Efforts to Improve LLM Accuracy (March 2026)

English summary

This presentation describes approaches for improving the accuracy and reliability of large language models in research. Because language models predict the probability of the next token from preceding tokens rather than relying on human-like semantic understanding, they can produce fluent but factually incorrect outputs. The presentation focuses on retrieval-augmented generation, stepwise and multi-path reasoning, domain-specific models, and agent or multi-agent collaboration.

Key points

• Language models sequentially predict statistically likely tokens, and later tokens do not retroactively change tokens that have already been generated. In self-supervised learning, the next word in large unstructured text collections serves as the learning signal.
• Hallucination is the generation of nonexistent or false information with high linguistic confidence. In academic research, fabricated journals, authors, volumes, page numbers, and citations may appear plausible. Web-search-based models may also rely on surface-level internet pages rather than structured academic databases.
• Retrieval-augmented generation retrieves relevant passages from curated scientific or clinical databases, reranks them, and places them in the model input. This can allow knowledge updates by refreshing the external database instead of retraining the base model, while supporting source traceability and switching between regional or institutional guidelines. OpenSciLM was presented as an example.

Audience

Researchers, clinicians, medical professionals, and scientists applying LLMs to research, clinical, medical, and scientific workflows.

Scope and limitations

This summary is based only on the supplied evidence notes from a presentation held in Seoul on March 20, 2026. The cited examples and performance-related statements were not independently checked against the original papers, replication studies, detailed methods, or source data; therefore, their effects and generalizability should not be treated as established by this presentation alone.

Research context

The content is thematically connected to medical AI and data science through its focus on literature retrieval, evidence-grounded generation, clinical data use, and model reliability in medical research.