2,852 Trajectories Beat 32,000: How TerminalHorizon Makes Long-Horizon Agent Data Count
The standard practice for training terminal agents is to stack quantity: tens of thousands of trajectories, each with a dozen-plus interaction turns, teaching the model to type commands and call tools. Last week, a team from USTC and Alibaba DAMO Academy delivered a different answer: 2,852 trajectories, each averaging 52.8 interaction turns. The fine-tuned model scored 55.8% on Terminal-Bench 2.0 — three percentage points higher than Terminal-Universe, the representative of the conventional approach, which used 32,000 trajectories, more than ten times as many.
The few beating the many comes down to what each trajectory contains.
Short Trajectories Teach Actions, Not Problem-Solving
A short trajectory demonstrates "how to type a command": given some output, what command to run. But real terminal tasks aren't chains of commands — they're problem-solving processes: first explore to understand the environment, then diagnose where the problem is, then implement, then verify. Along the way, plans must be revised based on feedback, and findings from one step must inform judgments in the next.
The paper did foundational work: it decomposed and enumerated the capabilities needed in long-horizon tasks, yielding five major behavioral families and twenty-seven concrete patterns. For example, "relation tracking" — following chains of associations between objects and using what's found to support later decisions; or "feedback-driven iterative correction" — information exposed by a failed round changing strategy in the next. These patterns simply never get the chance to appear fully in a dozen-turn short trajectory.
The real problem lies on the data synthesis side. To create long trajectories, you must first create task environments — and if an environment designer slips, the critical work of the task gets done in advance. If a verification script directly provides the checking method for the correct answer, the exploration step is skipped. If the environment pre-loads simplifying assumptions, diagnosis becomes unnecessary. The task remains, but the problem-solving process is hollowed out. Train on that, and the model learns to go through the motions, not to solve problems.
The paper's solution is a dual-loop. The outer loop manages breadth: examining coverage gaps in already-generated tasks and deliberately adding new specialized directions. The inner loop manages depth: an independent third-party agent executes candidate tasks, specifically hunting for pre-simplified parts of the environment — what was done ahead of time, what was skipped — then reconstructing those parts, returning the work to the agent. Every stage has clear responsibilities and acceptance criteria.
A Twelve-Hour Marathon
The most persuasive experiment comes last. The team gave the model a twelve-hour research budget to autonomously complete a real task: multi-view point cloud fusion.
The base model's performance is dark comedy: it ran for 12.1 minutes, designed a proxy metric for itself, tested it — the score looked good — so it rated itself excellent, submitted voluntarily, and called it done. What it didn't know: a proxy metric aligned to only one reference view doesn't mean the full scene was reconstructed. Surfaces visible only from other views were all lost.
The fine-tuned model ran the full course. It didn't settle for its initial plan; it went back to check what each view contributed, introduced a "leave-one-view-out" diagnostic — removing one view to see how much reconstruction quality dropped, exposing under-covered regions — then improved cross-view alignment accordingly, performed global refinement, and filtered points by confidence. Its hidden evaluation score beat the base model by 27.03%. Same twelve hours: one was "satisfied" after twelve minutes; the other used the time.
There's a point in this comparison worth highlighting on its own: what the model lacks isn't time budget, but the patience to use the time. Models trained on short-trajectory data learn to "submit as fast as possible."
Two Dimensions of Data
Placing this paper alongside others in the field reveals a bigger picture. Earlier today, ACE-Brain-0.6 used one million samples to stack coverage breadth; its own conclusion was that "adding more same-type data gains less than a point; new scenarios are what's valuable." This paper, TerminalHorizon, works the other dimension: not stacking quantity, but making each trajectory long and complete. One horizontal, one vertical — together they complete the picture of data efficiency: breadth expands the frontier, depth preserves the process.
Filling in the numbers: Terminal-Bench 1.0/2.0/2.1 scores are 55.0/55.8/60.67, improvements of 8.7/14.2/13.85 over the base model (46.3/41.6, authors' reproduced values). On the real productivity side, AutomationBench pass rate rose 11.7% and partial score rose 23.3%; VitaBench rose 11.3%. Across twelve benchmarks covering coding, productivity, and general reasoning, all results are positive. The dataset is called TerminalHorizon-3K: 2,852 trajectories, median feedback chain of 25 turns, with roughly 90% of trajectories approaching 100 turns.
One honest detail worth noting: the base model's scores were reproduced by the authors themselves (marked with a dagger in the tables), not copied from official reports — the kind of number-printing rigor that reproduction-focused work should have.
---
*Source: arXiv 2610.13166 (2026-10-09), "Scaling Agentic Data for Long-Horizon Terminal Intelligence"; authors Shun Zou et al., University of Science and Technology of China, Alibaba DAMO Academy, Southern University of Science and Technology; base model Qwen3.5-27B; baseline Terminal-Universe-27B (Wu et al. 2026, 32.0k trajectories); twelve benchmarks including Terminal-Bench 1.0/2.0/2.1, AutomationBench, VitaBench, and others.*