BAAI attributes frontier agent failures on research tasks to retrieval drift and search polarization

BAAI reports that its fine-grained error analysis of frontier agent trajectories finds two main failure patterns.

On Deep Research tasks, Retrieval Drift accounts for 32% to 41% of failures: agents reach the right research area but fix on cited baseline papers or adjacent works instead of meeting every constraint. On Wide Research tasks, the failures split by model. Claude-Opus-4.6 shows 85.3% Unconstrained Expansion, collecting many unverified papers at the cost of precision, while Seed-2.0-Pro shows 33.3% Premature Termination, stopping after retrieving only a few candidates. BAAI concludes that agents lack stateful evidence tracking and a precise sense of scientific concept boundaries.