How yifanzhang-pro/recurrent-looped-tranformer works
This repository is a research-paper and results showcase for Recurrent Looped Transformer, not an implementation package. Its reader workflow is to learn the model’s token-update architecture, compare feedback variants, then inspect benchmark and generalization results. The README documents the model relationships; no source-code excerpts were provided, so the graph represents documented behavior rather than verified runtime calls.
Main entry points: README.md Project overview · experiments-depth8/README.md Depth-eight results · results-refresh-20260920/README.md Depth-sixteen results · Recurrent_Looped_Transformer.pdf Research paper
Components 16
RLT Architecture
- Causal encoder
README.md - Global KV memory
README.md - Gated merge
README.md - Recurrent decoder
README.md - Layerwise SWA cache
README.md - Hidden-state feedback
README.md
Feedback Variants
- RLT-0 control
README.md - RLT-2 chunks
README.md
Experiments
- Depth-eight results
assets/experiments-depth8/README.md - Depth-sixteen results
assets/results-refresh-20260920/README.md - Mod-5 variants
assets/results-refresh-20260920/README.md
Research Materials
- Project overview
README.md - Research paper
Recurrent_Looped_Transformer.pdf - Architecture diagrams
assets/architecture-detail.pdf - Chunk schedule
assets/rlt2/chunk-schedule.pdf
Other components
- Research reader
Connections 16
- Research reader → Project overview: reads overview
README.md - Project overview → Research paper: links paper
README.md - Project overview → Architecture diagrams: documents architecture
README.md - Causal encoder → Global KV memory: projects KV
README.md - Causal encoder → Gated merge: provides representation
README.md - Hidden-state feedback → Gated merge: provides prior state
README.md - Gated merge → Recurrent decoder: feeds decoder
README.md - Global KV memory → Recurrent decoder: supplies context
README.md - Layerwise SWA cache → Recurrent decoder: retains local KV
README.md - Recurrent decoder → Hidden-state feedback: updates state
README.md - Project overview → RLT-0 control: describes control
README.md - Project overview → RLT-2 chunks: describes chunks
README.md - RLT-2 chunks → Chunk schedule: illustrated by
README.md - Project overview → Depth-eight results: reports experiments
README.md - Project overview → Depth-sixteen results: reports results
README.md - Project overview → Mod-5 variants: reports variants
README.md