Rewind Before the Uncertainty Peak
REWIND improves language-model reasoning by revising earlier uncertainty transitions, but its accuracy gains must be weighed against additional computation and unsuccessful repairs.
REWIND improves language-model reasoning by revising earlier uncertainty transitions, but its accuracy gains must be weighed against additional computation and unsuccessful repairs.
How BIEvolve adds batch-invariance checks to AI-generated GPU kernels, improving numerical consistency while retaining most benchmark speedup.
WAC shows how consequence evaluation can improve action selection by shaping the predictive representations a planner learns, while exposing the limits of benchmark scores and counterfactual supervision.
Compact-Bench shows why preserving technical details during context compaction does not guarantee that a coding agent can finish its work.
A confidence score can look well calibrated yet struggle to identify unreliable LLM answers, changing how teams should evaluate automated routing.
Shared low-rank adaptations improve customer-level electricity forecasts, but the reported efficiency gains still require operational verification.
Why statistically valid runtime monitoring can miss changes in stochastic controllers, and how spatial evidence improves detection without guaranteeing it.
Why an AI agent's memory of a caught attack can reduce subsequent sabotage without necessarily improving the safety of its monitoring system.
Why accepting more draft tokens can increase LLM inference work, and how to evaluate decoding speedups without overlooking longer responses or accuracy losses.
CauProphet explores whether agents make better decisions by retaining and revising causal hypotheses, and what its benchmark gains reveal about the costs and limits of explanatory memory.