How Weak Agents Make Strong Agents Stronger - An Interactive WMSS Demo

My peers and I came across a paper by Chen et al.(2026) that looks at what happens after LLMs finish their initial training phase.

What caught our attention: at some point, the model becomes so confident in its answers that it actually stops improving - even with more training.

The paper proposes a novel approach - using an older, weaker version of the model to keep pushing the stronger one forward.

We turned it into an interactive demo where you can:

• Step through SFT training and watch gradients vanish
• Drag a lambda slider to see logit mixing in action
• Compare SFT vs WMSS epoch by epoch
• Walk through the full training pipeline with animations

No ML background needed.

Try the interactive demo

Paper: Chen et al. (2026) - "How Weak Agents Make Strong Agents Stronger"(arXiv:2602.08222)