graph-augmented tree search cuts llm costs in agent planning
gats uses a layered world model and ucb1 tree search to eliminate llm calls during inference, achieving 100% success on synthetic tasks.
topic
research
drift lipschitz budget controls diffusion policy expressivity
a new study shows that the drift lipschitz constant k determines how well diffusion policies approximate optimal actions and how they behave statistically with finite data.
explaining temporal graph networks via memory backtracking
a new method explains temporal graph network predictions by tracing how past events shape node memory vectors through backtracking and topological attribution.
how ai agents exploit trust in public communication
a new framework explains how humans and large language models can strategically undermine trust in scaffolded public assertions, going beyond echo chambers and misinformation.
score matching simplifies bayesian experimental design
a new method separates score matching from policy training to make adaptive experiment design cheaper and more scalable.
who gets missed in chest x-ray ai decisions
study shows how threshold choices in long-tailed chest x-ray models can miss rare-positive patients, especially in subgroups, and how tail-aware methods reduce underdiagnosis.
how llms handle medical reasoning
a survey maps clinical reasoning levels to ai methods and tests 18 models on a new benchmark.
prediction-powered active testing cuts labeling costs
a new method uses model predictions to reduce variance in risk estimation while keeping estimates unbiased, and it adaptively selects which test points to label.
recolora keeps llm fine-tuning from forgetting past tasks
a new spectrum-aware method recursively consolidates low-rank adapters so each task builds on previous ones without overwriting them.
conformal prediction fails minority classes in drug screening
standard conformal prediction under-covers rare classes in imbalanced drug discovery data, but a class-conditional fix restores coverage.
muse spark 1.1 api released with agentic improvements
meta's muse spark 1.1 model now offers an api with better tool calling and computer use, plus a new cli plugin for easy access.
sensorfm: a foundation model for wearable health data
sensorfm is a large foundation model pre-trained on over a trillion minutes of wearable sensor data from five million people, learning a general representation of human physiology that transfers across 35 health tasks.
quantum belief updates tested on ibm heron for sequential decisions
a hardware study checks if a quantum processor can reliably update beliefs across a sequential decision task without corrupting the planner's posterior.
value of information with imprecise probabilities
a new approach to value-of-information analysis when probabilities are not precisely known, using decision-rule-specific values and fixed-measure envelopes on credal sets.
triroute jointly routes attention, experts, and kv-cache
a single lightweight controller coordinates attention mode, expert selection, and cache bit-width per token to reduce inference cost without losing quality.
bun rewritten in rust using ai agents
jarred sumner details how he used ai coding agents to rewrite the bun javascript runtime from zig to rust in 11 days, achieving a safe, stable port with minimal user impact.
agent harnesses for arc reasoning on a budget
a two-stage agent pipeline using an open-weight model without arc-specific fine-tuning achieves competitive abstract reasoning under strict compute limits.
foundation models for automatic cad generation
a study tests large language and vision models for generating 3d mechanical designs from text, using a framework with iterative refinement.