structured nonparametric vi for dependent latents
a new variational inference method uses splines to capture complex latent dependencies, moving beyond mean-field assumptions.
topic
research
export controls block ai bug fixing
export restrictions on claude fable 5 stem from a misinterpreted jailbreak that was actually a defensive code review request.
policy regret for embedding model routing
a new algorithm tackles dynamic routing of queries to multiple embedding models under adversarial conditions and bandit feedback.
audited conformal prediction for classification under distribution shift
a new method uses a small target dataset to train an audit model that improves prediction set coverage for pretrained classifiers facing unknown distribution shifts.
what makes a good explanation for ai outputs
a new definition of good explanations considers the listener's prior beliefs, highlighting why explaining large language model outputs is especially hard.
dr-dci: scaling direct corpus interaction with dynamic workspace expansion
a new framework combines retrieval with direct corpus interaction, letting agents pull relevant documents into a local workspace for flexible search and verification without full-corpus slowdowns.
personality clashes sent anthropic's models offline
axios reports internal friction and export control issues at anthropic, with key staff meeting the commerce department amid model safety concerns.
measuring trust between ai agents
researchers propose a behavioral measure of trust between ai agents using costly verification in a survival game, finding that larger models reduce verification with reliable teammates but differ in recovery after failures.
gpu workflow for hypersonic flow emulators
a fully gpu-based method speeds up building neural emulators for hypersonic flows with physics-aware refinement and uncertainty quantification.
recursive diffusion training converges to a smoothed limit
even with perfect learning, early stopping in diffusion models causes geometric drift toward a unique limiting distribution that acts as a low-pass filter on the data.
localized multidirectional correction for refusal suppression in moe models
a new method called lomc suppresses refusals in mixture-of-experts models using compact, support-gated edits that preserve general performance.
user portrait planning adapts dialogue strategies in real time
a new framework uses user portraits and nested rollout to let large language models dynamically tailor dialogue policies without offline training.
optimal transport recovers geometric shifts for domain adaptation
a study shows optimal transport with p-norm cost can recover rotation, translation, or scaling between source and target domains in 2d, enabling linear regression adaptation with scarce target data.
decision trees unified via bregman divergences
a new framework extends decision trees to any bregman divergence loss, generalizing classic cart and enabling flexible, interpretable models for diverse data types.
transformer policy for open shop scheduling via deep reinforcement learning
a transformer-based scheduling policy trained with deep reinforcement learning solves open shop scheduling problems and scales to large instances without retraining.
mobile npu framework speeds up diffusion llm inference
llada.cpp is the first npu-aware framework for accelerating diffusion large language models on smartphones, using multi-block speculative decoding and other techniques.
sqlite column provenance for datasette queries
simon willison explores methods to map sql query result columns back to their source table and column in sqlite, enabling richer data display in datasette.
publishing wasm wheels to pypi for pyodide
pyodide now supports installing packages with webassembly wheels directly from pypi, removing the need for maintainers to manually host them.