google introduces tabfm for zero-shot tabular predictions
tabfm is a foundation model that predicts on new tables without training, using in-context learning and synthetic data.
topic
research
mmm data model for knowledge interoperability
a new data model called mmm aims to improve how knowledge is structured, shared, and reused across systems, moving beyond traditional documents.
how low-dimensional data structure shapes classification
a mathematical framework extends cover's function-counting theory to analyze how low-dimensional data structure affects binary classification capacity and generalization.
humanity's last exam is a distraction
a look at the hle benchmark, why it was created, and the divided expert opinions on its value for evaluating ai systems.
filtered mixture-of-generators for synthetic survival data
fogs selects plausible synthetic samples from multiple tabular generators to improve downstream survival model training on small clinical datasets.
fourier neural operators learn pde solutions efficiently
new theory proves fourier neural operators can approximate and learn time-dependent solutions of dissipative partial differential equations with polynomial sample complexity.
deep multitask learning for mixed outcomes with shared sparsity
a new method uses deep neural networks and rank-based optimization to handle mixed outcome types in multitask learning with shared predictor selection.
bounded morality: a computational framework for finite agents
a new framework models moral cognition as a tradeoff between moral breadth and depth under limited resources, recasting ethical theories as efficient strategies.
semiscope: tuning gains in semi-supervised security classification
a study disentangles whether improvements in semi-supervised learning for security come from tuning the classifier alone or from joint optimization with the ssl pipeline.
ai alignment must handle shifting human preferences
a new framework treats alignment as controlling how preferences evolve through interaction, not just satisfying fixed goals.
grpo, dr. grpo, and dapo are three settings of one dial
three popular language model training methods all adjust the same number: the standard deviation of correctness marks across sampled answers.
google ai updates from june 2026
google announced gemini 3.5 live translate, android 17 features, and a new home speaker, plus tools for research and education.
when feedback actually helps ai improve
a controlled study finds that multi-turn improvement often comes from resampling or extra computation, not from feedback itself, with only strong external teachers providing real gains.
when does learning to stop help reasoning models
a study shows learned stopping rules can improve reasoning model efficiency on math tasks but not on multiple-choice benchmarks.
bayesbench tests if llms update beliefs like bayesian reasoners
bayesbench evaluates how large language models update beliefs across multi-turn conversations, comparing their trajectories to rational bayesian inference.
discovering shared pdes from multiple datasets
a new method uses competitive optimization to find governing partial differential equations from several datasets with different conditions.
training llms as zero-shot workflow generators
metaflow trains language models to produce reusable task-level workflows instead of instance-specific solutions, improving reliability and interpretability.
hierarchical global attention for long-context transformers
hierarchical global attention replaces dense causal attention in pretrained transformers, enabling 64k-token context on a single 32gb gpu without retraining.