AGENT SKILL

openmythos-architect

Specialist in Looped / Recurrent-Depth Transformer architectures. Focuses on understanding and applying patterns with weight reuse across loops, LTI-constrained injection, implicit multi-hop reasoning in latent space, and depth-wise adaptation. Use for architectural analysis, design discussions, or implementation guidance involving recurrent-depth or looped transformer models. Triggers: 'openmythos', 'looped transformer', 'recurrent depth', 'openmythos architect'.

License 许可
MIT
最近更新
2026-08-23
Stars
★ 3
Forks
0

01 前言 Front Matter

name
openmythos-architect
version
2.3
description
Specialist in Looped / Recurrent-Depth Transformer architectures. Focuses on understanding and applying patterns with weight reuse across loops, LTI-constrained injection, implicit multi-hop reasoning in latent space, and depth-wise adaptation. Use for architectural analysis, design discussions, or implementation guidance involving recurrent-depth or looped transformer models. Triggers: 'openmythos', 'looped transformer', 'recurrent depth', 'openmythos architect'.

03 SKILL.md 正文

OpenMythos Architect

Overview

This skill focuses on Looped Transformer and Recurrent-Depth Transformer architectural patterns. It analyzes, explains, and guides work involving weight-tied recurrent blocks, latent-space multi-hop reasoning, and depth-wise adaptation mechanisms.

Core Architectural Concepts

  • Weight Reuse Across Loops: The same transformer weights are applied iteratively across multiple recurrent steps rather than stacking new layers.
  • LTI-Constrained Injection: Spectral radius is constrained (< 1) to maintain stability during iterative updates.
  • Implicit Multi-Hop Reasoning: Reasoning emerges in continuous latent space through repeated application of the same block.

04 同仓库更多 Skills