Sleep-time Compute: Beyond Inference Scaling at Test-time
What it is
This repository contains the code and data necessary to reproduce the empirical results from the 'Sleep-time Compute' research paper by Letta and UC Berkeley. It provides scripts for running stateful AIME and GSM-Symbolic experiments using a Letta server to demonstrate how agents can perform background computation during idle periods to improve test-time performance.
The problem it solves
Current LLM inference scaling relies heavily on increasing test-time compute (e.g., longer chains of thought), which is expensive and slow; this work addresses the inefficiency by leveraging 'sleep-time' for background processing.
What you could use it for
- Reproducing the AIME 2024/2025 benchmark results for sleep-time compute
- Evaluating the impact of background agent processing on GSM-Symbolic tasks
- Researching alternative scaling laws for LLM inference beyond test-time compute
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.