pith. sign in

arxiv: 2605.28919 · v1 · pith:MLE5FI4Nnew · submitted 2026-05-27 · 💻 cs.LG · cs.AI· cs.CL

CosmicFish-HRM: Adaptive Reasoning via Hierarchical Recurrent Mechanisms in Compact Language Models

Venkat Akhil Lakkapragada This is my paper

Pith reviewed 2026-06-29 14:00 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CL
keywords adaptive reasoninghierarchical reasoning modulecompact language modelsdynamic computationtransformer architecturerecurrent mechanismsinference efficiencyvariable reasoning depth
0
0 comments X

The pith

A compact language model learns to vary the number of reasoning steps it applies to each input.

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

The paper introduces CosmicFish-HRM, a small transformer-based language model equipped with a Hierarchical Reasoning Module that runs iterative high-level and low-level reasoning cycles. The model learns a halting condition so that it spends more cycles on complex inputs and fewer on simple ones. This produces non-uniform reasoning behavior across tasks rather than applying the same fixed computation to everything. The authors argue that the extra infrastructure adds overhead at small scale but should become relatively cheaper and more useful as the base model grows larger. The central demonstration is that the learned allocation of reasoning effort already occurs without being explicitly programmed.

Core claim

By integrating a Hierarchical Reasoning Module into a compact language model that also uses Grouped Query Attention, RoPE, and SwiGLU, the system learns to perform a variable number of high-level and low-level reasoning cycles and to halt based on input complexity, resulting in different amounts of computation being allocated to different tasks and inputs instead of uniform fixed-depth processing.

What carries the argument

The Hierarchical Reasoning Module (HRM), which executes iterative high-level and low-level reasoning cycles and learns when to stop according to input complexity.

If this is right

  • The model automatically assigns more reasoning cycles to harder inputs and fewer to easier ones.
  • The overhead of the additional reasoning infrastructure becomes proportionally smaller at larger model sizes.
  • Adaptive reasoning depth provides an alternative route to improved reasoning performance besides simply increasing parameter count.
  • Reasoning capability can emerge from learned control over computation depth rather than from fixed architecture depth alone.

Where Pith is reading between the lines

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

  • Measuring total compute across an entire benchmark set would show whether the variable depth actually reduces average cost compared with fixed-depth baselines.
  • The same halting logic could be tested in other iterative architectures such as recurrent vision models or planning agents.
  • If the learned stopping rule transfers across domains, it might reduce the need for task-specific prompt engineering that forces extra reasoning steps.

Load-bearing premise

The added cost of the hierarchical reasoning module will become relatively smaller and more advantageous as the size of the base language model increases.

What would settle it

Training larger versions of the model and measuring whether they reach higher reasoning accuracy per total floating-point operation than standard transformers of matched size on the same tasks; if the efficiency gap does not appear or reverses, the adaptive approach does not deliver the claimed scaling benefit.

Figures

Figures reproduced from arXiv: 2605.28919 by Venkat Akhil Lakkapragada.

Figure 1
Figure 1. Figure 1: Overview of the CosmicFish-HRM architecture. Input tokens are embedded and processed by [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Training and validation loss curves for CosmicFish-HRM. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Average HRM reasoning steps during training and validation. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Mean HRM reasoning steps across benchmark tasks. More complex or retrieval-heavy tasks tend to [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗
read the original abstract

Large language models have achieved strong reasoning capabilities, though often at the cost of massive parameter counts and expensive inference. In this work, we explore a different direction: adaptive reasoning depth in compact language models. We present CosmicFish-HRM, a compact language model built around a Hierarchical Reasoning Module (HRM) that dynamically allocates computational effort during inference. Instead of applying fixed computation to every input, the model iterates through high-level and low-level reasoning cycles and learns when to halt based on input complexity. CosmicFish-HRM combines this adaptive reasoning core with modern transformer components including Grouped Query Attention, RoPE, and SwiGLU activations. While the additional reasoning infrastructure introduces overhead at small scale, we hypothesize that this tradeoff becomes increasingly favorable as model size grows and the relative cost of the HRM core diminishes. Our results show that the model learns non-uniform reasoning behavior, allocating different numbers of reasoning steps across tasks and inputs. These findings suggest that adaptive reasoning depth may offer a promising alternative to relying solely on parameter scale for reasoning capability.

