Model
palm
Appears alongside
Featured in threads
Tracks
- Ideas & essays 2
- Benchmarks & progress 2
- Models & capabilities 1
- Compute & infrastructure 1
"Let's think step by step" elicits zero-shot reasoning
A single prompt phrase, with no worked examples, lifted GSM8K accuracy from 10.4% to 40.7% — chain-of-thought without the exemplars.
Ideas & essays · Benchmarks & progress
Google announces PaLM at 540 billion parameters
Trained on the Pathways system across two TPU v4 pods, it posted large gains on reasoning benchmarks and explained its own jokes.
Models & capabilities · Compute & infrastructure
Chain-of-thought prompting is described
Asking a model to show its working improved reasoning benchmarks sharply, with no retraining — the seed of the later reasoning models.
Ideas & essays · Benchmarks & progress