one transformer for density and score across distributions
discoformer estimates density and score in one forward pass without retraining, outperforming kernel density estimation in high dimensions.
topic
research
inferring directed graphs from diffusion data
a new method identifies directed network topology from nodal measurements by solving quadratic matrix equations for graph filter coefficients.
when rag fails: four alternatives
retrieval-augmented generation breaks in predictable ways at scale, and engineers are turning to long-context prompting, memory compression, structured retrieval, and graph-based reasoning instead.
gemini 3.5 flash gets built-in computer use
gemini 3.5 flash now includes native computer use, letting developers build agents that see, reason, and act across browser, mobile, and desktop environments.
proper positive-only learning solved with new condition
a new combinatorial condition called uniform exterior separability characterizes when a concept class can be properly learned from positive-only samples.
darts unifies zero-shot time series foundation models
darts introduces a unified interface for foundation models like chronos-2 and timesfm 2.5, simplifying zero-shot forecasting in python.
training agents to simulate future plans
a new three-stage method teaches language model agents to internally predict outcomes before acting, improving long-horizon planning.
benchmark saturation hides key agent performance dimensions
a study on core-bench shows that after accuracy saturates, measuring construct validity, generalization, efficiency, reliability, and human-agent collaboration reveals new insights into ai agent capabilities.
2,000 people tried to hack an ai assistant and failed
a public challenge to leak secrets from an ai assistant via email saw 6,000 attempts but no successful breaches, highlighting improved model defenses against prompt injection.
frozen multi-token prediction speeds up gemini nano on pixel
google retrofits multi-token prediction onto frozen gemini nano models to accelerate on-device inference without separate drafters.
which tokens hybrid models predict better
a study comparing token-level predictions of a hybrid model and a transformer reveals where each architecture excels.
persona gates refusal in chat models
refusal in ai chat models depends on a compliant persona direction, not just a standalone refusal mechanism.
detecting sycophancy with cascading linear features
a new method uses graded data samples to better isolate and control sycophancy in language models.
local mass view of bayesian inference
new tools measure how bayesian updating changes probability mass in small regions, revealing behavior missed by global divergences.
xmse-aware adaptive empirical bayes estimation
a new method blends maximum likelihood and empirical bayes estimators using excess mean squared error to improve second-order risk.
neuro-symbolic framework grounds amr predictions in biology
kg-trace combines neural networks with a knowledge graph to make antimicrobial resistance predictions that align with known biological pathways.
video games train ai agents for real world
general intuition raised $320m to build ai agents trained on gameplay data that can control robots and navigate physical spaces.
linear elastic caching cuts cloud memory costs
google research introduces linear elastic caching, a dynamic approach that treats memory as a utility to minimize total cost of ownership in cloud databases.