Jian Luan

Research profile

Research summary

Jian Luan is a researcher at Xiaomi (China) whose OpenAlex-resolved profile shows a sharp recent acceleration: 76 works within the last three years, rising from 2 indexed works in 2024 to 7 in 2025 and 67 in 2026. Reviewed output splits evenly across three directions — diffusion language models with reinforcement-learning post-training, GUI agents built on vision-language models, and long-term conversational memory with retrieval-augmented generation — each holding exactly one third of the field distribution. Evolve has reviewed three 2026 arXiv papers: Informed Masking, which applies rollout-derived mask biasing to RL post-training of diffusion LLMs across three frameworks on LLaDA-8B; a masked trajectory prediction method that unifies omni-dimensional GUI agent navigation with 2-3 point gains on five benchmarks; and MemLoc, a retrieve-localize-generate pipeline reporting state-of-the-art results on four memory-QA benchmarks. Luan is last author on two and middle author on one, a senior-position pattern consistent with supervising parallel project teams inside a 32-coauthor network with 12 frequent collaborators. OpenAlex topics widen the picture further, led by multimodal machine learning applications (36), topic modeling (21), GANs and image synthesis (19), plus speech recognition/synthesis (10) and speech and audio processing (8). Citation-linked impact data remain unusable (h-index 1, i10-index 0), most plausibly a name-disambiguation artifact rather than a true measure of influence.

Research keywords

Highlights

• Structure-aware masking for RL in diffusion LLMs

  Informed Masking introduces rollout-derived mask biasing that improves RL post-training across three frameworks on LLaDA-8B, with consistent gains that remain single-seed and lack confidence intervals.

• Masked trajectory prediction unifies GUI agent navigation

  The MaP approach trains vision-language agents for omni-dimensional GUI navigation via masked trajectory prediction, delivering 2-3 point gains over mixture training across five benchmarks, though without significance tests.

• MemLoc: retrieve-localize-generate for conversational memory QA

  MemLoc combines multi-granularity retrieval, personalized PageRank, and GRPO-trained evidence localization, reporting state-of-the-art results on four long-term memory-QA benchmarks while its reward design and baseline scope warrant scrutiny.

Timeline

1. 2026Reviewed arXiv paper 'Informed Masking' proposes structure-aware mask perturbation for RL in diffusion LLMs, evaluated on LLaDA-8B across three frameworks (last author).
2. 2026Reviewed arXiv paper on masked trajectory prediction unifies omni-dimensional GUI agent navigation, with 2-3 point gains over mixture training on five benchmarks (last author).
3. 2026Reviewed arXiv paper MemLoc presents a retrieve-localize-generate pipeline reporting state-of-the-art results on four long-term conversational memory QA benchmarks (middle author).
4. 2026OpenAlex-indexed output peaks at 67 works in a single year, coinciding with all three reviewed preprints — the visible high point of an accelerating 76-work three-year run.

Papers in the Evolve catalog

• 2026-09-22

  Informed Masking: Structure-Aware Perturbation for Reinforcement Learning in Diffusion Large Language Models

• 2026-09-22

  Towards Omni-dimensional GUI Agent Navigation with Masked Trajectory Prediction

• 2026-09-07

  Where to Look and What to Use: Retrieve-Localize-Generate for Long-Term Conversational Memory Question Answering

Frequent co-authors