AI research targets agent reliability and inference efficiency

AI research is addressing model reliability through work on failure prediction, confidence calibration, financial reasoning benchmarks and forecasting evaluation. FinTradeBench focuses on financial reasoning, while OracleProto proposes benchmarking forecasting through knowledge cutoffs and temporal masking. Other work examines whether situation reports are grounded and why agents fail to load relevant skills.

Multimodal research covers visual token budgeting, training-free token pruning, video highlight recognition and long-horizon robotic manipulation. Applications include cyclone forecasting, clinical prediction, remote sensing and underwater image restoration. Agent-focused studies explore autonomous scientific discovery, protocol vulnerabilities, encrypted skill exploits and accountability through verifiable insurance.

Inference research emphasizes energy-aware scheduling, cache management, quantization and deployment on consumer hardware and mobile devices. MultiAlloc addresses budget-aware precision allocation for cache quantization, while SparPR examines runtime cache recovery under resource constraints. Persistent memory research covers error attribution, scope-aware storage rules, cross-model drift, systematic erasure and safeguards against propagating misalignment. These topics extend evaluation beyond model responses to the infrastructure, retained information and workflows supporting deployed systems.