FACET: Terminal Training Tasks That Actually Run
The dirty secret of synthesized agent training data: a lot of the tasks are unsolvable, and the verifiers grade fiction. When instruction, solution and checker get generated in separate stages from inconsistent assumptions, you end up training agents on tasks no one could complete and scoring them with tests that check the wrong thing. FACET, from a USTC team, is squarely aimed at that — 112 upvotes on Hugging Face's daily papers board.
The method inverts the usual order. Instead of writing a task and hoping an environment exists for it, FACET reconstructs agent skills into coherent scenarios, then realizes and repairs the actual execution environment first. The live container state becomes the shared ground truth that instruction, reference solution and verifier are all derived from — with execution-based validation and targeted repair keeping them consistent. If it doesn't run, it doesn't ship.
The results back the premise: fine-tuning models across scales on FACET-generated tasks consistently improves Terminal-Bench 2.1 scores, and ablations confirm the load-bearing parts are exactly the environment-grounded construction and the solution-verifier alignment.
This is the second environment-side paper to top the boards in three days — Google's EnvHarness rewrites environments to teach agents what they're weak at; FACET makes the environment the single source of truth for training data. Same shift from two directions: the unit of agent data engineering is becoming the container, not the text.
Paper: https://arxiv.org/abs/2608.18580 project: https://stokou.github.io/FACET-Terminal/
← Back to all articles
The method inverts the usual order. Instead of writing a task and hoping an environment exists for it, FACET reconstructs agent skills into coherent scenarios, then realizes and repairs the actual execution environment first. The live container state becomes the shared ground truth that instruction, reference solution and verifier are all derived from — with execution-based validation and targeted repair keeping them consistent. If it doesn't run, it doesn't ship.
The results back the premise: fine-tuning models across scales on FACET-generated tasks consistently improves Terminal-Bench 2.1 scores, and ablations confirm the load-bearing parts are exactly the environment-grounded construction and the solution-verifier alignment.
This is the second environment-side paper to top the boards in three days — Google's EnvHarness rewrites environments to teach agents what they're weak at; FACET makes the environment the single source of truth for training data. Same shift from two directions: the unit of agent data engineering is becoming the container, not the text.
Paper: https://arxiv.org/abs/2608.18580 project: https://stokou.github.io/FACET-Terminal/
Comments