Interactive Post-Training for Vision-Language-Action Models

Shuhan Tan, Kairan Dou, Yue Zhao, Philipp Krähenbüh

• Introduces RIPT-VLA, a scalable third-stage reinforcement learning method for VLA models, enhancing performance through interactive training with sparse binary rewards.
• Achieves SOTA performance across diverse benchmarks, including LIBERO-90 (94.3%), LIBERO-LONG 5-shot (71.4%), MetaWorld45 5-shot (76.0%), and OpenVLA-OFT (97.5%).
• Employs dynamic rollout sampling and leave-one-out advantage estimation to significantly enhance generalization, stability, and effectiveness across challenging tasks and scenarios.

EviNote-RAG: Enhancing RAG Models via Answer-Supportive Evidence Notes

Yuqin Dai*, Guoqing Wang*, Yuan Wang*, Kairan Dou, Kaichen Zhou, Zhanwei Zhang, Shuo Yang, Fei Tang, Jun Yin, Pengyu Zeng, Zhenzhe Ying, Can Yi, Changhua Meng, Yuchen Zhou, Yongliang Shen, Shuai Lu

• Introduces EviNote-RAG, an agentic RAG framework that restructures the pipeline into a retrieve–note–answer process. It trains LLMs to generate Supportive-Evidence Notes (SENs) and leverages an entailment-based Evidence Quality Reward (EQR) to improve evidence selection and reasoning.
• Achieves state-of-the-art performance across in-domain and out-of-domain QA benchmarks, with notable relative F1 gains: +20% on HotpotQA (+0.093), +40% on Bamboogle (+0.151), and +91% on 2Wiki (+0.256).
• Employs structured note-taking and reward-guided filtering to mitigate noisy retrievals, enhance training stability, and improve generalization and efficiency, significantly outperforming CoT, prompt-based, and RL-based baselines.

From Words to Worth: Newborn Article Impact Prediction with LLM

Penghai Zhao, Xinghua Xing, Kairan Dou, Jinyu Tian, Ying Tai, Jian Yang, Ming-Ming Cheng, Xiang Li

• Proposed the “Newborn Article Impact Prediction” (Newborn AIP) task and introduced the TNCSIsp metric, achieving an NDCG@20 score of 0.901.
• Constructed TKPD and NAID datasets, including over 12,000 samples for training and validation.
• Used LoRA to fine-tune and test 5+ large language models on server to evaluate prediction performance.

JobForce.AI

LLM-Powered Multi-Agent Platform for Job Seekers

Kairan Dou, Mingxi Tang , Yiya You, Yufeng Yan

• End-to-end LLM-powered pipeline automates both resume rewriting and job matching in a single workflow.
• Context-aware resume optimization tailors each experience using representative job descriptions and LLM-generated language.
• Semantic embedding–driven selection ranks and condenses content based on normalized cosine similarity for maximum relevance.
• Real-time, personalized job recommendations match optimized resumes against live postings filtered by user preferences.
• Quantitative alignment evaluation delivers a clear 100-point score and visual feedback on resume–JD fit improvements.