Orthogonalized Read Improves Recurrent Memory Training, Not Capacity

Keston Aquino-Michaels· July 23, 2026 View original

Summary

A recent study found that orthogonalizing the mLSTM memory matrix during read operations significantly improves noisy associative recall, but this effect is due to re-conditioning the learning problem during training, not an actual memory capacity enhancement. The orthogonalized read acts as a removable training scaffold.

A previous report highlighted that orthogonalizing the memory matrix in mLSTMs during read operations, specifically through Newton-Schulz iterations, substantially boosted noisy associative recall. This new research replicates the effect but clarifies its underlying mechanism. It demonstrates that the improvement is not a fundamental enhancement of memory capacity. Instead, the orthogonalized read functions as a "training scaffold," re-conditioning the learning problem during the initial plateau phase of training, which is typically characterized by a long period of chance performance before a sharp improvement. The effect is self-consistent, uniform across learning rates and hardness levels, and crucially, removable. The study shows that the orthogonalized read can be annealed away after the model escapes the plateau, leaving a numerically standard mLSTM with full accuracy. This suggests that recall benchmarks used for architecture selection may inadvertently measure trainability rather than pure capacity, and that "emergence" in these systems can arise from censored metrics over gradually accumulating internal structure.

Why it matters

AI researchers and engineers designing recurrent neural networks should understand that certain architectural "improvements" might primarily act as training facilitators rather than intrinsic capacity enhancements, guiding more effective model development and evaluation.

How to implement this in your domain

  1. 1Investigate training techniques that re-condition the learning problem in recurrent networks, rather than solely focusing on architectural changes.
  2. 2Implement annealing schedules for training scaffolds like orthogonalized reads to optimize training efficiency without impacting final model performance.
  3. 3Rethink how recurrent memory benchmarks are designed and interpreted, ensuring they differentiate between trainability and inherent capacity.
  4. 4Apply variance decomposition to understand the true sources of performance gains in novel RNN architectures.

Who benefits

AI/ML ResearchSoftware DevelopmentRoboticsNatural Language Processing

Key takeaways

  • Orthogonalized reads improve mLSTM performance by re-conditioning the learning problem, not by increasing memory capacity.
  • This technique acts as a "training scaffold" that can be removed after the model learns.
  • Recall benchmarks may inadvertently measure model trainability rather than intrinsic memory capacity.
  • Understanding the true mechanisms behind performance gains is crucial for effective RNN development.

Original post by Keston Aquino-Michaels

"arXiv:2607.19390v1 Announce Type: new Abstract: A recent report finds that orthogonalizing the mLSTM memory matrix at read time (five Newton-Schulz iterations, trained through) substantially improves noisy associative recall. The effect replicates, but it is not a memory improvem…"

View on X

Originally posted by Keston Aquino-Michaels on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses