risk control for ai agents under retrieval and tool drift
a new method controls risk in ai agents that retrieve documents and use tools, even when their behavior changes over time.
topic
research
gaussian mixture attention cuts transformer cost
a new attention method uses probabilistic routing through learned gaussian components to achieve linear-time sequence mixing, avoiding the quadratic cost of standard attention.
vlm framework reduces hallucinations with citation-grounded reasoning
a new vision-language model pipeline uses closed-loop retrieval and verification to enforce evidence-based outputs and measure step-level faithfulness.
glm-5.2 open weights llm tops benchmarks
z.ai's glm-5.2, a 753b parameter mixture-of-experts model with 1m token context, leads open weights benchmarks and ranks second in code arena webdev.
language-guided 3d motion forecasting with molmomotion
molmomotion predicts future 3d object motion from video, points, and text instructions, outperforming existing methods and aiding robotics and video generation.
diffusion model generates irregular clinical time series with missingness
a diffusion-based method jointly models lab values and observation patterns to create realistic irregular clinical time series from mimic-iii data.
glm-5.2: open model for long coding tasks
glm-5.2 is an open-source model with a solid 1m-token context, improved coding, and flexible effort control, ranking highest among open models on long-horizon benchmarks.
model choice matters in causal drug safety detection
a study compares classification models in the inferbert framework for pharmacovigilance, finding that simpler models can suffice and domain-specific pretraining helps.
editing multimodal models fails when inputs split
knowledge edits in multimodal large language models often revert to old facts when image-text pairs are separated, due to modality-specific knowledge storage.
fixing shape-prior shortcuts in long-range fringe profilometry
a study uses mechanistic interpretability and uncertainty quantification to diagnose and repair neural network failures in single-shot 3d imaging at long distances.
bayesian boolean matrix factorization for cancer genomics
a new bayesian method for boolean matrix factorization provides interpretable binary patterns with uncertainty estimates, applied to copy number analysis in cancer.
tighter bounds for ai benchmark uncertainty
new concentration inequality for exchangeable sequences gives sharper error estimates when comparing ai model scores on benchmarks.
skill-constrained mpc for resilient supply chains
a model predictive controller balances production and worker training to handle skill shortages and disruptions in manufacturing supply chains.
self-evolving agent boosts legal case retrieval with bm25
a rule-based agent iteratively creates and refines query rewriting rules to improve bm25 retrieval on legal cases without training.
kv cache as editable, composable notes
prefix caching fails when one field changes, but treating the kv cache as a notebook of memoized conclusions enables editing and composing precomputed skills.
ai planning tool aims to halve uk home application times
google deepmind and uk government prototype uses gemini to speed up householder planning decisions by 50%.
google maps hidden hedgerows for restoration
google research releases a vectorized dataset that turns high-resolution satellite maps into an actionable inventory of fine-scale woody features like hedgerows and copses across the uk.
structured nonparametric vi for dependent latents
a new variational inference method uses splines to capture complex latent dependencies, moving beyond mean-field assumptions.