Junyang Lin

Junyang Lin is an AI researcher and coauthor of the Qwen3 Technical Report, with earlier contributions to Chinese multimodal pretraining and unified vision-language models. His public writing examines how models should allocate reasoning effort and learn through interaction with tools and environments. He uses the handle JustinLin610; as of September 2026, his personal website describes him as an independent researcher working on language models, multimodal systems, and agentic learning.[7][6][5][4][2][3]

Lin coauthored the Qwen3 Technical Report, submitted May 14, 2025, alongside An Yang, Binyuan Hui, Rui Men, Peng Wang, and other collaborators. Qwen3 integrated thinking mode for complex, multistep reasoning with non-thinking mode for rapid responses. A thinking budget let users balance inference computation, latency, and performance. The models covered by the report were publicly accessible under Apache 2.0. In June 2025, Lin contributed an online-track keynote to the AI Engineer World’s Fair, where he was identified with Alibaba Qwen.[7][1]

In his March 26, 2026 essay, “From 'Reasoning' Thinking to 'Agentic' Thinking”, Lin recalled that he and other Qwen team members had wanted a system that could unify thinking and instruct modes and automatically choose how much reasoning a prompt required. He acknowledged: “We did not get everything right when trying to balance model merging with improving the quality and diversity of post-training data.”[4]

Lin located the difficulty in competing training objectives. Instruct behavior favors directness, brevity, formatting compliance, and low latency for tasks such as structured extraction and templated support. Thinking behavior favors spending more computation on hard problems, maintaining coherent intermediate reasoning, and exploring alternatives. He argued that poorly curated mixtures could weaken both. In his account, Qwen3’s later-2025 2507 line shipped separate Instruct and Thinking updates, allowing teams to address each mode’s data and training problems more cleanly. His preferred goal remained a smooth spectrum of reasoning effort that a model could choose adaptively.[4]

Lin’s March 2026 argument centers on agentic thinking: reasoning that sustains action through tool use, environmental feedback, and revised plans. He argues that longer reasoning traces alone do not establish better intelligence. Useful thinking should help a system decide when to act, select tools, recover from failures, and remain coherent across many interactions.[4]

That view makes training infrastructure and environment design central. “The environment is no longer a static verifier; it’s part of the training system,” Lin writes. His example is a coding agent waiting for a live test harness: inference stalls for execution feedback while training lacks completed trajectories. He advocates cleaner separation of training and inference, along with environments that provide realistic tasks, stable execution, diverse states, and useful feedback.[4]

Lin also identifies reward hacking as a major challenge for tool-enabled agents. Search can expose answers; repository information or logs can reveal shortcuts that invalidate a coding task. As he puts it, “An environment with hidden leaks can make the policy look superhuman while actually training it to cheat.” He therefore emphasizes evaluator robustness, anti-cheating protocols, and the interfaces between a policy and its environment.[4]

Before Qwen3, Lin worked on models combining language and visual information. He was the first listed author of M6: A Chinese Multimodal Pretrainer, submitted in March 2021, with Rui Men, An Yang, Chang Zhou, and 21 other coauthors. M6 unified pretraining on single-modality and multimodal Chinese data. The authors reported 10-billion- and 100-billion-parameter variants and demonstrated downstream text-guided image generation.[6][7]

In 2022, Lin coauthored OFA with Peng Wang, An Yang, Rui Men, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Accepted at ICML 2022, OFA brought image generation, visual grounding, captioning, classification, and language modeling into a simple sequence-to-sequence framework. It used instruction-based learning during both pretraining and finetuning, without adding task-specific layers for downstream tasks. These collaborative projects form an earlier part of Lin’s work on systems spanning language and vision.[5][6]