LLM Reliabilityedit
Reliability of large language model systems in realistic use.
LLM Reliability concerns whether large language model systems behave consistently, safely, and truthfully under realistic use. Reliability includes knowing when an answer lacks adequate support, not only maximizing the number of answers scored as correct.
Research contextedit
During his 2025 research internship at NUSRI-CQ, Qiao worked 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. This also matters for synthetic data: 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 NeurIPS 2026 paper compares LLM and human susceptibility to unsupported relational inference from random, unrelated, or insufficient evidence. It connects his LLM-reliability work to a broader research concern: whether AI systems ground their conclusions in adequate evidence when that evidence is limited, distributed, generated, or potentially misleading.
See alsoedit
- Illusory Pattern Perception Drives Spurious Inference in Large Language Models
- NUSRI CQ
- Synthetic Data
- Collaborative Evaluation
- Trustworthy AI
Footnotesedit
-
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). ↩
-
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. ↩