SwarmRSI

The SwarmRSI framework comprehensively sets new SOTA across 118 tasks on the three benchmarks OpenProblems, NatureBench and MLEBench Lite.

The programs generated by SwarmRSI have been used by biologists in real intestinal and spleen immunology research.

At the core of SwarmRSI is a shared graph structure that records all inheritable research states, including each agent's hypotheses, programs, experimental evidence and unresolved requirements.

On NatureBench tasks, removing the H mutation reduces the proportion matching SOTA by 10 percentage points, and removing the P mutation reduces it by 13.3%.

Turning off the recombination operation directly drops the proportion exceeding SOTA from 15.6% to 7.8%.

SwarmRSI uses a discounted UCB metric to score the two frontier sets of modifying hypotheses and modifying programs, comprehensively accounting for token cost, number of experiments and unresolved-requirement pressure.

With DeepSeek V4 Pro as the backbone, SwarmRSI achieves an average score of 74.2 on OpenProblems, exceeding the previous SOTA of 67.1.

With DeepSeek V4 Pro as the backbone, SwarmRSI's proportion matching SOTA on NatureBench is 21.1 percentage points higher than Claude Code with the same backbone.

With DeepSeek V4 Pro as the backbone, SwarmRSI achieves a medal rate of 81.8% on MLEBench Lite.

After switching to Gemini 3.1 Pro Preview, SwarmRSI's score on OpenProblems rises to 76.2.

Search trajectory analysis shows that subsequent optimization by SwarmRSI converts 35% of failed implementations into new best-performing valid programs, whereas the compared methods AEDA and ARBOR achieve only 12% and 5% respectively.

Among the 90 tasks in NatureBench, 77% of the final best programs came from at least one recombination, crossover or feedback-triggered update.

In cross-task transfer experiments, the research states inherited from three source tasks raised the average base value of 12 held-out NatureBench tasks from 12.01 under cold start to 21.89, and the proportion matching SOTA jumped from 41.7% to 86.1%.

Two programs generated by SwarmRSI have actually been adopted by biologists in intestinal and spleen immunology research, where in the intestinal study the Spearman correlation coefficients of the post-processing representation MHC-II scores reached 0.77 and 0.69, higher than the 0.65 and 0.40 of RNA PCA; in the spleen study the metric improved from 0.045 to 0.102.

The paper's failure analysis annotated 644 failed attempts, of which 93.3% were faithful implementations of modifications that brought no improvement, and only 6.7% were implementation or execution problems.