# ⚡️ SparseTech

**Redefining LLM reliability for edge AI through variance reduction, sparse knowledge distillation, and probability-domain manifold correction.**

Standard benchmarks measure whether a model is correct; SparseTech measures whether a model is *reliable*. We believe that in agentic and edge AI, **hallucinations live in variance**. Our mission is to crush stochastic variance and stabilize reasoning without relying on massive, server-side inference ensembles.



---

## 📚 Foundational Research
We believe models should be built upon a rigorous axiomatic framework recently published in early 2026. You can read our core methodology here:

### The Core Theory
* **[Hallucinations Live in Variance (Jan 2026)](https://huggingface.co/papers/2601.07058)**
  *Introduces Semantic Stability (SS) and Paraphrase Consistency (PC@k) as the true metrics for LLM reliability.*

### The Distillation Framework
* **[Sparse Knowledge Distillation: A Mathematical Framework... (Jan 2026)](https://huggingface.co/papers/2601.03195)**
* **[Multi-Teacher Ensemble Distillation (Jan 2026)](https://huggingface.co/papers/2601.09165)**
* **[Recursive Meta-Distillation: An Axiomatic Framework... (Jan 2026)](https://huggingface.co/papers/2601.13100)**
* **[Adaptive Weighting in Knowledge Distillation (Jan 2026)](https://huggingface.co/papers/2601.17910)**

### Advanced Manifold Correction
* **[Post-Training Probability Manifold Correction via Structured SVD Pruning... (Jan 2026)](https://huggingface.co/papers/2602.00372)**