Adaptive Multi-Task Learning

EvoNashMTL

EvoNashMTL is a research prototype that asks a narrow question with a concrete mechanism: can Nash bargaining-based multi-task weighting be improved by a slower evolutionary scheduler that adapts task influence over time?

The current repository is strongest as a method and stability project. The toy harness shows that the outer scheduler can move, can alter final task weights relative to plain Nash-MTL, and can avoid the old one-hot collapse mode after repair.

Core Idea

Two levels of task weighting

The method keeps Nash-MTL as the inner bargaining rule at each training step, then adds a slower replicator-style scheduler that updates task shares over time. Final weights are formed by modulating Nash weights with those scheduler shares.

Current Evidence

Mechanism first, benchmarks later

The clearest current evidence is toy-level: the scheduler is active, the final optimization signal changes relative to plain Nash-MTL, and the hard stability case can be stabilized. Real benchmark superiority is still unconfirmed.

What This Site Covers

About

The project framing, research question, and current scope.

Results

All five toy figures, plus the strongest takeaways from the notes in imgs/.

Method

The two-level weighting design, tuned toy setup, and why the scheduler modulates Nash instead of replacing it.

Repo

Where the code lives, how to run the prototype, and what to inspect during experiments.