Hao Wang

Beyond Single-View Indexing: Structure-Aware Multi-View Retrieval for Knowledge-Based VQA

• Proposed a coarse-grained retrieval framework for knowledge-based visual question answering, using manifold diffusion and graph-similarity-based deduplication to leverage complementary information across knowledge base modalities and improve retrieval accuracy.
• Established the ideal upper bound of single-vector retrieval in multimodal settings; experiments show our method reaches 80–90% of this bound. Compared to SOTA single-vector coarse-grained retrieval, we improve recall by over 10% while adding less than 1% retrieval time.

Momentum-integrated Multi-task Stock Recommendation with Converge-based Optimization

• Addressed short-term noise and overfitting in stock prediction with a multi-task training framework using momentum labels and overfitting-aware gradient allocation.
• On A-share and US stock markets, the proposed method improves RNN prediction performance and significantly mitigates overfitting during training.

Work I — Systematic Study of Semantic Uncertainty and Hallucination Detection

• Proposed a systematic framework for semantic uncertainty methods in LLM hallucination detection, covering 19 datasets, 17 semantic uncertainty computation methods, and experiments with 13 LLMs for generation and hallucination detection.
• Found that current semantic uncertainty methods still struggle on complex semantic tasks such as summarization and translation. Explored blind spots in black-box LLM hallucination detection and constructed blind-spot datasets. Discovered a scaling law in LLM hallucination via uncertainty quantification: stronger models exhibit greater overconfidence.

Work II — RAG Uncertainty Measurement and Context Refinement

• Addressed RAG frameworks that rely solely on LLMs to judge context relevance by proposing RAG uncertainty measurement using perturbed data and semantic matrices, along with a two-stage RAG context refinement framework.
• Experiments show strong correlation between the proposed uncertainty measure and RAG answer accuracy; the refinement method significantly improves RAG context quality.

Work III — Financial Investment RAG Framework

• LLMs show conservative biases on finance and investment questions, limiting actionable advice for investors. Proposed a knowledge-graph and step-wise reasoning RAG framework with time-semantics hybrid weight sampling and decoupling final recommendations from the LLM to improve stock prediction performance.