Ouro / LoopLM — looped latent reasoning (CONSOLIDATED → SIGMA0-OURO-CODER.md)

🔁 Consolidated. The loop mechanism now lives in the coder doc.

→ SIGMA0-OURO-CODER.md § The loop mechanism is the single source of truth. It carries the full Q-exit math, both loop implementations, and the codebase map that used to live here.

This page is retained as a redirect so existing references (research notes, handoffs) still resolve.

One-paragraph summary

Ouro / LoopLM (Scaling Latent Reasoning via Looped Language Models, arXiv:2510.25741; PDF: docs/research-papers/ouro-looped-llm-2510.25741.pdf) builds reasoning into computation depth by reusing weight-tied layers R times in latent space, with a learned Q-exit gate that exits at the first recurrent step where CDF(t) ≥ q. Lantern is the recurrent-depth engine behind the local coder. Two implementations:

• Native latent loop — src/sigma0/loop_lm.py (Sigma0LoopLM)

applies Q-exit over Ouro's pretrained weight-tied block + exit gate (inference-time only; reports realized mean_depth). Served drop-in via scripts/ouro_serve.py on :11434; deep mode is OURO_NATIVE=1.

• API re-prompt loop — lib/loop-reasoner.js

(loopedReason()) approximates the loop for any plain model by re-prompting up to MAX_LOOPS and exiting via cdfExit() (threshold_met / converged / max_loops).

For the full treatment — the paper mechanisms, both loops, the "where it maps in the codebase" table, the inference modes, and honest scope — see SIGMA0-OURO-CODER.md.