PSGD-QEQ equals KL-AIRM up to step size: YouJiacheng corrects gradient uniqueness claim
YouJiacheng · x · 2026-09-10
YouJiacheng issued a correction on his PSGD optimization analysis: the gradient isn't really unique, but is "coincidentally" offset by geometry difference. With the derivation dp=d(q²)=2qdq=-2q²r=-2pr, PSGD-QEQ equals KL-AIRM up to a step size, since the study of p happens in the eigen space where everything commutes.
Related event: Correction: PSGD and KL-AIRM Gradients Differ Only by a Step Size(2 posts)→
More from Research
- TorchRL 0.14 ships DreamerV3 faster than Jax implementations and async collectors — jsuarez · 2026-09-10
- Bryan Johnson responds to Michael Levin's peer-reviewed Platonic Space paper: bodies as collective intelligence — AllThingsApx · 2026-09-10
- OpenBMB open-sources training data: 400B tokens of code and 500K agent samples — zibuyu9 · 2026-09-10
- RobustSGPO lifts agent harness completion from 60% to 80% on held-out tasks — dair_ai · 2026-09-10
- Jack Clark proposes pre-registering AI economy forecasts to score predictors in a year — jackclarkSF · 2026-09-10
- TokenPrint: open-source interactive explorer for tokens, attention, hidden states and KV cache seeks contributors — Rich-Fruit-326 · 2026-09-10