Yuhe Wu

Research profile

Research summary

Yuhe Wu is an early-career natural language processing researcher at The Hong Kong University of Science and Technology (Guangzhou). This profile is name-resolved from public OpenAlex bibliographic data and three first-author arXiv papers reviewed in the Evolve catalog. All three 2026 papers attack one problem from different angles: how large language models can be measured and deployed reliably. 'Count Evidence, Not Sentences' introduces tempered evidence fusion, which entropy-weights sentence-level log-odds of LLM judgments and improves long-text value measurement by roughly 4.5 accuracy points over naive voting. 'Caught in the Story' contributes a 5,078-scenario benchmark showing that one-sided multi-turn narration shifts LLM moral judgment by about 25 percentage points. 'BizSage' presents a self-evolving multi-agent framework using section-level knowledge-graph retrieval for business research automation. OpenAlex records 9 works over the last three years (1 in 2025, 8 in 2026), 24 distinct coauthors, and topic mass led by Topic Modeling (7) and Multimodal Machine Learning Applications (6). h-index and i10-index currently read zero, consistent with a short 2025-2026 publication window rather than measured influence. Evolve reproducibility scoring averages 43/100 across the three papers, with verdicts flagging weak baselines, synthetic data, and LLM-judge-only evaluation.

Research keywords

Highlights

• Tempered evidence fusion beats naive LLM voting by ~4.5 accuracy points

  First-author 2026 paper 'Count Evidence, Not Sentences' proposes TEF, which entropy-weights sentence-level log-odds of LLM judgments for long-text value measurement. Evolve calls it effective and simple, while noting that weak baselines and synthetic data temper confidence. Reproducibility readiness: medium.

• 5,078-scenario benchmark shows narration shifts LLM moral judgment ~25pp

  First-author 2026 paper 'Caught in the Story' quantifies narrative captivity in multi-turn LLM conversation: one-sided narration shifts LLM moral judgment by about 25 percentage points, with direct implications for LLM-as-a-judge reliability. Evolve finds the DPO attribution strains against Table 2. Readiness: medium.

• BizSage: self-evolving multi-agent framework for business research

  First-author 2026 framework combining section-level knowledge-graph retrieval (Personalized PageRank) with self-evolving LLM agents for business research automation, reporting zero-hallucinated-citation evidence. Evolve notes scoring relies on LLM judges only. Reproducibility readiness: low.

Timeline

1. 2025First recorded work appears in the OpenAlex countsByYear record (1 work in 2025), marking the start of the visible publication window.
2. 2026Publishes 'Count Evidence, Not Sentences' as first author on arXiv, proposing entropy-weighted evidence fusion for long-text value measurement (~4.5 accuracy points over naive voting).
3. 2026Publishes 'Caught in the Story' as first author, a 5,078-scenario benchmark showing that multi-turn one-sided narration shifts LLM moral judgment by about 25 percentage points.
4. 2026Releases BizSage as first author, a self-evolving multi-agent framework for business research with section-level knowledge-graph retrieval.
5. 2026Logs 8 works in 2026 versus 1 in 2025 (9 works over three years) with 24 distinct coauthors — the most productive stretch in the current OpenAlex window.

Papers in the Evolve catalog

• 2026-09-22first author

  Count Evidence, Not Sentences: Tempered Evidence Fusion of LLM Judgments for Long-Text Value Measurement

• 2026-09-03first author

  Caught in the Story: Narrative Captivity in Multi-turn LLMs Conversation

• 2026-09-03first author

  BizSage: A Self-Evolving Multi-Agent Framework for Business Research with Efficient Knowledge Retrieval

Frequent co-authors