Research Advances AI Efficiency, Safety, and Reliability Across Diverse Domains While Reducing Costs and Errors

Recent research advances AI efficiency, safety, and reliability across diverse domains. Jev-style models exhibit 'option-label bias' while PACE reduces token usage by 17-25%; GraphMemory and MRVQ cut token and memory consumption by 81-85% and 17.8-22.0x respectively. TradeGrad optimizes trading strategies, JOVE reduces cost/latency by 3.17x, and TACD achieves 7.7-11.9x speedups. In healthcare, HEAR lowers heart rate error to 1.6 BPM, FD-SCoPE answers clinician questions with 90.7-97.9% accuracy, and Breast cancer AI reaches 0.79 AUROC. THPL boosts aquaculture decision accuracy from 33% to 97%, and SoftGene enhances gene annotation.

Safety auditing via Positive-Unlabeled learning outperforms baselines by 5-17 points, and HACKTRACE detects reward hacking with 0.997 AUC. LUMOS reveals a retrieval gap where 84% of internal encoding fails to translate to 54% behavioral expression. Law&Order achieves 100% tax law autoformalization accuracy, yet LLM judges fail to match worker acceptance rates (3.0%-97.9% vs 61.1%). Jailbreak Benchmark reports unsafe rates rising to 18.65%, while PlurPO cuts endorsement of harmful actions by 89%. Algorithmic extraterritoriality describes AI-driven financial surveillance, and UK AI risk disclosures rose to 41.2% but substantive ones remain rare at 4.3%.

Efficient reasoning reduces CoT faithfulness but preserves monitorability, whereas Loop models may reduce CoT monitorability under stress. Dynamic routers underperform due to 'difficulty blindness', and Near-Zero Monitor Readout does not confirm behavioral control. Ego2World improves action validity by 4.15 points, and VIGOR achieves 43.6% gain on Robosuite. Dual-Stream OSSE-LSTM achieves 96.36-96.72% accuracy, and MetaRubric improves medical QA accuracy by 20.40 points. FlashSinkhorn 2 solves discrete EOT for 1.34×10⁸ particles in under 2.5 hours, and CALM improves text-to-image safety. GHOST hazards occur in 11.5% of GPT-5.5 interactions, and CreateScore reduces token use by 65.2% though it misses 8B model errors.