vedant33/supersede-demo-zerogpu

Supersede: base vs trained, live

Pick a multi-session conversation in which a fact about the user changes, then run base Qwen2.5-3B and the GRPO-trained adapter side by side under the exact bounded-memory protocol the paper evaluates: each model keeps a capped NOTES memory, never re-sees raw sessions, and answers from memory alone. Watch the base model keep the stale value while the trained model overwrites it with the current one.

Requires ZeroGPU hardware (Settings, Change hardware, ZeroGPU). Greedy decoding, so results are deterministic.

📄 Paper · 🤖 Model · 💻 GitHub · 🌐 Prime Intellect Hub