Research map / Vision-language and multimodal / Multimodal LLMs
LocalAgent: Collaborative Agentic Verification for Fine-Grained Instance-Level Consistency
jiachen Guo, Xinshan Zhu, Boyun Wang, Yanyan Liang, Jun Wan, Feng Lin, Ajian Liu
Read the paperCheck your idea against 137,217 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.
Proposes LocalAgent, a collaborative agentic framework for fine-grained instance-level consistency verification that checks whether two images show the same physical object under varying viewpoint, illumination, and background.
Problem
MLLMs struggle with fine-grained instance-level consistency verification, deciding whether two images depict the exact same physical object under viewpoint, illumination, and background changes; the task needs both global instance understanding and precise local examination.
Approach
A collaborative agentic verification design that combines global understanding of the target instance with precise local examination of details to judge instance identity; details beyond the abstract are not stated in the provided text.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LEV tools with auxiliary-capped RFT on top of CoT-RFT | Qwen3-VL-4B · six-source mean accuracy | mean accuracy (%): 75.8 -> 80.7 | Improved |
| add LEV (LAC+LGC) instead of a crop tool (ablation) | Qwen3-VL-4B · six-source mean | mean accuracy (%): 75.64 -> 80.70 | Improved |
| combine LAC and LGC instead of using a single verifier (ablation) | LocalAgent · six-source mean | mean accuracy (%): 78.82 (LAC) / 79.94 (LGC) -> 80.70 | Improved |
| add weak behavioral prior labels to the tool reward (ablation) | LocalAgent · six-source mean | mean accuracy (%): 79.39 -> 80.70 | Improved |
| use auxiliary cap instead of hard gating (ablation) | LocalAgent · six-source mean | mean accuracy (%): 80.23 -> 80.70 | Improved |
| loosen auxiliary bounds to (0.10, 0.20) (ablation) | LocalAgent · six-source mean | mean accuracy (%): 80.7 -> 77.1 | Hurt |
| remove auxiliary reward on incorrect answers (bounds 0,0) (ablation) | LocalAgent · six-source mean | mean accuracy (%): 80.7 -> 79.5 | Hurt |
| learn selective tool invocation instead of always calling the verifier (ablation) | LocalAgent · inference latency | mean latency (ms/sample): 610.41 -> 336.86 | Improved |
| add LocalAgent selective verification on TWIN-Eval | Qwen3-VL-4B · TWIN-Eval total | accuracy (%): 82.9 -> 74.5 | Hurt |
| add LocalAgent on training-emphasis-aligned hard TWIN-Eval cases | Qwen3-VL-4B · TWIN-Eval hard cases subset | accuracy (%): 72.1 -> 80.9 | Improved |
| add LocalAgent on other TWIN-Eval cases | Qwen3-VL-4B · TWIN-Eval other cases subset | accuracy (%): 88.8 -> 69.2 | Hurt |
| crop to MLLM-localized regions before DINO or XFeat matching (ablation) | DINO and XFeat matching · six-source mean | mean accuracy (%): 74.23 -> 75.06 (DINO), 69.23 -> 71.11 (XFeat) | Improved |
Closest papers by meaning
- Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.90)
- VLM-R³: Region Recognition, Reasoning, and Refinement for Enhanced Multimodal Chain-of-Thought (NeurIPS 2025 · similarity 0.89)
- RegionReasoner: Region-Grounded Multi-Round Visual Reasoning (ICLR 2026 · similarity 0.89)
- Socratic-Geo: Synthetic Data Generation and Cross-Modal Geometric Reasoning via Multi-Agent Interaction (CVPR 2026 · similarity 0.88)
- ATLAS: Agentic or Latent Visual Reasoning? One Word is Enough for Both (NeurIPS 2026 · similarity 0.88)
- Locality Alignment Improves Vision-Language Models (ICLR 2025 · similarity 0.88)
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.
