Haifeng Wang · page 2
Papers
- DuReader_robust: A Chinese Dataset Towards Evaluating Robustness and Generalization of Machine Reading Comprehension in Real-World Applications (2020 · 13 citations)
Neural machine reading comprehension (MRC) models often fail in real-world use. They may answer paraphrased questions inconsistently (over-sensitivity), get distracted by passage spans that share many words with the question (over-stability), or struggle outside their training domain (generalization). Earlier… - ERNIE 5.0 Technical Report (2026 · 7 citations)
Most large multimodal systems can read images, video or audio but still answer mainly in text. When generation is added, it usually comes from separate decoders attached to a language model, which can weaken core language skills. Baidu's ERNIE 5.0 addresses this with a single autoregressive model trained from scratch… - BeamLoRA: Beam-Constraint Low-Rank Adaptation (2025 · 6 citations)
Low-Rank Adaptation (LoRA) is an efficient way to fine-tune large language models, but its accuracy still has room to improve. - Reinforced Informativeness Optimization for Long-Form Retrieval-Augmented Generation (2026 · 2 citations)
Reinforcement learning has improved LLMs on math and short-answer QA, where outputs can be checked against a single correct target. - Distributional Clarity: The Hidden Driver of RL-Friendliness in Large Language Models (2026)
Under identical RL with verifiable rewards (RLVR) pipelines such as GRPO and DAPO, some model families (Qwen) improve markedly while others (Llama) gain little. - PEC-Home: Interpretation of Progressively Elliptical Commands in Smart Homes (2026)
Smart-home assistants built on LLMs usually expect each command to be fully specified. In real households, people who interact repeatedly tend to drop details they assume are already shared, so their requests get shorter and more implicit over time. The authors frame this as a new task: interpreting progressively… - RRAtention: Dynamic Block Sparse Attention via Per-Head Round-Robin Shifts for Long-Context Inference (2026)
Attention cost grows quadratically with sequence length, which makes long-context prefill expensive. Dynamic sparse attention computes only the query–key blocks that matter for a given input, but each existing method gives something up. SeerAttention and MInference need offline training or pattern search. MInference…
This page in Markdown → https://md.research.elamur.ai/free/en/authors/haifeng-wang/page/2/