Summary
Only the abstract was ingested: the full text is on OpenReview, which could not be fetched (see the note in the raw file). SAGE (Self-guided hierArchical exploration for Generalist wEb agents, NeurIPS 2025) targets the reliance of web-agent training on predefined task datasets and curated demonstrations. It explores at three levels: a pre-exploration phase that builds a structural understanding of a web environment, a top-level exploration that generates a self-evolving curriculum of tasks from easy to hard, and a low-level exploration that combines planning-based rollouts with step-wise learning. The authors present this as a scalable, supervision-free framework and report that it significantly outperforms prior approaches on WebVoyager and WebArena.
Key points
- No predefined tasks or demonstrations: the agent builds its own understanding of the site before generating tasks.
- Self-evolving curriculum: top-level exploration proposes tasks from easy to hard as the agent improves.
- Low-level exploration pairs planning-based rollouts with step-wise learning, aimed at policy efficiency on long tasks.
- Evaluated on WebVoyager and WebArena; the abstract reports significant gains over prior approaches but no numbers.
Relevance
- Web Agents — a supervision-free training recipe evaluated on WebArena.
- Environment Generation — curriculum by self-proposed tasks is a task-generation loop that adapts difficulty to the learner; its verification relies on the agent’s own judgement, which is the weak point we flagged.
- Not to be confused with SAGE, a different paper on cold-start user simulators.
Notes
From paper club, 13 February 2026:
With SAGE, the system creates an easy task, then increases task difficulty until the agent fails, then regresses to a simpler task before trying again. It’s like finding the Goldilocks zone for agent learning.
We’re honestly still a bit skeptical that LLMs-as-a-judge can verify complex state-changes as reliably as raw back-end data (especially for production-grade browser agents). Despite that, the core technique is too sound to ignore. By isolating exactly where a trajectory fails, we can stop throwing away “good enough” data that just needs a specific mid-course correction.