The method manages long-context search histories during extended discovery.
Automated discovery is a search problem over combinatorially large spaces. It requires reasoning to explore promising ideas while handling long contexts.
Introducing ARTS, a search algorithm that (i) reasons to guide exploration and (ii) test-time trains to handle long context
Users thank the researchers for introducing the ARTS algorithm thanks to its awesome collaborators and advisors who made the automated discovery work possible.
No Digg Deeper questions have been answered for this story yet.
This search over hypotheses and experiments requires making two crucial decisions: what to try next, and when to persist versus when to pivot.
We ask whether a Large Reasoning Model is better suited to take these decisions compared to the heuristics they have long relied on.
Thanks to the awesome collaborators @arnavkj95 @deepaknathani11 and advisors @xwang_lk @WilliamWangNLP for making this happen!
Code: https://github.com/UCSB-AI/arts Arxiv: https://arxiv.org/abs/2606.21891 Webpage: https://automating-discovery.github.io/arts
While AI scientists are getting better, two core challenges remain: (i) deciding what to explore next and (ii) handling the growing context.
ARTS addresses both with reasoning-guided exploration and test-time RL to learn from long search histories.
Automated discovery is a search problem over combinatorially large spaces. It requires reasoning to explore promising ideas while handling long contexts.
Introducing ARTS, a search algorithm that (i) reasons to guide exploration and (ii) test-time trains to handle long context
Hourly progression under a fixed 8-hour budget shows interesting results. ARTS trails for the first hour or two, catches up mid-run, and finishes above the strongest contemporary method. We find slow start is because it spends early budget exploring broadly rather than committing to the first high scoring hypothesis.
Across MLGym and MLEBench, ARTS outperforms leading methods like AIRA and MLEvolve by 15.3%. On tasks like Battle of Sexes and RSNA Brain Tumor it even surpasses the human-best.
And it is. ARTS reasons over prior logs, code, and training curves to propose strong hypotheses, separate a weak hypothesis from a weak execution, and decide when to persist or pivot. Verbalized sampling keeps the hypotheses diverse and TTT instills the history into the weights
When the search history outgrows the context window, ARTS* test time trains the scientist on its own search using reinforcement learning, applying GRPO with a percentile reward to instill the search into the weights. A 4B scientist then matches o3 at much lower inference cost.
Interestingly, on Meta Maze the trained 4B scientist discovers optimized recurrent-memory solution. The o3 scientist recognizes the task needs recurrent memory, but drifts off as its history outgrows context. Prior discovery methods abandon recurrent memory after early failure.
