Research

One underlying question: how does training signal behave, and can it be steered?

Hydra (Thread 1: AI for Software Optimization) treats a model's own per-metric training loss as a live feedback signal, reading it every epoch to decide which objective (latency, memory, energy) a code LLM should sample next, so multi-objective preference learning self-corrects instead of relying on fixed weights. Variance Collapse (Thread 2: Optimization Dynamics) asks the inverse question: when does that same kind of signal, the per-unit gradient gate, disappear on its own? It derives, from a single training-free quantity, whether gate density survives or collapses before training even begins. Hydra acts on training signal; Variance Collapse derives when that signal exists to act on.