Research question
Can targeted post training improve difficult desktop agent behavior while preserving the capabilities that already work?
Abstract
Computer use agents can perform well on isolated actions yet fail over longer tasks. An ineffective click can trigger a loop. A dialog can interrupt progress. An export can appear complete without the requested file ever being saved. Specialization can also damage capabilities already present in the base model.
JxAgent studies this tradeoff using a failure focused multimodal training mixture and an explicit regression guard. The planned Run 1 mixture contains 95,503 samples from public computer use and grounding resources, with a bounded preservation replay cohort. LoRA training uses a frozen vision stack and a native interaction contract derived from the pinned model revision.
Base and adapted checkpoints are evaluated on identical task IDs, preprocessing, tool schemas, generation settings, and frozen preservation samples. Checkpoint selection considers computer use improvement, capability preservation, and serving cost together.
Scope
- Data selection, normalization, and contamination checks.
- Native model interfaces and a controlled training protocol.
- Paired evaluation, preservation checks, and checkpoint selection.
- Reproducibility requirements and planned ablations.
Status
The paper documents a research design and implementation. It reports no final model quality improvement. Controlled GPU training, paired checkpoint evaluation, and ablations remain the basis for future claims.
Citation
@techreport{pfaff2026jxagent,
author = {Julius Pfaff},
title = {JxAgent: Regression-Aware Post-Training for Reliable Long-Horizon Computer Use Agents},
year = {2026},
month = {August},
type = {Technical White Paper},
version = {0.1},
url = {https://github.com/JxAgentReal/JxAgent}
}