acceptodds
Under review as a conference paper at ICLR 2027

Scaling Agentic Data for Long-Horizon Terminal Intelligence

Abstract

Terminal agents are moving beyond local operations toward completing real-world professional tasks. Short trajectories can demonstrate commands and tool calls, but they do not fully show how agents accumulate evidence through exploration, use it to inform judgments, continually adjust strategies based on feedback, and turn intermediate results into final deliverables. High-quality long-horizon trajectories capture these agentic behavioral patterns and their connections within a single problem-solving process, providing richer experience for sustained autonomous work. However, environment synthesis may complete or simplify the key work underlying these patterns in advance and thus remove the need for exploration, implementation, and verification that the agent was meant to perform. To address this, we introduce TerminalHorizon, a fully automated data synthesis engine for long-horizon terminal agents. Starting from real-world professional work, it uses independent execution to identify key work omitted during construction and uses the resulting evidence to guide environment reconstruction. To extend these patterns across stages, it progressively expands tasks under the same goal, introducing new requirements that build on prior results and call for further exploration and improvement. Using this engine, we construct 1.5K task environments and perform supervised fine-tuning of Qwen3.5-27B on the resulting dataset to obtain TerminalHorizon-27B. We achieve substantial performance gains on all 12 benchmarks spanning coding-agent tasks, agentic tasks in real-world production settings, and general reasoning. Multiple RSI case studies, each with a 12-hour research budget, further demonstrate the model's ability to continue experimenting beyond an initial solution, improve its methods based on feedback, and retain effective improvements in its final deliverables.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agent, Long-Horizon agent, Terminal agent, Coding agent, LLM

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_yem8OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper is trying to provide a new method for auto-generating terminal tasks, and then collect how the agents finish the tasks, and based on the trajectories to train the model. And the issue to fix is clear: the current environment synthesis may already give the key hints to the agents, so that the agents may skip exploring and verifying, which may lead to useless training data. That's why the authors propose TerminalHorizon, and there are 2 steps: 1. rebuild the environment, where agents can try to finish the task first and check if the key work has already been completed or not. 2. Extend the tasks, based on one goal, to add new requirements so that the later work can rely on previous results and give multi-stage solutions. Using this methodology, the authors generated 1500+ task envs, filtered 2800+ trajectories, and used them to fine-tune the Qwen3.5 model. For the experiment part, they tested the trained model based on the teminal operation, agent tasks and general reasoning tasks and demonstrate the validity of their methods.

Critical strengths
  1. This paper tries to solve a real issue and also provide a proper solution. For complex tasks, the agent may not need to complete them; the environment may already provide the results.
  2. The authors not only evaluate the model on terminal tasks, but also extend it to coding-agent tasks, agentic tasks in real-world production settings, and general reasoning, and show good performance improvement, which means their method works for more than one evaluation system.
Critical weaknesses

Lack of demonstration of how the pipeline components actually help improve performance. Although the entire pipeline works, I don't know how each part of the pipeline plays its part. For example, if we ablate the env reconstruction, will that impact the final results? Or if we ablate the extension, will the performance get better or worse? Their Individual contributions to downstream training are not clear.

Questions for authors

I like the general idea of this paper, and I don't have any other questions.

Reviewer_MZVtOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces TerminalHorizon, a pipeline for constructing long-horizon terminal environments and training trajectories. Its main components are challenge-guided reconstruction and progressive horizon expansion, designed to produce tasks requiring sustained planning, tool use, and iterative debugging. The authors construct 1,535 environments and collect 2,852 training trajectories. They evaluate supervised fine-tuning across several Qwen3.5 model sizes using terminal benchmarks, transfer evaluations, and recursive self-improvement experiments.

Critical strengths

A concrete pipeline addressing an important data bottleneck. Constructing executable terminal environments with challenging tasks and verifiable outcomes is costly. The combination of task reconstruction and progressive horizon expansion provides a systematic approach to generating such training data.

