The authors first constructed executable tasks with verified reference solutions.
Co-evolving agent harnesses improve abstention without sacrificing task completion
HERA iteratively adapts an agent’s control system and its training environments, producing a harness that better distinguishes solvable tasks from those requiring abstention.
Top University
Han Luo · Bingbing Wen · Guang Yang · Zora Zhiruo Wang · Pan Lu · Lucy Lu Wang
University of Washington · University of Leeds · Carnegie Mellon University · Stanford University · Allen Institute for AI
Research Digest··3 min read
Luo and colleagues introduce HERA, a framework that generates matched feasible and infeasible tool-use tasks, then uses observed failures to alternately improve the agent harness and create new challenges.
Why this paper
From Allen Institute for AI and 4 others
In one line
Co-evolving an agent harness with adversarially mutated environments improves abstention, task completion, cross-model transfer, and cost efficiency in tool-using LLM agents.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ✓Reports numbers on named benchmarks (2 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§