annealed entropic allocation for ranking and selection
a new annealed weighted soft-min framework improves sequential budget allocation in ranking and selection by smoothing the maximin objective and adding saddlepoint corrections.
topic
research
jeremy howard on slowing recursive ai self-improvement
jeremy howard argues that if a lab wants to slow frontier ai progress, it should not use its own top model for that research, while others should have access.
google deepmind launches robotics accelerator for european startups
google deepmind selects 15 robotics startups for a three-month program offering mentorship and ai tools to build real-world applications.
topological features as survival data for testing and learning
a new method treats persistence diagrams as survival data, enabling hypothesis testing, effect sizes, and stable feature vectors for machine learning.
prohiflo: hierarchical flow matching for protein generation
a new hierarchical flow matching framework generates proteins with functional guidance and fewer sampling steps.
boltzmann margin gives knn near-exponential convergence
a new margin condition bridges the gap between polynomial and exponential rates for knn classifiers.
how audio and visual signals flow through multimodal llms
a study traces how audio-visual large language models route and integrate sensory information, revealing task-dependent pathways and a shift in routing for interleaved inputs.
new framework for auditing machine unlearning
google research introduces regularized f-divergence kernel tests to verify machine unlearning with higher sensitivity and fewer samples.
predictive ai compresses search before exploration
a geometric model shows how predictive ai stabilizes problem-solving trajectories early, reducing exploratory diversity and potentially trapping users in narrow strategy regions.
memory tools can make ai models worse
new research shows that ai memory systems can degrade model accuracy by amplifying user misconceptions and irrelevant preferences.
diffusiongemma speeds up local text generation 4x
google deepmind releases an experimental open model that generates text in parallel blocks, achieving up to 4x faster inference on dedicated gpus.
ai optimization can reduce adaptive responsiveness
a new theory shows how ai-assisted optimization may trap systems in local efficiency, reducing their ability to explore and adapt.
ai models fail classic attention test on longer tasks
top ai models lose focus and accuracy on a classic psychology test when lists get longer, revealing a fundamental weakness in sustained attention.
synib: training objective for multimodal synergy
a new training objective called synergistic information bottleneck targets task-relevant information that only emerges from combining multiple modalities.
how memory design shapes privacy in ai agents
a study measures how memory settings in foundation-model agents affect personalization, data extraction risk, and deletion fidelity.
conformal risk prediction for fatty liver disease
a machine learning framework combines gradient boosting with conformal prediction to give reliable individual risk estimates for non-alcoholic fatty liver disease.
uncertainty-aware subspace rectification for trustworthy mllm decoding
a new training-free method reduces hallucinations in multimodal language models by adaptively projecting hidden states onto a language-prior subspace, preserving helpful priors while suppressing harmful ones.
business world model for autonomous planning
a business world model encodes states, dynamics, and actions to let ai agents simulate and optimize business initiatives.