LLM Reliabilityedit

LLM Reliability concerns whether large language model systems behave consistently, safely, and truthfully under realistic use. In this wiki the term is connected to synthetic data, evaluation, and trustworthy systems rather than to a separate product-building track. Reliability includes knowing when an answer lacks adequate support, not only maximizing the number of answers scored as correct.

Role in this wikiedit

This page gives context for Qiao's 2025 research internship at NUSRI-CQ, where the biography records work on trustworthy LLM systems and synthetic-data evaluation. Reliability is used here as an umbrella for problems such as hallucination, data contamination, evaluation leakage, recursive synthetic-data use, and miscalibrated trust in generated outputs. The page is intentionally linked to Synthetic Data because generated text or multimodal data can become part of future model-training pipelines.

Recent evidence sharpens two distinctions. First, false but fluent outputs are not explained only by missing knowledge: TruthfulQA showed that language models can reproduce widely held human misconceptions, while a 2026 Nature study argued that accuracy-only evaluation can reward guessing over abstaining when evidence is weak.¹ Second, reported benchmark performance is not the same as generalization. If evaluation examples overlap with pre-training data, scores can be inflated; a 2025 ICML paper treats this overlap as measurable dataset leakage rather than as an abstract concern.² Reliable evaluation should therefore examine factual support, abstention behavior, benchmark freshness, and possible contamination together.

Connection to Qiao's workedit

Qiao's public publication pages currently emphasize machine unlearning, AI and networks, and synthetic-data model collapse rather than a standalone LLM paper. This page therefore stays conservative: it records the research context and links LLM reliability to the methods that are already visible in the wiki. The relevant methodological bridge is evaluation under imperfect evidence, especially when data are generated, distributed, or selected before training.

See alsoedit

• NUSRI CQ
• Synthetic Data
• Collaborative Evaluation
• Trustworthy AI

Footnotesedit

1. Lin, Hilton, and Evans introduced TruthfulQA to measure whether models imitate common false beliefs. Kalai et al. later showed that next-word prediction and accuracy-only evaluation can reward unsupported guessing, and proposed evaluation rules that make abstention incentives explicit in Nature (2026). ↩

2. Choi et al., "How Contaminated Is Your Benchmark? Measuring Dataset Leakage in Large Language Models with Kernel Divergence", ICML 2025, report that benchmark overlap with pre-training data can inflate evaluation metrics and study a controlled method for measuring that leakage. ↩