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 introduces CosmicFish-HRM, a compact language model built around a Hierarchical Reasoning Module (HRM) that performs iterative high-level and low-level reasoning cycles and learns a halting condition based on input complexity. It integrates this module with standard transformer components (Grouped Query Attention, RoPE, SwiGLU) and reports that the resulting model exhibits non-uniform reasoning behavior by allocating different numbers of reasoning steps across tasks and inputs; a scaling hypothesis is stated that the overhead of the HRM core becomes relatively less costly at larger model sizes.

Significance. If the empirical claims hold, the work would be significant as a concrete demonstration that adaptive computation depth can be learned inside compact models rather than relying exclusively on parameter count, providing a potential efficiency route for reasoning tasks.

major comments (1)
  1. Abstract: the central claim that 'our results show that the model learns non-uniform reasoning behavior' is stated without any accompanying data, tables, figures, error bars, ablation studies, or methodological details sufficient to evaluate support for the claim.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their thoughtful review and the opportunity to clarify the presentation of our results. Below we respond directly to the single major comment.

read point-by-point responses

  1. Referee: Abstract: the central claim that 'our results show that the model learns non-uniform reasoning behavior' is stated without any accompanying data, tables, figures, error bars, ablation studies, or methodological details sufficient to evaluate support for the claim.

    Authors: We agree that an abstract should not make unsupported claims. The manuscript body (Section 4 and Figures 2–4) contains the supporting evidence: histograms of per-input reasoning-step counts across tasks, tables reporting mean and variance of step allocation with error bars from multiple runs, and ablations isolating the halting mechanism. The abstract, however, is written as a high-level summary and therefore omits these specifics. We will revise the abstract to either (a) qualify the claim (“as shown in Section 4”) or (b) add one sentence referencing the observed distribution of reasoning depths, keeping the abstract within length limits. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The manuscript text consists of high-level architectural description and empirical observations without any equations, derivation chains, fitted parameters presented as predictions, or load-bearing self-citations. The claim that the model learns non-uniform reasoning behavior is stated as an observed result from training, not derived by construction from its own inputs. The scaling tradeoff is explicitly labeled a hypothesis rather than a proven result. No steps match any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No specific details on parameters, axioms or entities available from the abstract alone.

pith-pipeline@v0.9.1-grok · 5713 in / 1045 out tokens · 54737 ms · 2026-06-29T14:00:22.121810+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

31 extracted references · 17 canonical work pages · 15 internal anchors

  1. [1]

    GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints

    Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zeiler, Sumit Sanghai, and Yuexin Xu. GQA: Training generalized multi-query transformer models from multi-head checkpoints.arXiv preprint arXiv:2305.13245, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  2. [2]

    PonderNet: Learning to Ponder

    Andrea Banino, Jan Balaguer, and Charles Blundell. PonderNet: Learning to ponder.arXiv preprint arXiv:2107.05407, 2021

    work page arXiv 2021

  3. [3]

    Pythia: A suite for analyzing large language models across training and scaling

    Stella Biderman, Hailey Schoelkopf, Quentin Anthony, Herbie Bradley, Kyle O’Brien, Eric Hallahan, Moham- mad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. Pythia: A suite for analyzing large language models across training and scaling. InProceedings of the 40th International Conference on Machine Learning, 2023

    2023

  4. [4]

    PIQA: Reasoning about physical commonsense in natural language

    Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. PIQA: Reasoning about physical commonsense in natural language. InProceedings of the AAAI Conference on Artificial Intelligence, 2020

    2020

  5. [5]

    On the Measure of Intelligence

    Franc ¸ois Chollet. On the measure of intelligence.arXiv preprint arXiv:1911.01547, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1911

  6. [6]

    Think you have solved question answering? Try ARC, the AI2 reasoning challenge

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? Try ARC, the AI2 reasoning challenge. 2018

    2018

  7. [7]

    Universal Transformers

    Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Łukasz Kaiser. Universal transformers. arXiv preprint arXiv:1807.03819, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  8. [8]

    Adaptive Computation Time for Recurrent Neural Networks

    Alex Graves. Adaptive computation time for recurrent neural networks.arXiv preprint arXiv:1603.08983, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  9. [9]

    Textbooks Are All You Need

    Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio C´esar Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, et al. Textbooks are all you need.arXiv preprint arXiv:2306.11644, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  10. [10]

