The Tarka Architecture: An Integrated Framework for Principled, Self-Correcting, Self-Evolving Language Models
Year: 2026
Type: preprint
Abstract: A systems-integration and measurement paper, not a capability paper. Every technique required to build a language model that reasons before it answers, calibrates and resolves its own uncertainty, corrects itself, discovers tools at runtime, remembers across shutdowns, and updates its own weights is already published; no published system integrates them into one coherent, reproducible, self-evolving whole. We specify that the system — Tarka — build it at a small scale (a ~10M-parameter base model, a single consumer GPU, total training cost under $50), and measure the full stack. Contributions: (1) a complete, reproducible architecture specification integrating ten principled components; (2) a three-layer evolution taxonomy (persist/consolidate/adapt) that scopes "self-evolving" to bounded, measured behavior and explicitly excludes AGI claims; (3) a 14-benchmark reproducible suite, including a registry-reuse ratio for agentic tool use, a persistence-across-power-cycle benchmark (priority not claimed against concurrent work), and a sycophancy benchmark scoring truth-over-agreement as a constitutional behavior; and (4) honest small-scale measurements, with negative and modest results pre-registered rather than dropped. The central claim is architectural, not capability: the architecture is scale-invariant; results scale with parameters. No component is claimed as invented; every constituent technique is cited to its origin.
Author: Verma, Abhijeet
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7168344599