SeekVLN-C2PO-RFT

Official Counterfactual Contrastive Policy Optimization reinforcement-fine-tuned checkpoint for Seek Before You Move: Evidence Seeking for Progress Grounding in Vision-Language Navigation.

Paper: https://arxiv.org/abs/2609.37353

SFT initialization: https://huggingface.co/specialsuzuka/SeekVLN-FRG-SFT

Training dataset: https://huggingface.co/datasets/specialsuzuka/SeekVLN-Dataset

Collection: https://huggingface.co/collections/specialsuzuka/seekvln-6abcad5a8c398bb292e51252

This repository is being prepared for release and remains private. It contains the exported inference weights (language model, vision tower, multimodal projector) and tokenizer, not a resumable FSDP training checkpoint. Public inference instructions will follow with the code release. Licensing remains to be reviewed before public release.

The policy selects <nav> for direct navigation or <seek> to obtain left, front, and right observations before progress reasoning and navigation. This is the 20-update checkpoint used for the paper's main results, not the later 67-update or RxR-fine-tuned variants.

Evaluation

Paper-reported Val-Unseen results:

Dataset	NE	OSR (%)	SR (%)	SPL (%)	nDTW (%)
R2R-CE	3.7	75.2	67.5	61.4	-
RxR-CE	4.9	-	59.7	50.3	63.6

The archived evaluation records were checked during release preparation, but inference was not rerun. They cover 1,839 R2R episodes and 1,223 RxR guide-en-US episodes. R2R SR/SPL are 67.5367/61.4336; RxR SR/SPL are 59.6893/50.3374. The archived R2R OSR is 75.1495%, which rounds to 75.1 rather than the paper's 75.2; this small reporting discrepancy remains to be reconciled. RxR SPL counts the 33 broken-simulator-state episodes as zero, while NE excludes undefined geodesic distances.

Files

• llm/: language model weights, config, and tokenizer.
• vision_tower/: vision encoder weights and preprocessing config.
• mm_projector/: multimodal projector weights and config.
• Root config and tokenizer files preserve the original inference export layout.