What it is
Referring-expression grounding that drafts a box in one pass, then verifies coordinate tokens in parallel to recover the model's greedy token box in fewer passes. Grounds text spans in images by predicting a one-pass draft box, then refining it with parallel token verification to return either the fast draft or the model's greedy coordinate-token box.
Key results
- Exact mode reaches the model's greedy token box in 5.60 passes avg
- Locus-2B exact runs at 102.2 ms vs 151.6 ms for vLLM token decoding
- Locus-0.8B typical gains +4.52 [email protected] over Jag-0.8B on same backbone
Shows that box regression can be used as a speculative draft to accelerate token-based visual grounding without changing the final greedy output.