Research map / Vision-language and multimodal / Human and embodied modeling

REVERSE: Reinforcing Evidence Verification and Search for Agentic Image Geolocation

NeurIPS 2026 · Vision-language and multimodal › Human and embodied modeling · approach: Scene graph generation

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Trains a 4B Qwen3-VL agent with zoom, image search, and text search tools for image geo-localization, using SFT, agentic cold start, and GRPO with process rewards for crop IoU, query utility, and MCC-based evidence discrimination over an offline search cache.

Problem

Agentic image geo-localization often fails because models crop the wrong region, query poorly, or trust distractor search results; existing methods supervise only final coordinates.

Approach

Make three intermediate decisions (where to look, how to query, what to trust) observable via zoom boxes, search-result usefulness tags, and teacher-annotated labels, then supervise them with process rewards (IoU, search utility, MCC) during GRPO against an offline search cache.

Key results

• REVERSE (4B) reaches 48.3% at 25km and 22.5% at 1km on Im2GPS3k, above Geo-R's 41.5% at 25km and untuned Qwen3-VL-8B's 37.2%.
• On YFCC4k, REVERSE reaches 27.5% at 25km, below G3 at 35.9%.
• Full-curriculum RL raises 750km accuracy from 68.4% to 73.5% on Im2GPS3k in the training ablation.

Surprising findings

• Equipping the untuned Qwen3-VL-4B with tools without training slightly raises 25km accuracy but lowers 750km accuracy (62.6% to 60.9%).
• Removing the IoU reward leaves 25km accuracy unchanged, suggesting the MCC signal already guides crop selection.

Limitations

• Offline static search cache means the model learns against a fixed snapshot rather than live web results.
• Cold-start trajectories depend on a large proprietary teacher (Kimi-K2.6).
• Image search does not filter GPS-exposing domains such as Flickr, risking coordinate leakage; YFCC4k performance trails retrieval baselines.

Experimental setup

Datasets: MP16-Pro Im2GPS3k YFCC4k

Benchmarks: Im2GPS3k YFCC4k

Metrics: accuracy at 1km 25km 200km 750km 2500km

Baselines: kNN PlaNet CPlaNet ISNs GeoToken GeoCLIP Img2Loc PIGEON G3 Translocator GeoDecoder GeoBayes

Builds on: GRPO Qwen3-VL Geo-R G3

Closest papers by meaning

• Locate-Then-Examine: Grounded Region Reasoning Improves Detection of AI-Generated Images (CVPR 2026 · similarity 0.85)
• Vision-Language Reasoning for Geolocalization: A Reinforcement Learning Approach (AAAI 2026 · similarity 0.85)
• Asking like Socrates: Socrates helps VLMs understand remote sensing images (CVPR 2026 · similarity 0.85)
• OpenSearch-VL: An Open Recipe for Frontier Multimodal Search Agents (NeurIPS 2026 · similarity 0.84)
• DeepScan: A Training-Free Framework for Visually Grounded Reasoning in Large Vision-Language Models (CVPR 2026 · similarity 0.84)
• Perceive, Interact, Reason: Building Tool-Augmented Visual Agents for Spatial Reasoning (NeurIPS 2026 · similarity 0.84)