Charlot Lab · Quantum information at the edge · tensor-network compression
A policy's weights, like an entangled quantum state, need not be stored in full if their structure is low. A tensor network (here the simplest one, a rank-r factorization) writes a big weight as small cores. This trains a little control policy live in your browser, then asks the honest question two ways. Compress-after (red): squeeze the finished dense policy to rank r. Train-compact (teal): train the policy in factored form from the start. Slide the rank and watch which one keeps the policy. No quantum computer — the classical mathematics of entanglement, run on your own device.
What you are seeing. The curves are measured live, not asserted. Compress-after stays low at every rank because a from-scratch dense weight sits near its high-rank random start, so truncating it throws the policy away. Train-compact jumps to full accuracy at a small rank, because the control function itself is low-complexity and the factored form has room to fit it. That is the whole deployable lesson: train the tensor-network form, don't squeeze a dense one (on a pretrained model, truncate then fine-tune). The same result at scale is what the tensorization literature reports. Offline in the lab's research record: a pure-Rust factored kernel that cuts the on-device multiply count by the same factor, and this exact bench in Python. More: Quantum information at the edge.