Hunmin-CUA says a low-rank model transfer improved computer use without repeating a grounding weakness

A team presenting Hunmin-CUA says it transferred selected weights from a stronger computer-use model into its own 397-billion-parameter model, improving agent performance while avoiding a weakness in high-resolution screen grounding. The results make an interesting case for selective model transfer, though the team’s benchmarks are its own and several key results come from single runs.

Watch Desk analysis

What happened

In a community article published on Hugging Face, the Hunmin-CUA team describes post-training a vision-language model based on Qwen3.5-397B-A17B. Rather than copy the full difference between its base model and Qwen-CUA, the team used truncated singular value decomposition to transfer selected parts of that difference at reduced rank.

The authors say they ran the project on one machine with eight B200 GPUs. Their final comparison reports an OSWorld score of 70.45 for Hunmin-CUA, against 48.16 for the base and 77.12 for Qwen-CUA. On ScreenSpot-Pro, a benchmark for locating things in graphical interfaces, Hunmin-CUA scored 75.61, above both the base at 72.74 and Qwen-CUA at 62.20. The team also reports a Korean knowledge score of 81.43, close to the base’s 82.59 and above Qwen-CUA’s 78.07.

Key findings

• A selective transfer, not a full copy
  The team retained low-rank portions of selected model-weight differences, leaving other parts at their base values.
• Better computer-use scores, with a trade-off
  Hunmin-CUA scored 70.45 on OSWorld, below Qwen-CUA’s 77.12 but well above the base’s 48.16.
• Grounding held up in the reported comparison
  Hunmin-CUA scored 75.61 on ScreenSpot-Pro, compared with 62.20 for Qwen-CUA and 72.74 for the base.
• Not every result has the same evidential weight
  The authors report three-run means for ScreenSpot-Pro and Korean benchmarks, but single runs for OSWorld and WindowsAgentArena.

Why it matters

The practical point is that a model’s useful capabilities may not need to be transferred as an all-or-nothing package. In the team’s tests, selective transfer delivered much of a source model’s advantage on computer-use tasks without reproducing its reported high-resolution grounding deficit. That could matter for teams trying to specialise large models without paying for every improvement with losses elsewhere.

These are results from the authors’ own evaluations, not an independent replication. The OSWorld comparison is especially provisional: it is based on one run, and the authors note variation of about 2.8 points in separate repeat evaluations.

Our read

This is a thoughtful, technically specific approach to a real post-training problem, and the reported grounding result is the intriguing part. But the benchmark table is a promising result, not a verdict. The method deserves a closer look if replicated tests show the trade-off holds across more tasks, models and runs.

What to watch

• Whether the team publishes model weights, evaluation code or enough detail for others to reproduce the results.
• Whether repeated OSWorld runs confirm the reported improvement.
• Whether the grounding and Korean-language trade-offs persist on independent evaluations.

Discussion spark: If selective transfer improves an AI agent while preserving other abilities, what level of independent replication should be required before teams rely on it?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.