huggingface.co web signal

ZeroVLN Hits 51.3% on SAGE-3D With VLM Brain Frozen

TL;DR

  • ZeroVLN reaches 51.3% success on SAGE-3D and 52.7% on R2R-CE, beating NaVILA's 46.0% and 46.5% on the same episode pool.
  • The VLM brain's weights shift by only 0.01497%, versus 3.21% for DualVLN and 7.67% for NaVILA under supervised fine-tuning.
  • General visual benchmarks hold at 59.16% mean accuracy against a 59.47% Qwen2.5-VL-7B baseline, while SFT baselines collapse to roughly 0%.

A new navigation paper on Hugging Face argues you can beat supervised-fine-tuned VLN systems without touching the vision-language model's weights at all. ZeroVLN reports 51.3% success on SAGE-3D, 52.7% on R2R-CE and 51.4% on RxR-CE, versus 46.0%, 46.5% and 48.7% for NaVILA on the same 34,445-episode training pool.

The trick is a slow–fast split. The VLM brain, Qwen3.6-35B-A3B in the authors' setup, is frozen and given a semantic tool-calling interface; a separate 12-layer flow-matching action expert carries out the movement requests. The abstract puts it this way: rather than teaching the brain to predict geometric goals, the system "equips it with a semantic tool-calling interface for conveying movement decisions to an action expert in natural language," then "jointly improves the brain's movement decisions and the action expert's continuous execution through closed-loop reinforcement learning."

The weight-preservation numbers are the sharper part of the claim. ZeroVLN reports only 0.01497% displacement in the brain's parameters, against 3.21% for DualVLN and 7.67% for NaVILA. On general visual benchmarks, ZeroVLN keeps 59.16% mean accuracy versus the 59.47% base rate, while the paper reports the SFT baselines score "0% (zero scores on most benchmarks)." Code is posted at github.com/hzx122/ZeroVLN.

The benchmark story is not uniformly flattering. ZeroVLN trails NaVILA on SPL (29.2% vs 41.8% on R2R; 28.1% vs 43.0% on RxR), meaning its successful runs take longer paths. The results are also simulator-only, with no physical-robot evaluation reported. The paper sits alongside a run of navigation and embodied-agent work on our radar this week, including Mind2Act's drop of the top agent to 12% on compositional tasks.