REVERSE** (Reinforcing Evidence Verification and Search for Agentic Image Geolocation)

Paper Link (arXiv:2605.26861)

The model learns to iteratively zoom into image regions, search with image crops, and query text search engines — then synthesizes evidence to predict GPS coordinates.

*(A) Agentic geo-localization fails when a model crops the wrong region and follows misleading results, or retrieves the right candidate but cannot filter distractors. (B) REVERSE fixes both: spatial [...]

---

Overview

REVERSE/
├── eval/               # Evaluation pipeline (Im2GPS3k, YFCC4k)
├── data_pipeline/      # Data processing: SFT coldstart + RL parquet builders
├── script/
│   ├── sft/            # SFT training scripts
│   └── rl/             # RL training scripts (GRPO, configs, dist launcher)
└── verl/               # Full verl with modifications for relative image paths
                        # and offline search cache support

---

Environment

We use the official verl Docker image:

docker pull verlai/verl:sgl056.latest

Install verl from this repo:

cd verl
pip install --no-deps --no-build-isolation -e .
pip install --no-deps fastmcp

---

Dataset

The dataset is hosted on HuggingFace: Jack11211/REVERSE

Download and set DATA_ROOT:

export DATA_ROOT=/path/to/REVERSE_data

Expected structure:

$DATA_ROOT/
├── images/          # All training images (relative path root)
├── sft/             # SFT coldstart parquet files
│   ├── train_sft_coldstart_v5.parquet
│   └── val_sft_coldstart_v5.parquet
├── rl/              # RL training parquet files
│   ├── train_rl_v5_easy.parquet
│   ├── train_rl_v5_full.parquet
│   └── val_rl.parquet
└── rl_cache/        # Offline search cache (image_search + text_search)
    ├── image_search_cache.parquet
    └── text_search_cache.parquet

---

Training

REVERSE training pipeline: direct SFT → cold-start multi-turn SFT → agentic RL with offline search cache and process-level rewards.

Data pipeline: Kimi-K2.6 generates trajectories over MP-16 Pro images, filtered for quality, annotated with evidence labels and corrected bounding boxes, then split into easy/full curriculum.

Stage 1: SFT on MP-16 Pro (~4M images)

Geographic knowledge pretraining on the MP-16 Pro dataset.

DATA_ROOT=/path/to/REVERSE_data \
bash script/sft/dist/run_sft_dist.sh

Stage 2: SFT Coldstart (multi-turn tool-use)

Trains the model on cold-start trajectories with zoom, image search, and text search tools.

DATA_ROOT=/path/to/REVERSE_data \
bash script/sft/dist/run_sft_coldstart.sh

Stage 3a: RL Training (Easy curriculum)

DATA_ROOT=/path/to/REVERSE_data \
TRAIN_FILE=train_rl_v5_easy.parquet \
TRAIN_SCRIPT=/path/to/REVERSE/script/rl/geoloc/run_qwen3vl-4b_geoloc_coldstart_v5_dist.sh \
bash script/rl/dist/run_dist.sh

Stage 3b: RL Training (Full curriculum)

DATA_ROOT=/path/to/REVERSE_data \
TRAIN_FILE=train_rl_v5_full.parquet \
MODEL_PATH=/path/to/stage3a_checkpoint/huggingface \
TRAIN_SCRIPT=/path/to/REVERSE/script/rl/geoloc/run_qwen3vl-4b_geoloc_coldstart_v5_dist.sh \
bash script/rl/dist/run_dist.sh

---

Evaluation

cd eval/
python eval_benchmark.py \
    --model_path /path/to/checkpoint \
    --data_root $DATA_ROOT

*Comparison of two agent trajectories on the same case. Qwen3-VL-4B-Instruct retrieves strong evidence but rejects it and follows a distractor. REVERSE correctly identifies the sculpture and predicts [...]

---

Environment Variables

Variable	Description
DATA_ROOT	Path to dataset root (images + parquet files)
WANDB_API_KEY	Weights & Biases API key
OXYLABS_USER	Oxylabs username (for live image search)
OXYLABS_PASS	Oxylabs password
TAVILY_API_KEY	Tavily API key (optional, for live text search)
COS_SECRET_ID	Tencent COS secret ID (optional, for image upload)
COS_SECRET_KEY	Tencent COS secret key (optional, for image upload)

For offline training with precomputed cache, only DATA_ROOT and WANDB_API_KEY are required.

---

Citation

If you find our work useful, please cite:

@misc{li2026reverse,
  title   = {REVERSE: Reinforcing Evidence Verification and Search for Agentic Image geo-localization},
  author  = {Yong Li and Furong Jia and Dacheng Yin and Kang Rong and Fengyun Rao and Jing Lyu and Fan Zhang},
  year    = {2026},
  archivePrefix = {arXiv},
  eprint  = {2605.26861}
}