MTA-Zero: Self-Improving Multimodal Deep Search via Online QA-Solver Co-training
Xiangyu Peng, Jiaxin Zhang, An Yan, Yu Li, Ran Xu, Zeyuan Chen, Chien-Sheng Wu and colleagues (Salesforce AI Research) present MTA-Zero, which co-trains a question generator and a multimodal deep-search solver online so that training questions track what the solver can almost answer.
Ask this paper
Online data. Starting from image-grounded VQA seeds, a trainable QA Generator retrieves evidence and builds verified multi-hop questions online, with no precomputed multi-hop corpus.
Goldilocks reward. The generator is rewarded from the solver's group rollout success rate. Questions solved by all rollouts or by none earn zero, so generation concentrates on the solver's current frontier.
Training. Generator and solver are optimized separately with GRPO and share no parameters or gradients.
Results. Across six multimodal deep-search benchmarks, a Qwen3-VL-8B solver rises from 36.91% to 51.58% mean accuracy, comparable to GPT-5 in the same agent workflow. A Qwen3.6-27B solver gains 10.82 points.
Abstract
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. 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. Starting from image-grounded VQA seed contexts, a trainable QA Generator retrieves evidence and constructs verified multi-hop questions online. A separate deep-search Solver then selects tools, gathers evidence, and answers the generated questions. 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 zero reward, whereas questions near the current frontier are favored. We optimize the two policies independently with GRPO, without sharing parameters or gradients. 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.