HarnessTTT: Learn to Optimize Harness at Test-Time for Scientific Discovery
What this work does
Trains a harness proposer at test time while keeping the executor fixed; generated prompts, skills and execution controls guide budgeted discovery rollouts whose feedback updates proposer weights.
Connection to discovery
A direct test-time learning mechanism for improving the procedure used to search for discoveries.
Evidence
The body separates executor-weight, context-only and proposer-weight adaptation, specifies RLOO/max-sharpening, and compares best valid solutions across computational discovery tasks.
Scope and limitations
Proposer training and evaluator access must be included in cost; the executor remains frozen, and improved verifier scores require independent solution-validity checks.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.