analiza aiqbench
This research reveals that token representations in Pythia models follow nonlinear dynamics governed by quadratic drift and significant rotational components, challenging standard linear approximations used to analyze transformer layers. By modeling these trajectories as a discrete Langevin process, the study demonstrates that flows do not descend their own log-density but instead follow distinct potentials while preserving specific angular ranks across network depths. These findings offer crucial mathematical insights into how large language models internally manipulate semantic information without relying on hype or marketing claims.
The Depth Flow of Token Representations Is Nonlinear and Does Not Descend Its Own Density
arXiv:2608.29706v1 Announce Type: new Abstract: A token's representation is carried through the network layer by layer. The whole vocabulary carried together forms a flow. We fit this flow's equation of motion as a discrete Langevin model over corpus-mean trajectories of Pythia-160M and Pythia-410M, and score the predicted steps on held-out tokens. Linear maps are often used as cheap surrogates for a layer. The flow they summarize is not linear: a quadratic drift beats the linear linear map at every transition of both models, and the Kramers--Moyal estimator agrees wherever its neighborhoods stay local. We then characterize the flow further. First, we show that it does not descend its own log-density. The drift instead descends a potential that is not the density. Second, the rotational component is not negligible, $4$ to $45\%$ of the explainable drift, and the circulation shows in what the flow preserves: a token keeps its angular rank across all thirteen layers while its norm rank is shuffled and its concentration rank is reversed by the last block.
Abstract:A token's representation is carried through the network layer by layer. The whole vocabulary carried together forms a flow. We fit this flow's equation of motion as a discrete Langevin model over corpus-mean trajectories of Pythia-160M and Pythia-410M, and score the predicted steps on held-out tokens. Linear maps are often used as cheap surrogates for a layer. The flow they summarize is not linear: a quadratic drift beats the linear linear map at every transition of both models, and the Kramers--Moyal estimator agrees wherever its neighborhoods stay local. We then characterize the flow further. First, we show that it does not descend its own log-density. The drift instead descends a potential that is not the density. Second, the rotational component is not negligible, $4$ to $45\%$ of the explainable drift, and the circulation shows in what the flow preserves: a token keeps its angular rank across all thirteen layers while its norm rank is shuffled and its concentration rank is reversed by the last block.
Computation and Language (cs.CL)
arXiv:2608.29706 [cs.CL]
(or arXiv:2608.29706v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2608.29706 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 260 cuvinte
Subiecte conexe
Symmetry-enriched topological order in tensor networks: Defects, gauging and anyon condensation
Quantum Journal · 1 sept. 2026
Representation Learning with Quantum Signal Processing
arXiv · 1 sept. 2026
How can LLM RL Work Despite Information-Theoretic Inefficiency
LessWrong · 16 oct. 2026
Contextual Representation Flow Across Depth in State-Space and Transformer Architectures
arXiv · 14 oct. 2026