Research map / Large language models / Agents, code and math
LaF-GRPO: In-Situ Navigation Instruction Generation for the Visually Impaired via GRPO with LLM-as-Follower Reward
Yi Zhao, Siqi Wang, Jing Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LLM-as-Follower reward to format+METEOR GRPO rewards (ablation) | SFT+GRPO Qwen2.5-VL-3B · NIG4VI intra-town (no pre-calc) | BLEU: 10.912 -> 10.921 | Improved |
| add LLM-as-Follower reward to format+METEOR GRPO rewards (ablation) | SFT+GRPO Qwen2.5-VL-3B · NIG4VI inter-town (no pre-calc) | BLEU: 10.076 -> 10.157 | Improved |
| add LLM-as-Follower reward to format+METEOR GRPO rewards (ablation) | SFT+GRPO Qwen2.5-VL-3B · NIG4VI intra-town | SPICE: 0.279 -> 0.274 (no pre-calc); 0.284 -> 0.286 (with pre-calc) | Mixed |
| add METEOR text reward on top of format reward (ablation) | SFT+GRPO Qwen2.5-VL-3B · NIG4VI intra-town (no pre-calc) | BLEU: 10.251 -> 10.912 | Improved |
| apply LaF-GRPO after SFT | Qwen2.5-VL-7B · NIG4VI intra-town (no pre-calc) | METEOR: 0.518 -> 0.545 | Improved |
| apply LaF-GRPO after SFT | Qwen2.5-VL-7B · NIG4VI inter-town (no pre-calc) | BLEU: 9.709 -> 9.002 | Hurt |
| apply LaF-GRPO after SFT | Qwen2.5-VL-3B · NIG4VI intra-town (no pre-calc) | BLEU: 9.099 -> 10.921 | Improved |
| apply LaF-GRPO directly without SFT | Qwen2.5-VL-7B · NIG4VI intra-town (no pre-calc) | METEOR: 0.211 -> 0.256 | Improved |
| increase training samples from 1k to 3k (ablation) | SFT+(LaF-GRPO) Qwen2.5-VL-7B · NIG4VI intra-town (no pre-calc) | METEOR: 0.529 -> 0.545 | Improved |
| increase training samples from 2k to 3k (ablation) | SFT+(LaF-GRPO) Qwen2.5-VL-7B · NIG4VI intra-town (with pre-calc) | METEOR: 0.550 -> 0.556 | No clear effect |
| use LaF reward instead of standard GRPO (LLM judge) (ablation) | Qwen2.5-VL · NIG4VI inter-town | navigational accuracy (%): 67.3% -> 68.1% | No clear effect |
| use LaF reward instead of standard GRPO (human study) (ablation) | Qwen2.5-VL · NIG4VI inter-town | preference rate: 24% -> 76% | Improved |
Closest papers by meaning
- Scene Map-based Prompt Tuning for Navigation Instruction Generation (CVPR 2025 · similarity 0.85)
- Towards Long-Horizon Vision-Language Navigation: Platform, Benchmark and Method (CVPR 2025 · similarity 0.84)
- Affordances-Oriented Planning Using Foundation Models for Continuous Vision-Language Navigation (AAAI 2025 · similarity 0.83)
- General Scene Adaptation for Vision-and-Language Navigation (ICLR 2025 · similarity 0.83)
- VLN-Video: Utilizing Driving Videos for Outdoor Vision-and-Language Navigation (AAAI 2024 · similarity 0.83)
- Lookahead Exploration with Neural Radiance Representation for Continuous Vision-Language Navigation (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
