Research map / Vision-language and multimodal / Multimodal LLMs

OVOD-Agent: A Markov-Bandit Framework for Proactive Visual Reasoning and Self-Evolving Detection

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 0 citations (OpenAlex)

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.

Proposes OVOD-Agent, an LLM-free Markov-Bandit framework that models visual-CoT refinement of open-vocabulary detector category prompts as a weak MDP, using UCB exploration and a dual-head reward-policy model, reporting gains on COCO and LVIS across four OVOD detectors.

Problem

Open-vocabulary object detectors rely on fixed category names at inference, making detection a static text-matching step; existing refinement methods are static or LLM-heavy, leaving rare-category recognition limited.

Approach

Treats visual-context refinement as a weakly Markovian decision process over eight state units, selects seven interpretable visual operators (dictionary, color, texture, background, geometry, lighting, spatial) with UCB bandit exploration, and distills the resulting trajectories and Markov transition priors into a lightweight reward-policy model that replaces the bandit at inference.

Key results

• On LVIS val, OVOD-Agent improves AP_r for GroundingDINO from 30.2 to 32.9 (+2.7) and for DINO-X Pro from 48.0 to 49.2 (+1.2).
• On COCO 2017 val, gains range from +0.6 to +1.3 mAP across the four base detectors.
• Adding the KL Markov regularization raises AP from 38.2 to 39.4 and AP_r from 19.0 to 20.3 on LVIS minival (ablation).

Surprising findings

• The LLM-free OVOD-Agent reaches 37.0 AP_r on LVIS minival-style comparison, exceeding DVDet's 36.2, despite using no LLM.

Limitations

• Degrades on non-canonical object appearances and sparse transition statistics for rare states.
• Struggles with tiny, occluded objects in clutter, where geometric and spatial actions give noisy rewards.
• Gains on COCO are modest (+0.6 to +1.3 mAP).

Experimental setup

Datasets: COCO LVIS

Benchmarks: LVIS val LVIS minival COCO 2017 val

Metrics: AP AP_r mAP latency Top-K@Stop Pareto-Win Rate

Baselines: GroundingDINO YOLO-World GroundingDINO 1.5 DINO-X Pro RALF CoT-PL DVDet LLMDet

Builds on: Chain-of-Thought Visual-CoT CoT-PL RALF

Closest papers by meaning

• Cyclic Contrastive Knowledge Transfer for Open-Vocabulary Object Detection (ICLR 2025 · similarity 0.87)
• Train the Agent, Not the Expert: Learning to Harness Heterogeneous Experts for Multi-Turn Visual Reasoning (NeurIPS 2026 · similarity 0.87)
• CAKE: Category Aware Knowledge Extraction for Open-Vocabulary Object Detection (AAAI 2025 · similarity 0.87)
• Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning (ICML 2026 · similarity 0.87)
• Open-Det: An Efficient Learning Framework for Open-Ended Detection (ICML 2025 · similarity 0.86)
• Learning Background Prompts to Discover Implicit Knowledge for Open Vocabulary Object Detection (CVPR 2024 · similarity 0.86)