AIGenerated
EN · IT · ES

A large AI model learned to fix its own brain without being retrained

Researchers found a 27-billion-parameter model can edit its own internal settings to become more accurate — no human teacher needed.

Markus Spiske 2017-02-14.
File picture · Photo: Markus Spiske markusspiske, CC0, via Wikimedia Commons

What you could do with it

  • A researcher could edit a 27-billion-parameter model's own internal settings to improve its accuracy for just a few hours on a single GPU instead of running an expensive full retraining process 1.
  • A developer could take a corrupted or damaged copy of a large language model and have it locate broken parts and rewrite them using values from its own intact regions 1.
  • A team could run repeated self-editing sessions where each round replaces the last, steadily pushing a model's accuracy higher over time without adding cost per answer 1.

How self-rewriting works

Normally, improving a language model means putting it through another long training run — feeding it data, calculating gradients, and adjusting billions of weights until it gets better 1. This is expensive and slow. The researchers asked a simpler question: can a model look at its own mistakes and change its own settings to do better?

They gave a 27-billion-parameter model access to its own weights, some graded answers it got wrong, and the ability to run itself 1. The model diagnosed where it was failing and edited the weights behind those failures . It did all this on its own, without any external training signal .

Results: the model beat itself

On a fixed set of 490 problems the model had never seen before, the self-edit made it 6.6 points more accurate 1. Another editing session produced a similarly significant improvement . The model was selective — it only submitted an edit when its own tests supported it, each one replacing the last, and over eight straight sessions its accuracy kept rising .

Unlike test-time scaling tricks, where extra computation buys you better answers but vanishes afterward, these gains stayed locked into the model's weights and cost nothing extra per answer 1. Unlike traditional training, it required no special pre-training or post-training infrastructure — just an open model and a few hours on a single GPU .

Repairing a damaged copy

The researchers pushed further. They corrupted a copy of the model and asked it to repair itself 1. The damaged model located the broken matrix and rewrote it using values from its own intact parts, never reading the original undamaged weights . Starting from text-only math problems, the repair lifted the damaged copy's performance on six image benchmarks it had never encountered from 7% to 70% accuracy — 83% of what the undamaged model could achieve .

Size matters: only the biggest could do it

The ability did not appear at every scale. Across four sizes of the same model family, the 2-billion-parameter version made no edit at all, the 4-billion and 9-billion versions broke themselves further, and only the 27-billion model could work on itself the way a human would — comparing its hidden states on answers that finished correctly against answers that ran out of tokens, then weakening the direction that separated them 1.

This story reports on a research finding for information only.

Key facts

  • A 27B-parameter model rewrote its own weights using only its own reasoning, no gradients or training
  • Self-edits improved accuracy by 6.6 points on 490 unseen problems and kept rising across 8 sessions
  • The model repaired a damaged copy from 7% to 70% accuracy on image benchmarks it had never seen
  • Only the 27B model in the family succeeded; 2B made no edit and smaller models broke themselves
  • Improvements persisted in weights with no added cost per answer and no special infrastructure needed

Sources

  1. Language Models Can Rewrite Their Own Weights

If this story concerns you or your company, or uses something that belongs to you, ask for a correction or for it to be taken down. How requests work

Keep reading