📄 Summary: This paper introduces CV-18 NER, the first publicly available dataset for extracting named entities directly from Arabic speech, created by annotating the Arabic Common Voice 18 corpus with 21 fine-grained entity types. The work demonstrates that end-to-end speech NER models substantially outperform traditional cascaded pipelines (ASR followed by text NER), achieving 37.0% CoER on Arabic speech despite the language's morphological complexity and resource scarcity.
💡 Key Insight: End-to-end neural models can skip the error-prone intermediate step of transcribing speech to text, especially for morphologically complex languages like Arabic.
📄 Summary: This study compares how radiologists and LLMs evaluate machine-translated medical reports, testing whether LLMs can reliably assess translation quality of chest CT reports from English to Japanese. The research validates whether LLM-as-a-judge approaches can replace human expert evaluation for medical translation accuracy and readability.
💡 Key Insight: Using AI to evaluate other AI's translations of critical medical documents requires validation against human expert judgment.
📄 Summary: This paper addresses how LLMs struggle with position bias and the "lost-in-the-middle" problem when ranking candidates for recruitment, proposing an implicit strategy that better captures relationships among candidates while reducing token consumption. The approach moves beyond pointwise ranking paradigms to create more efficient and position-robust talent recommendations.
💡 Key Insight: LLMs rank candidates worse when they appear in the middle of a list, so systems must account for this structural bias in hiring applications.
📄 Summary: This paper proposes Neuro-RIT, a fine-tuning framework that makes retrieval-augmented LLMs robust to irrelevant or noisy retrieved contexts by identifying and aligning individual neurons that process relevant versus irrelevant information. Rather than updating entire layers or modules, it performs precision-driven neuron-level adaptation through attribution-based neuron mining.
💡 Key Insight: You can make AI systems more robust to bad information by surgically targeting individual neurons instead of broadly retraining entire layers.