    Distilling the Knowledge in a Neural Network

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network.arXiv preprint arXiv:1503.02531, 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  11. [11]

    TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension

    Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. InProceedings of the 55th Annual Meeting of the Association for Computational Linguistics, 2017

    2017

  12. [12]

    Farrar, Straus and Giroux, 2011

    Daniel Kahneman.Thinking, Fast and Slow. Farrar, Straus and Giroux, 2011

    2011

  13. [13]

    Natural questions: A benchmark for question answering research

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. Natural questions: A benchmark for question answering research. volume 7, pages 453–466, 2019

    2019

  14. [14]

    Human-level control through deep reinforcement learning.Nature, 518(7540):529–533, 2015

    V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning.Nature, 518(7540):529–533, 2015

    2015

  15. [15]

    GPT-4 technical report

    OpenAI. GPT-4 technical report. Technical report, OpenAI, 2023. URL https://arxiv.org/abs/2303. 08774

    2023

  16. [16]

    Training language models to follow instructions with human feedback

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.arXiv preprint arXiv:2203.02155, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  17. [17]

    Using the Output Embedding to Improve Language Models

    Ofir Press and Lior Wolf. Using the output embedding to improve language models.arXiv preprint arXiv:1608.05859, 2017. 16

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  18. [18]

    Language models are unsupervised multitask learners.OpenAI Blog, 2019

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners.OpenAI Blog, 2019

    2019

  19. [19]

    Mixture-of-Depths: Dynamically allocating compute in transformer-based language models

    David Raposo, Sam Ritter, Blake Richards, Timothy Lillicrap, Peter Humphreys, and Adam Santoro. Mixture of depths: Dynamically allocating compute in transformer-based language models.arXiv preprint arXiv:2404.02258, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [20]

    WinoGrande: An adversarial winograd schema challenge at scale

    Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. WinoGrande: An adversarial winograd schema challenge at scale. InCommunications of the ACM, 2021

    2021

  21. [21]

    Tran, Yi Tay, and Donald Metzler

    Tal Schuster, Adam Fisch, Jai Gupta, Mostafa Dehghani, Dara Bahri, Vinh Tran, Yi Tay, and Donald Metzler. Confident adaptive language modeling.arXiv preprint arXiv:2207.07061, 2022

    work page arXiv 2022

  22. [22]

    GLU Variants Improve Transformer

    Noam Shazeer. GLU variants improve transformer.arXiv preprint arXiv:2002.05202, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2002

  23. [23]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. InInternational Conference on Learning Representations, 2017

    2017

  24. [24]

    RoFormer: Enhanced Transformer with Rotary Position Embedding

    Jianlin Su, Yu Lu, Shengding Pan, Ahmed Murtadha, Bo Wen, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding.arXiv preprint arXiv:2104.09864, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  25. [25]

    Llama 2: Open Foundation and Fine-Tuned Chat Models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  26. [26]

    Hierarchical Reasoning Model

    Guan Wang, Jin Li, Yuhao Sun, Xing Chen, Changling Liu, Yue Wu, Meng Lu, Sen Song, and Yasin Abbasi Yadkori. Hierarchical reasoning model.arXiv preprint arXiv:2506.21734, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  27. [27]

    On layer normalization in the transformer architecture

    Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tie-Yan Liu. On layer normalization in the transformer architecture. InProceedings of the 37th International Conference on Machine Learning, 2020

    2020

  28. [28]

    HellaSwag: Can a machine really finish your sentence? InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2019

    Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. HellaSwag: Can a machine really finish your sentence? InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2019

    2019

  29. [29]

    Root mean square layer normalization

    Biao Zhang and Rico Sennrich. Root mean square layer normalization. InAdvances in Neural Information Processing Systems, volume 32, 2019

    2019

  30. [30]

    TinyLlama: An Open-Source Small Language Model

    Peiyuan Zhang, Guangtao Zeng, Tianhao Wang, and Wei Lu. TinyLlama: An open-source small language model. arXiv preprint arXiv:2401.02385, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  31. [31]

    OPT: Open Pre-trained Transformer Language Models

    Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. OPT: Open pre-trained transformer language models.arXiv preprint arXiv:2205.01068, 2022. 17

    work page internal anchor Pith review Pith/arXiv arXiv 2022