The information filtering technology is usually used to track favorite topics and eliminate garbage content from information stream. The adaptive information filtering, which requires little initial training resource and can actively improve itself in filtering process, provides a better performance and convenience than the old way. But there are still some difficulties in training and adaptive learning. In this paper, an improved filtering model for adaptive text filtering is proposed. In this model, two retrieval/feedback mechanisms are used respectively. One is based on vector space model and Rocchio feedback algorithm, and another mechanism is derived from a latest language model IR system. Based on them, an incremental learning method using multi-step pseudo feedback is introduced in profile training to keep a minimal bias to the original topic, and an adaptive profile adjusting mechanism in filtering process, which newly takes into account the document distribution and the decay rate of the topic feature, is also developed. The running system constructed using the new model got a high evaluation score in related international contest, indicating that the improvements in the filtering model are effective.