Early-Signal Detector

architecture · first seen 4 Oct 2026

looped transformer

Also: LoopLM, looped model, layer looping, latent looping, parameter tying, layer repetition, looped reasoning, selective looping

A transformer that reuses shared blocks over repeated recurrent loops to increase depth without adding parameters.

Looped transformers reuse one shared block, or a small stack of blocks, over several recurrent passes, so effective depth grows without adding parameters. Recent arXiv work has shifted from asking whether looping helps reasoning to making it practical. A Hochreiter-group paper and a Qualcomm-affiliated paper target the per-loop KV-cache memory that blocks dynamic depth in models like Ouro. LoopFill applies speculative decoding to cut decoding latency. Rectified and flow-inspired variants stabilize hidden states across loops, and energy-based papers propose principled stopping rules. The idea is also spreading beyond language, into diffusion (LiFT), MRI reconstruction and multimodal embeddings, along with a few small open-source repos.

If the cache, latency and stability problems get solved, looped models could offer adjustable test-time compute per token from a single small checkpoint, an alternative to generating chain-of-thought tokens. The energy-based critiques suggest the design is still unsettled, so it is worth watching which formulation wins.

Analyst note (emerging), 17 Nov 2026.

Mentions per week

01219 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Rank 20 on 16 Nov 2026 · 37 mentions in 28 days · 34 independent groups · papers → code · first seen 6 weeks ago · acceleration +0.02

Earliest sources

Latest sources

Mentioned together with

speculative decodingmixture of expertsKV cacheadaptive reasoningnormalized transformertabular foundation modelKV cache quantizationlength generalizationadaptive-depth inferenceresidual connections