Qwen

This research paper analyzes expert activation patterns in state-of-the-art Mixture-of-Experts (MoE) models to address inference bottlenecks in multi-node deployments. The authors propose a workload-aware micro-batch grouping and expert placement strategy to reduce inter-node communication overhead and improve latency.

The paper explores enhancing vision-language models through neuro-symbolic language-based reasoning using reinforcement learning. By utilizing Qwen3-VL-2B-Instruct, the researchers achieved higher accuracy and a significant reduction in reasoning tokens compared to SymPy.

Researchers investigate whether Reinforcement Learning from Verifiable Rewards (RLVR) actually produces reliable reasoning chains in language models. The study finds that standard RLVR does not guarantee that reasoning steps are causally important or sufficient, but suggests that adding auxiliary rewards can remedy this issue.

Researchers investigated the efficiency of chain-of-thought reasoning in LLMs, finding that models often decide on an answer long before finishing the generation. The study demonstrates that early stopping strategies can significantly reduce token usage and latency with minimal impact on accuracy.

Researchers propose a new reinforcement learning framework called polyGRPO that treats language as a latent variable to optimize reasoning. The method leverages multilingualism to improve the reasoning capabilities of models like Qwen2.5, showing significant accuracy gains in both English and multilingual benchmarks.

Researchers introduce POP, a self-play framework designed to improve LLM performance on open-ended tasks like creative writing and healthcare QA. The method uses the model itself to generate evaluation rubrics and input-output pairs, reducing the need for human-labeled data.

Alibaba's Qwen team has released the Qwen3.6-27B open-weight model, which claims flagship-level agentic coding performance. The new model is significantly more efficient, weighing 55.6GB compared to the 807GB of the previous generation's MoE model.

Researchers have introduced a two-dimensional early exit strategy that optimizes both layer-wise and sentence-wise processing for LLM inference. This method achieves significant computational savings and speed-ups for classification tasks across various open-source models like Llama and Gemma.

Researchers investigated why Large Language Models generate fake citations, finding that author names are particularly prone to hallucination. The study identifies specific 'hallucination neurons' in the Qwen2.5-32B-Instruct model and demonstrates that suppressing these neurons can improve citation accuracy.

Researchers introduce AlignCultura, a two-stage pipeline designed to improve the cultural alignment of Large Language Models. The method utilizes a new dataset, CULTURAX, to ensure models produce responses that are contextually aware and respectful of global cultural diversity.

Researchers have identified spectral phase transitions in the hidden activation spaces of large language models during reasoning versus factual recall. The study analyzes 11 models across 5 architectures to show how spectral properties can predict reasoning steps and correctness.

Researchers present causal evidence that hallucinations in autoregressive language models result from early trajectory commitments driven by asymmetric attractor dynamics. The study uses activation patching on Qwen2.5-1.5B to show that once a model commits to a hallucinated path, it is significantly harder to correct than it is to corrupt a correct path.

Researchers have identified a new vulnerability in Large Reasoning Models (LRMs) where harmful content can be injected into the step-by-step reasoning process without altering the final answer. The study introduces the PRJA framework, which uses semantic triggers and psychological framing to bypass safety alignment mechanisms.

Researchers have introduced Ecom-RLVE, a framework that extends verifiable environments to multi-turn e-commerce conversations. The system uses a curriculum of 8 verifiable environments to train models like Qwen 3 8B to better handle complex, tool-augmented shopping tasks and constraint satisfaction.

This article explores the evolution and future trajectory of China's open-source AI ecosystem following the DeepSeek R1 release. It examines how Chinese AI organizations are utilizing open-source models, papers, and infrastructure to drive large-scale global deployment.