A new AI reasoning system keeps the order of scientific evidence intact and the payoff shows up in the numbers.
PathAnchor is a bounded scientific reasoning system that retrieves evidence as Material-Sensor-Signal-System trajectories, preserving role, direction, and the source behind each step, rather than unordered passages or concept graphs. A controller with three read-only tools searches paper-specific trajectories, traces paths across sources, and opens exact evidence before writing a claim-cited answer with a stated evidence boundary. Tested on 120 single- and cross-paper questions about flexible sensors, it scored 82.6%, ahead of six other systems. Researchers also ran a controlled comparison: under the same controller, corpus, and six-call budget, swapping unordered concept graphs for path-structured records pushed source recall from 61.3% to 82.9% and lifted fully-cited answers from 69.2% to 90.0%, while using fewer tool calls.
The result is a reminder that retrieval-augmented AI agents don't just need more data, they need it kept in order. Losing the sequence of material to sensor to signal to system is exactly the kind of error that makes an AI-generated literature summary quietly wrong, not just incomplete. For anyone building AI tools to speed up scientific review, that's a bigger lever than throwing a bigger model at the problem.
It's a narrow benchmark on flexible-sensor papers, so treat the numbers as a proof of concept, not a verdict on scientific AI agents generally.