Meta study finds quantized reasoning models struggle with overthinking

Meta AI's FAIR team finds quantized reasoning models overthink 7.3x more, abandoning correct answers. A training-free fix cuts errors by up to 58%.

Raccontata dacryptobriefing.com