Interlock: From Local Correctness to Task Completion in Long-Horizon Agents

Interlock. From Local Correctness to Task Completion in Long-Horizon Agents. By Linji Hao and eight others, from Peking University, the Hong Kong University of Science and Technology, Guangzhou, and two other institutions. Published on the fourteenth of October. No code repository is linked to it.

Large language model agents are increasingly asked to carry long-horizon tasks through complex environments, and those environments have a particular texture. Evidence arrives in fragments, states are left partially completed, and resources and artifacts depend on one another. The failure the authors point to is not an agent getting an individual step wrong. An agent can satisfy each requirement on its own and still fail to keep these relationships intact, producing an outcome that is locally correct and yet incomplete. Their diagnosis is that finishing a task depends on relationships: between the evidence in the environment, the decisions that carry consequences, and the effects those decisions are required to produce. The distance between satisfying the pieces and completing the whole is what they call the completion gap.

Interlock answers it on the data side. It is a framework for task synthesis, and it represents every task it creates as a typed completion graph. Three kinds of connection carry the relationships the diagnosis named. Grounding edges link public evidence to decisions. Coupling hyperedges impose joint constraints that span several components at once. Realization edges link each decision to all of the downstream effects it is required to bring about. A task built this way has its completion conditions written into its structure rather than left implicit.

The graph is then compiled into three things: a resettable workspace, a public request that the solver sees, and an evaluation rubric. Two reviews guard that compilation. A source-coverage review checks whether the original requirements survive generation. A public-input review checks whether the rubric is supported by the materials actually available to the solver, so that nothing is graded against information the solver could not reach. Last, an agent acts as judge. It inspects execution artifacts and runs verification scripts, and only trajectories judged to satisfy all mandatory requirements are retained.

The test is training utility. Two Qwen3.6 backbones are fine-tuned on their own accepted trajectories, then evaluated in external task environments. Across the two backbones, AutomationBench task pass rates rise by four to four and a half percentage points. JobBench weighted scores rise by between three point two and six point six four points. Harvey LAB criterion pass rates rise by between three point eight nine and four point one six percentage points.

Interlock's contribution is a way of making completeness explicit before any agent is trained. If incomplete outcomes come from lost relationships, then tasks that encode those relationships as edges, judged by an agent that demands every mandatory effect, yield experience in which finishing the whole job is what gets kept. Gains in three outside environments are the authors' evidence that this experience carries over.