Format Online paper / PDF "Addressing hallucination incentives requires changes to the widely used evaluation infrastructure, not only the creation of new specialised tests." View Source Location PART 3 - THE THIRD AXIS CHAPTER 15: DEPLOY NOW, PRAY LATER Can We Automate Uncertainty? Source title Why Language Models Hallucinate Source authors Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, Edwin Zhang Source year 2025 Confidence high Importance high
Format Web Page "Smaller models can sometimes find it easier to recognise information gaps and admit uncertainty than models with partial knowledge." View Source Location PART 3 - THE THIRD AXIS CHAPTER 15: DEPLOY NOW, PRAY LATER Can We Automate Uncertainty? Source title Why language models hallucinate Source authors OpenAI Source year 2025 Confidence high Importance high
Format Web Page "Penalising confident errors more heavily than appropriate uncertainty can encourage language models to abstain instead of guessing." View Source Location PART 3 - THE THIRD AXIS CHAPTER 15: DEPLOY NOW, PRAY LATER Can We Automate Uncertainty? Source title Why language models hallucinate Source authors OpenAI Source year 2025 Confidence high Importance high
Format Online paper / PDF "Reducing confident hallucinations requires distinguishing correct answers, incorrect answers and appropriate abstentions." View Source Location PART 3 - THE THIRD AXIS CHAPTER 15: DEPLOY NOW, PRAY LATER Can We Automate Uncertainty? Source title Why Language Models Hallucinate Source authors Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, Edwin Zhang Source year 2025 Confidence high Importance high
Format Web Page "Accuracy-only leaderboards and model evaluations can incentivise developers to favour models that guess rather than abstain." View Source Location PART 3 - THE THIRD AXIS CHAPTER 15: DEPLOY NOW, PRAY LATER Can We Automate Uncertainty? Source title Why language models hallucinate Source authors OpenAI Source year 2025 Confidence high Importance high
Format Online paper / PDF "Standard training and many accuracy-focused evaluations reward guessing over acknowledging uncertainty." View Source Location PART 3 - THE THIRD AXIS CHAPTER 15: DEPLOY NOW, PRAY LATER Can We Automate Uncertainty? Source title Why Language Models Hallucinate Source authors Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, Edwin Zhang Source year 2025 Confidence high Importance high
Format Web Page "Language models can generate plausible but false statements because learning the statistical distribution of language is not the same as verifying facts." View Source Location PART 3 - THE THIRD AXIS CHAPTER 15: DEPLOY NOW, PRAY LATER Can We Automate Uncertainty? Source title Why language models hallucinate Source authors OpenAI Source year 2025 Confidence high Importance high
Format Online paper / PDF "Even with error-free training data, statistical pressures in language-model pretraining can produce factual errors." View Source Location PART 3 - THE THIRD AXIS CHAPTER 15: DEPLOY NOW, PRAY LATER Can We Automate Uncertainty? Source title Why Language Models Hallucinate Source authors Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, Edwin Zhang Source year 2025 Confidence high Importance high
Format Online paper / PDF "During standard language-model pretraining, statements are not accompanied by explicit true-false labels." View Source Location PART 3 - THE THIRD AXIS CHAPTER 15: DEPLOY NOW, PRAY LATER Can We Automate Uncertainty? Source title Why Language Models Hallucinate Source authors Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, Edwin Zhang Source year 2025 Confidence high Importance high
Format Online paper / PDF "In a review of 1,000 chain-of-thought research papers, approximately one quarter explicitly treated CoT as an interpretability technique." View Source Location PART 3 - THE THIRD AXIS CHAPTER 15: DEPLOY NOW, PRAY LATER Can We Automate Uncertainty? Source title Chain-of-Thought Is Not Explainability Source authors Fazl Barez et al. Source year 2025 Confidence high Importance mid
