Autoresearch v2 Binder Discovery Results
Live public snapshot of strict, globally deduplicated binder discovery. The page keeps the observed tree and recorded hypotheses so the search trajectory can be inspected, while internal ranking mechanics are omitted.
Results
Cumulative Strict Global-Unique Discovery Curves
Compare methods by strict global-unique discoveries versus accumulated Proteina-Complexa GPU-hours. Curves are stepwise cumulative counts; archived 48h adaptive runs are included as result evidence.
Selected Target Snapshot
Run Table
Tree Search Snapshot
The tree view shows observed nodes, parent-child structure, stored intent/hypothesis, and measured outcomes. It is treated here as result evidence; internal ranking mechanics are not displayed.
Select a node
Definitions
Strict Global Unique
Sequences are deduplicated across the whole run before counting. A repeated successful sequence from a later node does not increase the cumulative count.
Baseline
Official-style seeded baseline runs use the clean official Proteina-Complexa code on one GPU per target. The public baseline repeats the paper-style beam-search setting: task-specific target, beam-search, n_branch=4, beam_width=4, step_checkpoints=[0,100,200,300,400], max_batch_size=16, and a new random seed for each repeat.
Adaptive Runs
Adaptive runs build a search tree over candidate generation settings and use measured outcomes from previous nodes to decide the next nodes.