Referring Expression Comprehension on RefCOCO+ (Accuracy, Average Tokens)
82AccuracyVisual Para-Thinker++
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Visual Para-Thinker++Backbone=Qwen2.5-VL-7B, Training Stage (MAC/RDMAO)=MAC+RDMAO2026.06 | 82 | 2,540 | |
| Visual Para-Thinker++Backbone=Qwen2.5-VL-7B, Training Stage (MAC/RDMAO)=MAC only2026.06 | 77.6 | 2,503 | |
| Base MLLM + Multi-Agent DebateBackbone=Qwen2.5-VL-7B2026.06 | 76.4 | 2,854 | |
| Base MLLM + Para-ThinkerBackbone=Qwen2.5-VL-7B, K (Number of Paths/Agents)=42026.06 | 76 | 2,456 | |
| Base MLLM + Self-ConsistencyBackbone=Qwen2.5-VL-7B, K (Number of Paths/Agents)=42026.06 | 75.1 | 2,410 | |
| Base MLLM + Long CoTBackbone=Qwen2.5-VL-7B2026.06 | 74.5 | 605 | |
| Base MLLMBackbone=Qwen2.5-VL-7B, Decoding Strategy=greedy2026.06 | 73.1 | 134 |