Datasets:
Languages:
English
ArXiv:
Tags:
large-language-models
nonlinear-dynamics
dynamical-systems
transformers
edge-of-stability
repetition
License:
|
Download DICTIONARY.md from maorfarid/three-clocks: direct link, hf CLI and curl.
- Browser
- Download file 4.51 kB
-
https://huggingface.co/datasets/maorfarid/three-clocks/resolve/main/DICTIONARY.md
- Command line
-
hf download hf://datasets/maorfarid/three-clocks/DICTIONARY.md
-
curl -L -o DICTIONARY.md https://huggingface.co/datasets/maorfarid/three-clocks/resolve/main/DICTIONARY.md
4.51 kB
The Three-Clock Dictionary
From Large Language Models as Nonlinear Dynamical Systems: The Three Clocks of Depth, Training, and Generation (Farid, 2026). Fifteen concepts from nonlinear dynamics, mapped onto the three clocks of a transformer language model. Evidence: P proved, M measured, C conjectured, ? open. Full citations are in the paper and in PAPERS.md. Corrections and new entries are welcome as pull requests.
| Concept | Clock 1: layer time | Clock 2: training time | Clock 3: generation time |
|---|---|---|---|
| Fixed point | Tokens collapse to a cluster; rank collapse [P] (Geshkovski et al. 2023; Dong et al. 2021). Massive activations nearly input independent [M] (Sun et al. 2024) | Convergence to a minimum; neural collapse at the end of training [M] (Papyan et al. 2020) | Period-one loops (one token repeated) as the simplest case of degenerate repetition [M] (Holtzman et al. 2019) |
| Linear stability, Jacobian spectrum | Dynamical isometry; layer Jacobians of trained models [P/M] (Pennington et al. 2017; Fernando et al. 2026) | Sharpness vs. 2/eta [M] (Cohen et al. 2021) | Derivative of the feedback map factorizes through 1/T [M] (Jones 2026) |
| Finite-time Lyapunov exponent | Ridges in input space separate classes [M] (Storm et al. 2023); per-layer profile over training [M] (this paper, Exp. 1) | Sensitivity of runs to minibatch noise early in training [M] (Frankle et al. 2019) | Divergence of nearby generations [M] (Jaca et al. 2026; Jones 2026) |
| Edge of chaos, criticality | Order-chaos transition at initialization sets trainable depth [P] (Poole et al. 2016; Schoenholz et al. 2016) | Edge of stability [M] (Cohen et al. 2021) | Critical temperature with power-law text statistics [M] (Nakaishi et al. 2024) |
| Limit cycle (periodic orbit) | ? | Period-2 orbits of GD [P] (Chen et al. 2022); SGD converges to cycles [P] (Chaudhari et al. 2017) | Repetition loops [M] (Xu et al. 2022); paraphrase 2-cycles [M] (Wang et al. 2025); loops converge in hidden-state space [M] (this paper, Exp. 3) |
| Period doubling, route to chaos | ? | Larger step sizes give period doubling and chaos in minimal models [P] (Zhu et al. 2022; Kong et al. 2020); period-2 orbit, then irregular oscillation, in a transformer [M] (this paper, Exp. 2) | Feigenbaum scaling in embedding feedback [M] (Jones 2026) |
| Fractal boundary | ? | Trainable vs. divergent hyperparameters [M] (Sohl-Dickstein 2024); interleaved trainable and divergent step sizes [M] (this paper, Exp. 2); fractal attractors of optimizers [M] (Tuci et al. 2026) | ? |
| Basin of attraction | Class regions bounded by FTLE ridges [M] (Storm et al. 2023) | Which minimum is reached depends on step size and batch size [M] (Jastrzębski et al. 2017) | Prompts that do or do not fall into loops [C] |
| Metastability, slow manifold | Multi-cluster states last exponentially long [P] (Geshkovski et al. 2024) | Plateaus [P] (Saxe et al. 2013); slow phase of pretraining [M] (Zhang et al. 2026; Wen et al. 2024) | Long coherent stretches before drift [C] |
| Saddle and ghost | ? | Saddles dominate [M] (Dauphin et al. 2014); abrupt learning from a ghost [P] (Dinc et al. 2025) | ? |
| Phase transition | ? | Grokking [P/M] (Rubin et al. 2023; Power et al. 2022); induction heads [M] (Olsson et al. 2022) | Low-T repetitive vs. high-T incoherent phase [M] (Nakaishi et al. 2024; Arnold et al. 2024); narrow coherent window between them, also in a reasoning model [M] (this paper, Exp. 3) |
| Noise-driven (Kramers) escape | ? | Escape of SGD from minima [P] (Xie et al. 2020; Mori et al. 2021) | Escape from repetition loops, logit gap as barrier [P/M] (this paper, Prop. 1, Exp. 3); locking temperature [P] (Prop. 3), horizon-dependent in practice [M] (this paper, Exp. 3) |
| Synchronization, consensus | Clustering as consensus; Kuramoto form [P] (Geshkovski et al. 2023; Li et al. 2026); synchronization by noise [P] (Agazzi et al. 2026) | ? | Collective dynamics of interacting models [P] (Krishnamurthy et al. 2026) |
| Koopman, linearization | Local linear description of each layer [M] (Fernando et al. 2026) | Koopman-based training and pruning [M] (Dogra et al. 2020; Redman et al. 2021); lazy regime [P] (Jacot et al. 2018) | ? |
| Conservation law, symmetry | Gradient-flow structure of attention [P] (Sander et al. 2021) | Symmetries give conserved quantities that finite steps break [P] (Kunin et al. 2020) | ? |