Symbol-Equivariant Recurrent Reasoning Models (Mar 2026)

Title: Symbol-Equivariant Recurrent Reasoning Models (Mar 2026)
Link: http://arxiv.org/abs/2603.02193v1
Date: March 2026

Summary:
This paper introduces Symbol-Equivariant Recurrent Reasoning Models (SE-RRMs), a new class of neural architectures designed for structured reasoning tasks like Sudoku and ARC-AGI. Unlike traditional Recurrent Reasoning Models (RRMs) that handle symbol symmetries through data augmentation, SE-RRMs bake permutation equivariance into the architecture itself. This allows them to outperform prior models on Sudoku, generalize to unseen grid sizes, and achieve competitive results on ARC-AGI with significantly fewer parameters and less data augmentation.

Key Topics:
- Recurrent Reasoning Models (RRMs)
- Symbol Equivariance
- Permutation Equivariance
- Sudoku Solving
- ARC-AGI
- Architectural Symmetry
- Generalization and Extrapolation

Chapters:
00:00 - SE-RRM Breakthrough Overview
01:26 - Analyzing Reasoning Bottlenecks
02:31 - Explaining Recurrent Models
03:56 - Deep Supervision Training
04:54 - Defining Symbol Equivalence
05:51 - Visualizing Sudoku Symmetries
06:43 - Implementing 3D Tensors
07:45 - Shared Symbol Embeddings
08:43 - Dual Attention Layers
09:45 - Calculating Scaling Complexity
11:03 - Zero-Shot Extrapolation Results
13:07 - Benchmarking ARC-AGI Performance
13:51 - Testing Architectural Robustness
15:16 - Scaling Inference Compute
16:51 - Future Efficiency Outlook

Stock video credits:
- Silviu Din - https://www.pexels.com/@silviu-din-1620549
- Soumya - https://www.pexels.com/@soumya-1446957
- Google DeepMind - https://www.pexels.com/@googledeepmind
- Colors Motion Graphics - https://www.pexels.com/@colors-motion-graphics-183847699
- Magda Ehlers - https://www.pexels.com/@magda-ehlers-pexels
- Adis Resic - https://www.pexels.com/@adis-resic-297996969
- Pressmaster - https://www.pexels.com/@pressmaster
- Max Fischer - https://www.pexels.com/@max-fischer
- olia danilevich - https://www.pexels.com/@olia-danilevich
- cottonbro studio - https://www.pexels.com/@cottonbro
- Bedrijfsfilmspecialist.nl - https://www.pexels.com/@bedrijfsfilmspecialist-nl-1284006
- Vlada Karpovich - https://www.pexels.com/@vlada-karpovich
- Pixabay - https://www.pexels.com/@pixabay
- Colin Jones - https://www.pexels.com/@larchmedia
- Pachon in Motion - https://www.pexels.com/@pachon-in-motion-426015731
- Oleg Gamulinskii - https://www.pexels.com/@oleg-gamulinskii-755060
- Pavel Danilyuk - https://www.pexels.com/@pavel-danilyuk
- Yaroslav Shuraev - https://www.pexels.com/@yaroslav-shuraev
- Anete Lusina - https://www.pexels.com/@anete-lusina
- fauxels - https://www.pexels.com/@fauxels
- Tima Miroshnichenko - https://www.pexels.com/@tima-miroshnichenko
- Kindel Media - https://www.pexels.com/@kindelmedia
- Thirdman - https://www.pexels.com/@thirdman
- Mikhail Nilov - https://www.pexels.com/@mikhail-nilov
- MART PRODUCTION - https://www.pexels.com/@mart-production
- Cyriac von Czapiewski - https://www.pexels.com/@cyriac-von-czapiewski-1601520
- tunnel motions - https://www.pexels.com/@tunnelmotions
- Ron Lach - https://www.pexels.com/@ron-lach
- Ketut Subiyanto - https://www.pexels.com/@ketut-subiyanto
- Caleb Oquendo - https://www.pexels.com/@caleboquendo
- Nino Souza - https://www.pexels.com/@ninosouza
- Tiger Lily - https://www.pexels.com/@tiger-lily
- StefWithAnF - https://www.pexels.com/@stefwithanf-1955763
- Engin Akyurt - https://www.pexels.com/@enginakyurt
- Anthony 🙂 - https://www.pexels.com/@inspiredimages

Видео Symbol-Equivariant Recurrent Reasoning Models (Mar 2026) канала AI Paper Slop