Scaling and testing approximate E-prop-like algorithms on realistic language and long-context tasks

Establish how approximate online credit-assignment algorithms based on E-prop perform when scaled and tested on realistic, challenging tasks such as language tasks requiring reasoning and long-context recall.

Background

Beyond mathematical derivations, the authors stress the need to evaluate approximate online learning methods like E-prop on real-world tasks. They specifically call out language tasks involving reasoning and extended context as benchmarks that can reveal whether such approximations remain effective outside simplified settings.

This open question targets practical validation and comparative assessment against BPTT or other strong baselines in applied domains.