Meta finds quantized reasoning models often doubt the right answer instead of finishing
rohanpaul_ai · x · 2026-07-22
A Meta paper argues that quantized reasoning models often fail not because they never reach the right answer, but because they hesitate after finding it. According to the authors, aggressive post-training quantization can make models more likely to reopen a problem mid-answer by preferring hesitation tokens such as “wait,” “but,” or “alternatively.”
The study evaluates 5 reasoning models, multiple quantization methods, and model sizes from 1.5B to 32B across math, coding, and science tasks. Main findings:
- Aggressive quantization increased overthinking failures by up to 52%.
- Adding a small penalty on 50 hesitation words reduced reasoning length by 12%–23%.
- Accuracy was often maintained or improved.
- Overthinking errors were reduced by up to 58% in some settings.
The authors frame this as an efficient decoding fix for compressed models that need to save memory and cost without sacrificing reasoning quality.
More from Research
- UK AISI: under $80 of fine-tuning fully restores filtered bio capabilities in open-weight models — connoraxiotes · 2026-10-27
- RFD4-Proteina runs on an RTX 3060: protein design from $0.001 per design — AllThingsApx · 2026-10-27
- University of Freiburg's Generative Intelligence Lab recruiting funded PhDs via ELLIS, apply by Oct 31 — yuyinzhou_cs · 2026-10-27
- Nature Physics: reprogrammable refractive index enables arbitrary multimode wave control for ML — MarioKrenn6240 · 2026-10-27
- Wakaru open-sources ontologize: labeling datasets from model activations — amplifiedamp · 2026-10-27
- EMNLP Oral paper traces multilingual representations in LLMs with cross-layer transcoders — ZhijingJin · 2026-10-27