- The problem
- When an AI assistant works on a long research question or a software repair, its running notes can fill up with old attempts and irrelevant details. That makes it harder to focus, and many systems cannot return cleanly to an earlier point without carrying all the later clutter along.
- The idea
- TreeFlex saves useful points in the work as a tree, like keeping several labeled bookmarks rather than one ever-growing notebook. The same model chooses which bookmark to reopen and what to do next; for example, it can try a different answer from an earlier point, then return to a promising line when new evidence helps. A separate training method teaches it these navigation and work decisions while respecting the distinct context it had at each step.
- What it got
- With the same working-context limit, TreeFlex scored 48.8% on a challenging research benchmark, compared with 41.5% for a standard linear-history assistant. On software repair, it resolved 54.9% of tasks, compared with 43.3% for the strongest reported FoldAgent comparison.