google deepmind and isomorphic labs share bioresilience approach
google deepmind and isomorphic labs outline a joint program using ai to prevent misuse, detect outbreaks, and speed up responses to biological threats.
topic
research
breaking the one-quarter barrier in two-time-scale optimization
new bias correction method improves convergence rate of non-expansive stochastic approximation from t^{-1/4} to t^{-1/3}.
auditing chain-of-thought reasoning with predicate substitution
a black-box method tests whether llm reasoning steps truly depend on stated premises by swapping predicates and checking if conclusions change.
what your model threw away and why you'll want it back
a new framework measures the symmetry information discarded by machine learning models under lie group actions, with implications for privacy and fingerprinting.
how pytorch computes gradients in physics-informed neural networks
a step-by-step trace of pytorch's automatic differentiation for pinn training, showing how it handles two levels of differentiation with explicit numerical values.
federated explainable ai survey
a systematic review of federated explainable ai, covering roles, architectures, evaluation, and open challenges.
joint tumor growth and dropout modeling with eb-vaes
a new method combines tumor growth and dropout predictions using empirical bayes variational autoencoders with genetic covariates.
tight minimax bounds for fair bandits
new research provides tight minimax regret bounds for fair multi-armed bandits when fairness is measured by negative p-means, resolving an open problem.
record-level data provenance for ai training datasets
originblame tracks author identity at record and token level, enabling precise forget sets for machine unlearning and reducing over-deletion.
optimal market making in perpetual futures
a theoretical framework for adaptive market making in perpetual futures markets with zero maker fees, deriving optimal bid-ask spreads and hedging strategies.
ai assistant for india's tinkering labs
google deepmind launches atl saathi, a gemini-powered web app that gives educators in india's atal tinkering labs a 24/7 planning and training assistant.
why diffusion models create new data
a study shows that diffusion models generate novel outputs because neural networks learn a smoothed score function, causing interpolation between training points.
voice ai still struggles in real conversations
a new benchmark with over 1 million human ratings shows voice models often fail at listening, emotion, and real-world noise despite high scores on traditional tests.
phase-controlled delta memory with chunk-wy kernels
a new attention method uses fourier phase control and chunk-wise factorization to improve state tracking and long-context memory in linear attention models.
viable path entropy measures bounded reflection in ai
mirror theory proposes studying intelligent systems by their capacity for coherent self-reflection, operationalized through viable path entropy, a measure of verified continuation diversity under budget constraints.
fixing shape-prior shortcuts in single-shot 3d scanning
a new network design prevents depth-from-boundary shortcuts in fringe projection profilometry by using a wrapped-phase representation and calibration layer.
ontology-amplified distillation for sovereign enterprise models
a study tests ontology-amplified distillation on a small language model for regulated finance, finding similar grounding to a frontier model but with limited statistical power.
testing causal graphs with outlier events
a new method uses outlier events to falsify incorrect causal graphs, providing statistical tests with false positive control and power guarantees.