Researchers find faster way to update AI model knowledge

Researchers discovered that large language models have a few specific internal layers where knowledge updates work best, and developed a method to find these layers quickly without trial-and-error. This makes it cheaper and faster to correct or update facts in deployed AI systems without breaking their other capabilities.

If AI systems can be efficiently corrected after deployment, it matters for any organization relying on them to stay accurate — but this is still a lab result on synthetic benchmarks with no evidence of real-world deployment impact or cost savings.