[KongchangAI]
Paper

Quiet-STaR

Quiet-STaR is a method proposed by Eric Zelikman and others in 2024 for training language models to perform internal reasoning. It extends the STaR framework by having the model generate multiple thought steps before producing each token, with the thoughts optimized through reinforcement learning or rejection sampling, improving performance on complex reasoning tasks.

Timeline (last 90 days)

Nov 2

The open-source code of the Quiet-STaR paper (Eric Zelikman et al., 2024) uses direct fine-tuning with bf16 weights and the AdamW optimizer and may have inadvertently triggered the weight-freezing phenomenon.

Unverified50%
Nov 2

Re-running Quiet-STaR's training recipe with an optimizer capable of applying every update caused model performance to diverge; the high-precision optimizer required lowering the learning rate by 3 to 10 times to obtain the same results and could not achieve better results.

Unverified50%

All Facts (2)

Source Articles