A systematic benchmark evaluating how far large language models are from real-world industrial telecommunication applications.

Proposes a dynamic policy optimization scheme that bridges SFT and RL post-training for robust LLM reasoning.

Augments retrieval-augmented generation with retrieved cell complexes to support topological reasoning over textual graphs.

Revisits multimodal RL post-training data sampling from a difficulty-distinguish perspective, with two complementary difficulty metrics (PISM & CMAB) and a hierarchical post-training framework.

A large-scale embodied planning model trained with multimodal post-training and Step-GRPO reinforcement learning, significantly outperforming RoboBrain2.0 on multimodal reasoning, spatial perception, and task planning benchmarks.

Introduces a curriculum reinforcement learning strategy from the reward-acquisition-difficulty perspective, enabling a Qwen2.5VL-3B model to outperform InternVL2.5-26B on multiple general benchmarks.

A knowledge-enhanced multi-intent transformer that integrates global heterogeneous information to model user intents while selectively filtering intent-irrelevant knowledge triples.

Disentangles user intents hierarchically for more accurate bundle recommendation.

Explores global information beyond the current session to improve session-based recommendation.

A multi-level interactive contrastive learning framework that enhances knowledge-aware recommendation with self-supervised signals.

Studies the sparse supervision problem in knowledge-aware recommender systems with a multi-level cross-view contrastive learning framework.

Disentangles latent user intents from multiple views with graph neural networks for bundle recommendation.