Stanford CS 329Z lecture argues filtering agent training data beats raw volume

Stanford's CS 329Z lecture 12 argues that filtering agent training data matters more than data volume, especially when data is scarce.

Meng Shao reposted the lecture on X on November 4, 2026. The lecture cites two examples. In SWE-bench Verified, 68.3% of tasks were filtered out. In LIMI, a model trained on 78 curated trajectories scored 73.5% on AgencyBench, compared with 47.8% for a model trained on 10,000 trajectories.

The figures come from the lecture as reposted by Shao. This report does not include independent verification of the benchmark results.