AI Research Radar

Oyster-II

Auto

“The same stream also quantifies the failure modes: a loose guard silently mislabels 48.9% of compiled answers, and an unfaithful deopt reference causes the verification gate to refuse recompilation.”

Multimodal Reward Hacking in Reinforcement Learning

“Scaling reduces but does not eliminate hacking: even the 32B model retains a 54.9% worse rate under outcome-only rewards, whereas answer-aware rewards improve the oracle trend at every scale.”

Governed Individuation

“An adversarial evaluation of monitors of increasing semantic depth shows false-allows falling from 75% (name-based gating) to zero (dynamic effect tracing), and refusal history transfers compliance to held-out red-line families.”

Your Agent's Memories Are Not Its Own

“Our evaluation also shows that SENTINEL reduces FARMA's attack success rate to as low as 0% with no false positives observed across 326 benign agent traces.”

Does AI Understand Imaging? A Systematic Benchmark of Agentic AI for Computational Imaging Tasks

WebRetriever

“Extensive experimental analysis reveals substantial performance disparities across evaluation protocols, demonstrating that navigation success alone is an insufficient predictor of real-world application effectiveness.”

RoboDojo

Native Video-Action Pretraining for Generalizable Robot Control

More Convincing, Not More Correct

“We show it fails structurally: conditioned on a candidate, a judge scores plausibility, not correctness, leaving false-positive basins a policy learns to exploit.”