CHART: A Harness-Rotation Curriculum for Harness-Robust Search Agents
What this work does
Rotates a small active set of harness variants according to mastery and learnability rather than training on a fixed prompt or uniformly large prompt pool.
Connection to discovery
Suggests evaluating discovery skills across paraphrased and structurally varied harnesses instead of measuring them only under the training interface.
Evidence
Search experiments distinguish parallel-search behavior from answer success and include cue-free harnesses and an out-of-distribution shopping setting.
Scope and limitations
Parallelization is one behavior in one search family. Harness robustness is not equivalent to new scientific domains, and curriculum thresholds are design choices.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.