Consistent empirical benefits across model sizes and evaluation settings. The authors evaluate multiple scales within the Qwen3.5 family and include transfer and recursive self-improvement experiments. These results provide evidence that the generated trajectories have practical training value beyond simply increasing interaction length.

Critical weaknesses
  1. The data-efficiency claim is not sufficiently supported by token- or compute-matched comparisons. The paper emphasizes strong performance using relatively few trajectories, but its trajectories are longer. Consequently, trajectory count does not adequately characterize the amount of training supervision or computational cost. The current comparisons do not establish whether the gains reflect more effective data or greater supervision per trajectory. Reporting the total tokens contributing to the training loss and training costs, together with comparisons under matched training-token budgets, would substantially strengthen this claim.

  2. The experiments do not sufficiently disentangle the contributions of longer horizons, overall data quality, and teacher capability. The proposed approach changes task construction, trajectory length, filtering, and data provenance simultaneously. Better downstream performance therefore does not establish that extending the task horizon is a key source of improvement. The same-teacher trajectory-length comparison is informative, but it does not replace downstream training comparisons using the corresponding control datasets. Differences from existing datasets may also partly reflect teacher capability. Comparisons using the same teacher, seeds, and comparable generation budgets, followed by student training under matched token budgets, would help establish the pipeline’s independent contribution. A comparison between shorter and longer trajectories of comparable quality would further clarify whether longer horizons are necessary for the observed gains.

  3. Two additional limitations narrow the evidence. All evaluated students belong to the Qwen family, leaving applicability across model families untested. Moreover, although the pipeline includes quality controls, the paper provides insufficient independent validation of the model judge’s quality assessments, such as human agreement or an analysis of false acceptances and rejections.

Questions for authors

What are the total numbers of tokens contributing to the training loss and the training costs for each compared dataset? Does TerminalHorizon retain its advantage under matched training-token budgets?

Are there downstream training results using control datasets generated with the same teacher, seeds, and comparable generation budgets through direct rollout or an alternative construction process? Under comparable data quality and matched training-token budgets, do longer trajectories still outperform shorter ones? These results would materially affect my assessment of the proposed construction process.

Have the judge’s quality assessments been validated through human inspection or an independent evaluation? If so, please report the sampling procedure, agreement, and main error types.

Reviewer_yv6tOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TerminalHorizon is an automated framework for generating long-horizon terminal-agent training data. It uses independent execution to identify work prematurely completed during environment construction, reconstructs environments to restore that work, and progressively expands tasks with interdependent requirements. The authors construct 1,535 environments and 2,852 trajectories, reporting SFT improvements across 12 benchmarks and presenting case studies of sustained autonomous research.

Critical strengths
  1. The paper identifies a concrete failure mode: seemingly complex environments may already provide solutions to key subtasks. Its execution–analysis–reconstruction loop directly addresses this issue, with illustrative examples in the appendix.
  2. Gains extend beyond terminal benchmarks to productivity tasks and general reasoning, with additional validation on two other Qwen3.5 configurations. For example, Terminal-Bench 2.1 improves from 46.82% to 60.67%.
Critical weaknesses
  1. Unclear benefit over direct distillation. Since the student is trained on teacher-generated trajectories, an important missing baseline is to collect and filter trajectories directly from the original environments, without reconstruction or progressive expansion. How much additional improvement do the proposed mechanisms provide over this simpler approach? A comparison using the same teacher, student, and filtering procedure would help distinguish the contribution of the environment-construction methods from the benefits of teacher-trajectory distillation itself.
  2. Insufficient evidence of data efficiency. The comparison between approximately 3K and 32K trajectories is informative, but the proposed trajectories are substantially longer. Fewer demonstrations therefore do not necessarily imply fewer training tokens or lower computational costs. Reporting total training tokens, data-generation and training costs, and results under matched training-token budgets would help establish whether the method offers an efficiency advantage beyond reducing the number of trajectories.
Questions for authors
  1. Why not collect naturally complex tasks while preserving available tools? How does the method distinguish “efficient solutions that perform the required work” from “work already completed during construction”? Removing useful components could introduce unnecessary reimplementation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.