Junyou Su 苏军又

Peking University

PlanGPT Series (Large Models for Urban Planning)

Collaborative project at Peking University. PlanGPT 1.5 accepted by ACL Industry Track (Oral), and PlanGPT-VL accepted by EMNLP Industry Track.

• PlanGPT 1.0: Developed a customized large language model and an efficient retrieval system to assist urban planning.
• PlanGPT 1.5: Extended the planning capability boundary based on version 1.0.
• PlanGPT-VL: Proposed the PlanAnno-V framework for generating high-quality visual QA data in planning maps; introduced the Critical Point Thinking mechanism to reduce hallucination via structured verification.
• PlanBench: Built benchmarks evaluating contextual reasoning, policy understanding, spatial logic, and value assessment in planning tasks.
• PlanBench-V: Constructed benchmarks for national land-use planning map understanding.

InstructDiff (Data Selection)

Accepted by ACL 2026 Main Conference. Proposed InstructDiff, a domain-adaptive data selection framework based on contrastive entropy for efficient LLM fine-tuning.

• Introduced warmup calibration plus distribution-aware filtering to identify the most learnable training samples.
• Unified cognitive expansion/compression signals across reasoning and general instruction domains.
• Achieved strong gains over full-data training while using only a small selected subset.

ASFT: Anchored Supervised Fine-Tuning (Model Post-Training)

Post-training framework with lightweight KL anchoring for stable supervised fine-tuning; accepted by ICLR. Merged into LLaMA-Factory.

• Proposed the ASFT framework, integrating dynamic weighting with lightweight KL anchoring to mitigate training drift in Dynamic Fine-Tuning (DFT).
• Unified DFT under the Reward-Weighted Regression theoretical framework and proved ASFT has a tighter lower bound and more stable convergence than standard SFT.
• Demonstrated significant improvements across math reasoning, medical QA, and code generation with ~3% of full RL compute.
• Introduced ASFT-LoRA, reducing VRAM usage by 50% while retaining performance gains.

TAG-INSTRUCT (Data Synthesis)

Accepted by Findings of ACL 2025. Proposed TAG-INSTRUCT to enhance controllability of instruction complexity via structured semantic compression and difficulty-guided expansion.

• Compresses instructions into compact tag spaces; expands via RL-guided complexity augmentation.
• Shows superior controllability and stability across multiple benchmarks.

FANNO (Data Synthesis)

Published in Findings of ACL 2025. FANNO is an open-source, end-to-end framework for automated LLM instruction data synthesis.

• Leverages pre-filtering, seed generation, and UCB-based dynamic enhancement to synthesize high-quality data without proprietary models.
• Introduced the “Think Different” prompting strategy and UCB-based dynamic weighting mechanism for diversity and complexity.
• Generated 10K instruction data outperforming Alpaca-GPT4, WizardLM, etc., on AlpacaEval 2.0 and Arena-Hard.

ByteDance Seed

LocationMind Inc.

Core Work: Research on spatiotemporal analysis integrated with large language models.

Beijing Urban Planning and Design Institute

Served as technical lead, developing an RAG-based planning knowledge QA model.

• Built a large-scale knowledge base using OCR pipelines for unstructured planning documents.
• Designed a hierarchical retrieval architecture for precise and efficient RAG generation.
• Integrated speech input for improved user interaction and accessibility.

China Academy of Urban Planning and Design (Shenzhen Branch)

Served as technical lead, developing a planning knowledge QA model based on RAG.

• Constructed an institutional knowledge base from internal planning documents using LLM-driven sub-knowledge segmentation.
• Implemented a hierarchical retrieval system for high-precision RAG generation.
• Built a knowledge-based QA and intelligent review system integrating institutional and open-domain information.