arXiv cs.CL· Xiangyu Peng, Jiaxin Zhang, An Yan, Yu Li, Ran Xu, Zeyuan Chen, Chien-Sheng Wu·· 12 小时前AI 评分41
MTA-Zero：通过在线 QA-Solver 协同训练实现自我改进的多模态深度搜索
MTA-Zero: Self-Improving Multimodal Deep Search via Online QA-Solver Co-training
AI 导读
MTA-Zero 是一个在线多模态 QA-Solver 协同训练框架，无需预计算多跳语料，即可让多模态深度搜索智能体自我改进。在六个多模态深度搜索基准上，它将 Qwen3-VL-8B Solver 的平均准确率从 36.91% 提升至 51.58%，在同一智能体工作流下性能与 GPT-5 相当；该框架还将 Qwen3.6-27B Solver 提升了 10.82 分。
正文
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.
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2610.33799 [cs.CL] |
| (or arXiv:2610.33799v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2610.33799 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Xiangyu Peng [view email]
[v1]
Sat, 24 Oct 2026 16:59:30 UTC (1,601 KB)
来源：arXiv cs.CL · arxiv.org