language models learn to forecast research idea success
a new method trains language models to predict which research ideas will work better without running experiments, using a dataset of 11,488 idea pairs from paperswithcode.
topic
research
uniform weak propagation-of-chaos in shallow neural nets
new bounds show how finite-width neural network outputs stay close to their infinite-width mean-field limit over all training time, without needing strong convexity or noise.
safer offline policy picks for ad auctions
a new framework helps ad platforms decide which reserve-price policies are worth testing, using logged data without overpromising gains.
local covariate selection without strong assumptions
a new method selects covariates for causal effect estimation without requiring pretreatment or causal sufficiency assumptions, using local learning to avoid global structure search.
no feature ranking is faithful, stable, and complete under collinearity
a new proof shows that feature rankings cannot be simultaneously faithful, stable, and complete when features are correlated, and proposes an ensemble method as a solution.
best small language models on hugging face now
a curated look at top small language models under 7b parameters on hugging face, with benchmarks and code to get started.
neural mutual information estimation in masked sequence models
a neural estimator learns pairwise conditional mutual information from pretrained masked diffusion models, enabling faster parallel decoding by identifying independent variable subsets.
bounded autonomous training control for stable language model runs
a new governance layer called lbw-guard improves training stability and reduces perplexity in language models under stress conditions.
fully looped transformer stabilizes iterative block reuse
a new transformer variant fixes training instability in looped models by distributing signals across layers and reusing attention states, enabling deeper iteration without extra parameters.
learning lagrangian multipliers for milp with guarantees
new research provides statistical learning bounds for using machine learning to predict lagrangian multipliers in mixed integer linear programming, showing generalization scales with problem size and sample count.
hellora adapts moe models by targeting hot experts
hellora fine-tunes mixture-of-experts models by adding low-rank adapters only to the most active experts, cutting parameters and compute while boosting performance.
dual-channel tensor neural networks with finite-sample guarantees
a new neural network processes tensor data by splitting it into low-rank and sparse parts, with proven error bounds and automatic structure selection.
conformal prediction via transported beta laws
a new framework uses beta distributions to measure calibration-conditional coverage in conformal prediction, providing bounds on coverage gaps and bad-calibration probabilities under non-i.i.d. settings.
ai speeds up hunt for anti-aging genes
google deepmind's co-scientist helps biologists find genetic factors that reverse cellular aging, cutting analysis time from months to days.
llms with personal health records improve patient answers
study tests gemini 3.0 flash on 2,257 patient queries using personal health records, finding better accuracy with full clinical notes.
calibrated uncertainty cuts llm cascade costs by 31%
a calibration-first router uses isotonic regression to map token uncertainty to error probability, enabling cost-optimal escalation thresholds without per-workload tuning.
google genie adds street view grounding
google deepmind's project genie now uses street view imagery to create interactive virtual worlds based on real locations, available to google ai ultra subscribers.
olmoearth v1.1 cuts compute costs up to 3x
a new family of remote sensing models reduces token sequence length to lower inference costs while matching previous performance.