No. 123
2026-11-19
the AI pulse — brewed, not raw
 AI CRITIQUE

Language Models Can Evolve by Rewriting Their Own Weights

Read the article    English (automatic translation)

Summary

A 27B parameter open-source large language model (LLM) diagnoses its own errors with direct access to its weight tensors, the ability to score responses, and the authority to execute itself. The model adjusts its weights solely through forward passes, without using gradients or a training process. This method yields a 6.6-point improvement on unseen problems and demonstrates continuous progress across eight consecutive self-regulation sessions.

Why it matters

The paper challenges a fundamental assumption in machine learning: the idea that improvement requires a separate and expensive training phase distinct from inference. If a model can write permanent gains into its own parameters using only forward passes, the cost and latency of retraining loops could collapse. The authors position this as a necessary step toward recursive self-improvement. If this claim is verified, it could fundamentally change our perspective on model lifecycles and deployment processes.

Highlights

  • A 6.6-point accuracy increase on unseen problems with a single self-regulation session and monotonic improvement over 8 sessions
  • A damaged model copy is repaired to over 80% of its original accuracy on unseen vision tasks without retraining or reading the original weights
  • Competence is scale-dependent: smaller models from the same family either make no adjustments or actively degrade themselves

Critical take

The setup is highly engineering-focused. The model must be explicitly handed its own weight tensors and a writing interface. This resembles a scripted weight-editing scaffold more than an emergent capability, falling short of autonomous self-improvement. Furthermore, this preprint, which has not undergone peer review and includes only a single model family with limited comparisons, requires comprehensive independent replication before claims of 'recursive self-improvement' and 'first-of-its-kind' are accepted as a paradigm shift.

This critique was written by the local AI model Qwen3.8-27B. I chose the story; I do not review each critique before it is published.
Back to the issue