=

    work page 2026

  14. [14]

    L ong B ench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks

    Bai, Yushi and Tu, Shangqing and Zhang, Jiajie and Peng, Hao and Wang, Xiaozhi and Lv, Xin and Cao, Shulin and Xu, Jiazheng and Hou, Lei and Dong, Yuxiao and Tang, Jie and Li, Juanzi. L ong B ench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks. Proceedings of the 63rd Annual Meeting of the Association for Computational...

    work page doi:10.18653/v1/2025.acl-long.183 2025

  15. [15]

    The First Workshop on System-2 Reasoning at Scale, NeurIPS'24 , year=

    Monte Carlo Tree Search Boosts Reasoning via Iterative Preference Learning , author=. The First Workshop on System-2 Reasoning at Scale, NeurIPS'24 , year=

    work page

  16. [16]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Tree of Thoughts: Deliberate Problem Solving with Large Language Models , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

    work page

  17. [17]

    2021 , eprint=

    Training Verifiers to Solve Math Word Problems , author=. 2021 , eprint=

    work page 2021

  18. [18]

    Advances in Neural Information Processing Systems , editor=

    Chain of Thought Prompting Elicits Reasoning in Large Language Models , author=. Advances in Neural Information Processing Systems , editor=. 2022 , url=

    work page 2022

  19. [19]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Self-Refine: Iterative Refinement with Self-Feedback , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

    work page

  20. [20]

    2024 , eprint=

    Training Language Models to Self-Correct via Reinforcement Learning , author=. 2024 , eprint=

    work page 2024

  21. [21]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Training Language Models to Reason Efficiently , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    work page

  22. [22]

    Second Conference on Language Modeling , year=

    L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning , author=. Second Conference on Language Modeling , year=

    work page

  23. [23]

    2025 , eprint=

    O1-Pruner: Length-Harmonizing Fine-Tuning for O1-Like Reasoning Pruning , author=. 2025 , eprint=

    work page 2025

  24. [24]

    2025 , eprint=

    Kimi k1.5: Scaling Reinforcement Learning with LLMs , author=. 2025 , eprint=

    work page 2025

  25. [25]

    2025 , eprint=

    The Markovian Thinker: Architecture-Agnostic Linear Scaling of Reasoning , author=. 2025 , eprint=

    work page 2025

  26. [26]

    2025 , eprint=

    InftyThink: Breaking the Length Limits of Long-Context Reasoning in Large Language Models , author=. 2025 , eprint=

    work page 2025

  27. [27]

    2025 , url=

    Chenxiao Yang and Nathan Srebro and David McAllester and Zhiyuan Li , booktitle=. 2025 , url=

    work page 2025

  28. [28]

    Proceedings of the 41st International Conference on Machine Learning , pages =

    Improving Factuality and Reasoning in Language Models through Multiagent Debate , author =. Proceedings of the 41st International Conference on Machine Learning , pages =. 2024 , editor =

    work page 2024

  29. [29]

    Forty-second International Conference on Machine Learning , year=

    Learning to Keep a Promise: Scaling Language Model Decoding Parallelism with Learned Asynchronous Decoding , author=. Forty-second International Conference on Machine Learning , year=

    work page

  30. [30]

    Skeleton-of-Thought: Prompting

    Xuefei Ning and Zinan Lin and Zixuan Zhou and Zifu Wang and Huazhong Yang and Yu Wang , booktitle=. Skeleton-of-Thought: Prompting. 2024 , url=

    work page 2024

  31. [31]

    and Luo, Hongyin and Glass, James R

    Schroeder, Philip and Morgan, Nathaniel W. and Luo, Hongyin and Glass, James R. THREAD : Thinking Deeper with Recursive Spawning. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.naacl-long.427

    work page doi:10.18653/v1/2025.naacl-long.427 2025

  32. [32]

    Atom of Thoughts for Markov

    Fengwei Teng and Quan Shi and Zhaoyang Yu and Jiayi Zhang and Yuyu Luo and Chenglin Wu and Zhijiang Guo , booktitle=. Atom of Thoughts for Markov. 2025 , url=

    work page 2025

  33. [33]

    2025 , eprint=

    ASPD: Unlocking Adaptive Serial-Parallel Decoding by Exploring Intrinsic Parallelism in LLMs , author=. 2025 , eprint=

    work page 2025

  34. [34]

    2nd AI for Math Workshop @ ICML 2025 , year=

    Instilling Parallel Reasoning into Language Models , author=. 2nd AI for Math Workshop @ ICML 2025 , year=

    work page 2025

  35. [35]

    2025 , eprint=

    ThreadWeaver: Adaptive Threading for Efficient Parallel Reasoning in Language Models , author=. 2025 , eprint=

    work page 2025

  36. [36]

    AutoGen: Enabling Next-Gen

    Qingyun Wu and Gagan Bansal and Jieyu Zhang and Yiran Wu and Beibin Li and Erkang Zhu and Li Jiang and Xiaoyun Zhang and Shaokun Zhang and Jiale Liu and Ahmed Hassan Awadallah and Ryen W White and Doug Burger and Chi Wang , booktitle=. AutoGen: Enabling Next-Gen. 2024 , url=

    work page 2024

  37. [37]

    2023 , url=

    Guohao Li and Hasan Abed Al Kader Hammoud and Hani Itani and Dmitrii Khizbullin and Bernard Ghanem , booktitle=. 2023 , url=

    work page 2023

  38. [38]

    Sirui Hong and Mingchen Zhuge and Jonathan Chen and Xiawu Zheng and Yuheng Cheng and Jinlin Wang and Ceyao Zhang and Zili Wang and Steven Ka Shing Yau and Zijuan Lin and Liyang Zhou and Chenyu Ran and Lingfeng Xiao and Chenglin Wu and J. Meta. The Twelfth International Conference on Learning Representations , year=

    work page

  39. [39]

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

    Leviathan, Yaniv and Kalman, Matan and Matias, Yossi , title =. Proceedings of the 40th International Conference on Machine Learning , articleno =. 2023 , publisher =

    work page 2023

  40. [40]

    The Twelfth International Conference on Learning Representations , year=

    AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors , author=. The Twelfth International Conference on Learning Representations , year=

    work page

  41. [41]

    2024 , eprint=

    APAR: LLMs Can Do Auto-Parallel Auto-Regressive Decoding , author=. 2024 , eprint=

    work page 2024

  42. [42]

    2023 , editor =

    Gao, Luyu and Madaan, Aman and Zhou, Shuyan and Alon, Uri and Liu, Pengfei and Yang, Yiming and Callan, Jamie and Neubig, Graham , booktitle =. 2023 , editor =

    work page 2023

  43. [43]

    2023 , html =

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle =. 2023 , html =

    work page 2023

  44. [44]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Toolformer: Language Models Can Teach Themselves to Use Tools , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

    work page

  45. [45]

    1994 , isbn =

    Gropp, William and Lusk, Ewing and Skjellum, Anthony , title =. 1994 , isbn =

    work page 1994

  46. [46]

    1998 , isbn =

    Snir, Marc and Otto, Steve and Huss-Lederman, Steven and Walker, David and Dongarra, Jack , title =. 1998 , isbn =

    work page 1998

  47. [47]

    Proceedings of the 41st International Conference on Machine Learning , pages =

    Zhuge, Mingchen and Wang, Wenyi and Kirsch, Louis and Faccio, Francesco and Khizbullin, Dmitrii and Schmidhuber, J\". Proceedings of the 41st International Conference on Machine Learning , pages =. 2024 , editor =

    work page 2024

  48. [48]

    2024 , eprint=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. 2024 , eprint=

    work page 2024

  49. [49]

    2022 , url=

    Eric Zelikman and Yuhuai Wu and Jesse Mu and Noah Goodman , booktitle=. 2022 , url=

    work page 2022

  50. [50]

    Neural Information Processing Systems , year=

    Attention is All you Need , author=. Neural Information Processing Systems , year=

    work page

  51. [51]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) , address=

    LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) , address=. 2024 , url=

    work page 2024

  52. [52]

    Kingma and Jimmy Ba , editor =

    Diederik P. Kingma and Jimmy Ba , editor =. Adam:. 3rd International Conference on Learning Representations,. 2015 , url =

    work page 2015

  53. [53]

    2023 , eprint=

    RoFormer: Enhanced Transformer with Rotary Position Embedding , author=. 2023 , eprint=

    work page 2023

  54. [54]

    Communications of the ACM , volume =

    Martin Davis and George Logemann and Donald Loveland , title =. Communications of the ACM , volume =

    work page

  55. [55]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics , year =

    LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics , year =

    work page

  56. [56]

    Recursive Language Models

    Recursive Language Models , author =. arXiv preprint arXiv:2512.24601 , year =

    work page internal anchor Pith review Pith/arXiv arXiv