1. 研究背景与挑战
Multimodal large language models (MLLMs) have become strong visual reasoners, yet remain limited in open-world, multi-step search that requires grounding images, selecting tools, and integrating web evidence. Existing approaches typically construct and filter fixed offline corpora whose question distributions cannot adapt as the target deep-search agent improves.
2. 方法框架：MTA-Zero 在线协同训练
Building on MTA-Agent's image-grounded QA construction, we present MTA-Zero, an online multimodal QA-Solver co-training framework for self-improving multimodal deep-search agents that requires no precomputed multi-hop corpus.
在线 QA 生成
- Starting from image-grounded VQA seed contexts, a trainable QA Generator retrieves evidence and constructs verified multi-hop questions online.
深度搜索求解
- A separate deep-search Solver then selects tools, gathers evidence, and answers the generated questions.
3. Goldilocks 奖励机制
- To focus generation on the Solver's learning frontier, we derive a Goldilocks reward for the QA Generator from the Solver's group-wise rollout success rate.
- Questions solved by all rollouts or by none receive 0 reward, whereas questions near the current frontier are favored.
- We optimize the two policies independently with GRPO, without sharing parameters or gradients.
4. 实验结果与结论
- Across six multimodal deep-search benchmarks, MTA-Zero improves a Qwen3-VL-8B Solver from 36.91% to 51.58% mean accuracy and achieves performance comparable to GPT-5 under the same agent workflow.
- The same framework also improves a newer Qwen3.6-27B Solver by 10.82 points.
- These results show that verified online question generation provides a policy-specific curriculum for continually improving multimodal deep-search agents.