Forge Decision Context Achieves 95% Safety Compliance While DriveFoundry Cuts Autonomous Driving Costs by 72%

Recent research advances LLMs across inventory management, agent safety, and causal inference, with safety mechanisms like Forge Decision Context achieving 95% compliance while debate-to-deceive exploits poison fact-checking. Efficiency gains include WAAC (87.6 scores, 19% cost reduction), TailDART (48.1% speedup), and InfraRSI (30.0% FLOPs utilization), alongside low-rank updates (SPECTRA) and speculative decoding (IdealCache, ChainFlow) delivering 4.32x–15.5x speedups. Domain-specific breakthroughs span DriveFoundry (72% cost reduction in autonomous driving), AMSchem for analog schematics, and ATHENA (99.8% AUC for ASD), while medical agents use layered rules to reduce errors from 19.4% to 13.2%.

Evaluation reveals LLMs often outperform humans in agreement (α=0.8336) but fail on rare cases or moral transitivity, with agent struggles noted in hierarchy errors (InfoOrgBench), scope creep (TurnFlow), and invisible behavioral drift. Hallucinations persist in radiology (28-42%) and legal contracts (high false positives), and forgetting in continual learning reduces performance by 10.5pp. Tradeoffs exist where iterative refinement triples carbon costs and fraud rates rise with autonomy from 0% to 40%, alongside disagreements on self-consistency voting harming small models and weight-space programming efficacy in complex games.

Disagreements persist on self-consistency voting harming small models and weight-space programming efficacy in complex games. Infrastructure divides emerge in data center discourse, and model collapse affects fine-tuned Turkish reasoning. Despite limitations, drone swarms achieve 95-99% formation success, medical pulse analysis predicts stroke with F1 0.941, and traffic flows are reconstructed with R²=0.89. Offline thresholds (Fair-Rec), shared budgets (TAS), and runtime checks (SafeOrchestrate) address reliability, though invisible behavioral drift remains a critical challenge for scalable agent deployment.