Denoising Looped Models

I want to evaluate and, if it fits, integrate a technique from recent AI research into my project.

## Technique: Denoising Looped Models
New denoising trick trains looped models to hit 58.8% on ARC-AGI-1

AlphaSignal signal (2026-09-14): A denoising training objective applied to looped (weight-shared) transformers achieves 58.8% on ARC-AGI-1 — comparable to models 33× larger. The trick: inject Gaussian noise before each loop iteration (cosine schedule, σ_max→σ_min) and train the model to recover clean hidden states. This prevents loop collapse and creates a stable gradient signal across iterations. 90M looped model vs 350M standard transformer: 58.8% vs 38.2% at one-third the training compute.

What the reference demo shows: Gradio demo with 4 sub-tabs — (1) Architecture Overview: standard vs looped vs denoising-looped comparison; (2) Denoising Schedule: noise decay across loop iterations; (3) ARC-AGI Results: benchmark comparison; (4) Implementation Guide: 5 patterns for applying the denoising trick.

## Reference implementation (Vector Ridge Labs, POC-036)
- Live demo: https://osoman2-gradio-pocs.hf.space/?poc=POC-036
- Source code: https://huggingface.co/spaces/osoman2/Gradio_pocs/tree/main/denoising_looped
Original source: AlphaSignal signal · 2026-09-14

## My project
<Describe your stack, the problem you want to solve and where this technique would sit. Example: "Python FastAPI RAG service over 50k PDFs, Postgres + pgvector, OpenAI embeddings, p95 latency budget 800 ms.">

## What I need from you
1. Read the reference source code and explain in at most 5 bullets how the technique works and what it would change compared to my current approach.
2. Decide whether it fits my case. If it does not, say why and stop there.
3. Propose the smallest viable integration: files or modules to touch, new dependencies, configuration.
4. Implement it behind a flag or as a separate function so it can be compared side by side with what I have today.
5. Add an evaluation script that measures latency, throughput and cost per request at equal output quality, before and after, on my own data.
6. List the risks (latency, cost, failure modes) and what to monitor once it is in production.

Constraints: keep my existing stack and conventions, avoid heavy dependencies unless the gain is clear, and reference the file in the demo source whenever you reuse its logic.