Litian Liu describes an approach to hallucination detection that treats next-token prediction in language models as a classification task and adapts out-of-distribution detection methods. The description says this yields training-free, single-sample detectors with strong accuracy on…
What Kind of Hallucination Is This? Decomposing Uncertainty for Generative Models
Why it matters
The talk presents a framework for estimating uncertainty in generative models, including how kernel scores can derive uncertainty measures from a finite set of outputs. It also describes separating ambiguity in a prompt from a model's lack of…
Benchmarking Hallucination Detection: From Long-Context to Long-CoT
Why it matters
The description identifies Leman Akoglu of CMU for a talk on trustworthy AI, framing the subject as a path from hallucinations toward reliable autonomy.
Discuss: What role should AI hallucination detection play in the path from unreliable outputs toward reliable…
Rethinking Uncertainty for Trustworthy Human–AI Interaction
Why it matters
Hamed Hassani’s talk frames uncertainty as a way to govern how AI affects human decisions: systems should preserve sound human judgments while helping where people are likely to err. It also examines agentic AI, where support from humans and tools is treated as part of the…
Understanding Language Model Hallucinations through Parametric Knowledge Integration
Why it matters
Yue Dong examines how language-model hallucinations can arise from failures in knowledge enrichment and answer extraction, then considers conflicts between a model’s parametric knowledge and information supplied in long contexts. The talk says s…
Quantifying Reliability in AI: Hallucinations and Uncertainty
Why it matters
The talk frames quantifying reliability as central to trustworthy AI, then describes feature-based conformal techniques for detecting and quantifying hallucinations in image reconstruction and probabilistic regression with diffusion models for uncertainty quantification.…
Understanding the Structure of Reasoning in Language Models
Why it matters
The talk explores how language models reason under uncertainty, allocate inference-time computation, and decide when to use external information or tools. It also examines why strong underlying capabilities may not translate into reliable reasoning behavior and how internal…
In Search of Missing Mass: Disentangling Discoveries from Hallucinations
Why it matters
Eric Nalisnick's talk examines when large language models can reliably contribute to scientific discovery. The description says he will discuss extending hallucination results to characterize discovery potential and how prompts likely to yield discoveries might…
Reliability Along the Way: Progress and Failure Signals in AI Agents
Why it matters
The talk examines how to detect whether LLM agents are progressing or failing during long-horizon tasks involving tools, irreversible actions, and unpredictable feedback. It presents the progress advantage from RL post-training as a step-level uncertainty signal,…
The talk analyzes language-model hallucinations mathematically and describes how next-word prediction training can induce them even when models are trained on clean data. Its description also outlines a reduction from a “fact or hallucination” classification task to generation or prompt com…
Aligning Evaluation Incentives to Reduce Hallucination
Why it matters
The talk explains how accuracy-based LLM evaluation can reward confident guessing while failing to credit useful uncertainty. It discusses evaluation approaches that account for abstentions and the differing costs of mistakes, correct answers, and unanswered questions.
LabOS: The AI-XR Co-Scientist That Sees and Works With Humans
Why it matters
The video presents LabOS as an AI-XR co-scientist designed to see and work with humans. Its description identifies the talk as addressing trustworthy AI, from hallucinations to reliable autonomy.
Discuss: What would an AI-XR co-scientist need to do to work reliably with…
The Hot Mess of AI: How Does Misalignment Scale With Model Intelligence, Task Complexity, and
Why it matters
Jascha Sohl-Dickstein of Anthropic presents a talk on trustworthy AI, framed around hallucinations and reliable autonomy. The title raises how misalignment may scale with model intelligence and task complexity.
Sequential Statistical Methods for Reliable Machine Learning
Why it matters
The talk presents sequential hypothesis tests for continuously auditing deployed machine-learning systems, with Type I error guarantees to detect performance degradation. It also describes parameter-free online conformal prediction for adaptive, group-conditional…
Grounding Diffusion Models: The Path to Reliable Synthesis Without Hallucination
Why it matters
The talk is titled “Grounding Diffusion Models: The Path to Reliable Synthesis Without Hallucination” and is presented under the broader theme of trustworthy AI, from hallucinations to reliable autonomy. The supplied metadata frames grounding as a pat…
The talk is titled “ATHENA: Steering Generative Models toward Better Visual Reasoning” and is framed within a series on trustworthy AI, from hallucinations to reliable autonomy. The supplied metadata does not specify the method or examples discussed.
Uncertainty aware Causal Decision Making via Effect Bound Decomposition
Why it matters
Under the Trustworthy AI theme, this talk is titled around uncertainty-aware causal decision making via effect-bound decomposition. The description identifies Murat Kocaoglu of Johns Hopkins University as the speaker.
When Can We Trust What a Model Has Learned? Identifiability, Causality, and Reliable AI
Why it matters
The talk examines why models can perform well yet behave unreliably under distribution shifts or interventions, linking the problem to non-identifiability. It discusses how sparsity, environmental variation and causal assumptions may help recover…
Sampling from spherical spin glasses: diffusions and simulated annealing
Why it matters
Simons Institute's description presents Brice Huang's analysis of diffusion sampling and simulated annealing for spherical spin glasses. It reports guarantees for a denoising-diffusion algorithm and Langevin dynamics above a temperature threshold. This is a…
We use essential storage to keep WittyWires working. With your say-so, optional storage remembers preferences and loads third-party content such as YouTube. Rejecting it will not stop you using the site. Read our Privacy Policy.
Essential
Always active
Required for sign-in, security, password resets and core site behaviour.
Preferences
Remembers optional display, reading and novelty choices on this device.
Statistics
Used to understand how the site is used.Used only for anonymous site statistics.
Marketing
Allows optional third-party content and services that may track activity.
