dorsal/arxiv
View SchemaEnergy-Entropy Regularization: The True Power of Minimal Looped Transformers
| Authors | Wai-Lun Lam |
|---|---|
| Categories | |
| ArXiv ID | 2601.09588vv2 |
| URL | https://arxiv.org/abs/2601.09588 |
| License | http://arxiv.org/licenses/nonexclusive-distrib/1.0/ |
Abstract
Recent research suggests that looped Transformers have superior reasoning capabilities compared to standard deep architectures. Current approaches to training single-head looped architectures on benchmark tasks frequently fail or yield suboptimal performance due to a highly non-convex and irregular loss landscape. In these settings, optimization often stagnates in poor local minima and saddle points of the loss landscape, preventing the model from discovering the global minimum point. The internal mechanisms of these single-head looped transformer models remain poorly understood, and training them from scratch remains a significant challenge. In this paper, we propose a novel training framework that leverages Tsallis entropy and Hamiltonian dynamics to transform the geometry of the loss landscape. By treating the parameter updates as a physical flow, we successfully trained a single-head looped Transformer with model dimension $d = 8$ to solve induction head task with input sequence length of 1000 tokens. This success reveals the internal mechanism behind the superior reasoning capability.
{
"annotation_id": "37ab356b-3c04-4f5a-9c33-5f518444f238",
"date_created": "2026-02-17T05:53:19.953000Z",
"date_modified": "2026-02-17T05:53:19.953000Z",
"file_hash": "0c6dff723dee1536572d348c9c4605356adf34f28da6f12ce9e2abf97ea5377e",
"private": false,
"record": {
"abstract": "Recent research suggests that looped Transformers have superior reasoning capabilities compared to standard deep architectures. Current approaches to training single-head looped architectures on benchmark tasks frequently fail or yield suboptimal performance due to a highly non-convex and irregular loss landscape. In these settings, optimization often stagnates in poor local minima and saddle points of the loss landscape, preventing the model from discovering the global minimum point. The internal mechanisms of these single-head looped transformer models remain poorly understood, and training them from scratch remains a significant challenge. In this paper, we propose a novel training framework that leverages Tsallis entropy and Hamiltonian dynamics to transform the geometry of the loss landscape. By treating the parameter updates as a physical flow, we successfully trained a single-head looped Transformer with model dimension $d = 8$ to solve induction head task with input sequence length of 1000 tokens. This success reveals the internal mechanism behind the superior reasoning capability.",
"arxiv_id": "2601.09588",
"authors": [
"Wai-Lun Lam"
],
"categories": [
"cs.LG"
],
"license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/",
"title": "Energy-Entropy Regularization: The True Power of Minimal Looped Transformers",
"url": "https://arxiv.org/abs/2601.09588",
"version": "v2"
},
"schema_id": "dorsal/arxiv",
"source": {
"execution_id": "bbdcce71-b1ab-498c-bb10-c6fed523a3b8",
"id": "arXiv Dataset",
"type": "Model",
"variant": "snapshot-2026-01-17",
"version": "0.1.0"
},
"user_id": 1000002
}