# Selecting Long-Horizon Trajectories for Reliable and Efficient Terminal-Agent Training

arXiv: https://arxiv.org/abs/2610.05831 · HTML: https://mlsanity.com/paper/2610.05831/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 12 of its 15 references that are in the map.
- First posted on arXiv 5 Oct 2026.

## Builds on (most cited first, 12 of 12)

- [LIMA: Less Is More for Alignment](https://mlsanity.com/paper/2305.11206/) (2023)
- [DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models](https://mlsanity.com/paper/2512.02556/) (2025)
- [Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces](https://mlsanity.com/paper/2601.11868/) (2026)
- [From Quantity to Quality: Boosting LLM Performance with Self-Guided Data Selection for Instruction Tuning](https://mlsanity.com/paper/2308.12032/) (2023)
- [Perplexed by Perplexity: Perplexity-Based Data Pruning With Small Reference Models](https://mlsanity.com/paper/2405.20541/) (2024)
- [Endless Terminals: Scaling RL Environments for Terminal Agents](https://mlsanity.com/paper/2601.16443/) (2026)
- [The Best Instruction-Tuning Data are Those That Fit](https://mlsanity.com/paper/2502.04194/) (2025)
- [On Data Engineering for Scaling LLM Terminal Capabilities](https://mlsanity.com/paper/2602.21193/) (2026)
- [Large-Scale Terminal Agentic Trajectory Generation from Dockerized Environments](https://mlsanity.com/paper/2602.01244/) (2026)
- [Terminal-World: Scaling Terminal-Agent Environments via Agent Skills](https://mlsanity.com/paper/2605.20876/) (2026)
- [ECHO: Terminal Agents Learn World Models for Free](https://mlsanity.com/paper/2605.24517/) (2026)
- [I Learn Better If You Speak My Language: Understanding the Superior Performance of Fine-Tuning Large Language Models with LLM-Generated Responses](https://mlsanity.com/paper/2402.11192/) (2024)

## Abstract (arXiv)

Terminal agents are commonly trained by imitating long teacher trajectories, yet how much of each trajectory to supervise remains unexplored. We study the supervision horizon, the number of trajectory tokens retained for training, and show that it is a key design axis for reliability and cost. Reliability improves with longer horizons but saturates: on Terminal-Bench, a 12K-token horizon solves more tasks than 16K (29±0.7 vs. 26±0.8) while requiring 30% less training time. The horizon also shapes agent behavior: short horizons cause premature termination, intermediate horizons yield productive error recovery, and long horizons induce over-persistence. We analyze this saturation through a bias--complexity bound, in which longer supervision reduces temporal supervision bias but increases finite-sample estimation error from more heterogeneous late-stage histories. Guided by this analysis, we propose selective long-horizon refinement, which first trains on short prefixes and then refines only on continuations that are most likely under the warm-start model. It consistently outperforms full long-horizon training. At 16K, it raises successful attempts from 110±2.7 to 126±2.1 and tasks solved in at least six of eight attempts from 9±0.7 to 14±0.6; with half of the long-horizon data, it still reaches 122±2.4 while cutting training time by 23%. The gains transfer across benchmarks, from 64±2.6 to 73±2.1 on Terminal-Bench v2.0 and from 137±2.7 to 155±2.2 on OpenThoughts-TBLite. For long-horizon supervision, selecting the right trajectories matters more than training on all of them.