ttomov/ambigqa_star

✨ AmbigQA*: Measuring Uncertainty under Ambiguity

The Illusion of Certainty: Uncertainty quantification for LLMs fails under ambiguity

AmbigQA* is a benchmark dataset for evaluating uncertainty quantification (UQ) in large language models (LLMs) under non-zero aleatoric uncertainty — that is, when multiple answers are simultaneously correct with different probabilities.

While existing QA datasets such as TriviaQA assume a single ground-truth answer, MAQA* explicitly models ambiguous questions and provides ground-truth answer distributions (p*(y | x)) estimated from large-scale corpus co-occurrence statistics. This enables, for the first time, a quantitative evaluation of epistemic uncertainty (KL(p* | p)) and the separation between epistemic and aleatoric uncertainty in language models.

Example:

“What is one essential component of the fire triangle?” → {Heat 0.35, Fuel 0.36, Oxygen 0.29}

---

📊 Data Fields

Field	Type	Description
id	int	original id from the AmbigQA dataset
question	string	Original question from the MAQA dataset
rephrased_question	string	Rephrased version so that a single answer is expected
answers	list[list[string]]	All correct answers to the question
statement	list[string]	Statements formed by combining the question and each answer, used for entailment checking
main_keywords	list[string]	Main keywords used in the co-occurrence search
additional_keywords	list[string]	Additional supporting keywords used in the search
counts	list[int]	Co-occurrence counts of each answer in the Wikipedia English corpus
probabilities	list[float]	Normalized probabilities derived from Wikipedia counts
counts_redpjama	list[int]	Co-occurrence counts of each answer in the RedPajama dataset
probabilities_redpjama	list[float]	Normalized probabilities derived from RedPajama counts
counts_thepile	list[int]	Co-occurrence counts of each answer in The Pile dataset
probabilities_thepile	list[float]	Normalized probabilities derived from The Pile counts

---

For more details, please refer to the paper.

📚 Reference

If you use our datasets or results, please cite:

@misc{tomov2025illusioncertaintyuncertaintyquantification,
  title         = {The Illusion of Certainty: Uncertainty Quantification for LLMs Fails under Ambiguity},
  author        = {Tim Tomov and Dominik Fuchsgruber and Tom Wollschläger and Stephan Günnemann},
  year          = {2025},
  eprint        = {2511.04418},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2511.04418}
}

🔗 Related Resources

• 📘 Paper (Preprint): [https://arxiv.org/abs/2511.04418]
• 💻 Code: https://github.com/timtomov/llm-uncertainty-under-ambiguity
• 🤗 Collection: https://hf.co/collections/ttomov/llm-uncertainty-under-ambiguity