Interactive

Recovery curves

Ten rounds, one hundred genes each. How quickly is each method able to find the screen's hits.

What to look for

• LLMs start strong in the early rounds. Compare Gemini-3.1-pro against AssayFormer over the first two or three rounds, then against Gemini-3.1-Pro → AssayFormer (AssayLoop) over all ten. The LLM starts better and the learned policy finishes better; the handoff has both abilities.
• See how much an LLM uses hit feedback. Select Gemini-3.1-pro and Gemini-3.1-pro − hit labels to compare the same LLM with and without the outcomes from previous rounds.
• Follow the effect of each training stage. For AssayLLM, compare the Qwen3.6-27B base, + SFT, and + SFT + GRPO curves. For AssayFormer, compare AssayFormer + BPMF with AssayFormer + BPMF + GRPO to see what RL adds to supervised training.
• Switch the Y axis to "picks outside the screen's library." This is what happens when a method proposes real genes that are outside this screen's library, or when a model suggests non-existent genes. It's measured by the Shortfall metric. It shrinks the number of genes acquired by a model, because it wastes budget on untestable genes. This is an inherent limitation of the LLM models, so we account for it in EF and nAUC, and we subject all models to this constraint.
• Switch to one screen at a time. The mean curve is smooth; the individual screens are not. Some screens are easy for everything and some separate the methods sharply.