Research
FACET Grounds Instruction, Solution and Verifier in One Container State to Stop Synthesizing Unsolvable Terminal Tasks
arXiv 2608.18580 (posted 2026-08-19, 74 upvotes on Hugging Face Daily Papers) argues that multi-stage terminal-task synthesis produces broken tasks because the instruction, environment, reference solution, and verifier get generated from inconsistent assumptions. FACET reconstructs related agent skills into information-rich scenarios, then realizes and repairs the execution environment first so the resulting container state is shared grounding for all downstream artifacts, with execution-based validation repairing only the specific artifact that failed rather than regenerating everything. Fine-tuning on trajectories from these tasks consistently improves Terminal-Bench 2.1 across model scales.
↳ Follow the thread