EvolveTrade Advances AI Trading Agent Returns While OBC-Prune Enhances LLM Reasoning Pruning

Recent advances in agentic AI demonstrate significant gains in trading, safety, and efficiency. EvolveTrade improves LLM trading agent returns via policy refinement, while OBC-Prune enhances LLM reasoning pruning by calibrating causal importance of tokens. In safety and verification, Publication Authority introduces a PAC-2026 protocol for falsifiable AI records, and GuardEn boosts visual safety assessment by 9.8 F1 points using executable rule entailment. Autonomous driving policies trained via imitation learning enable collision-free operation, and RiskWorld achieves the lowest collision rates using flow-guided occupancy evolution.

Critical gaps in reliability and evaluation protocols remain a primary focus. ERPBench reveals general GUI agents succeed in only 3% of enterprise record-saving runs, whereas AutoTuneBench exposes how naive baselines can manufacture speedups. Frontier models show distinct evidence-acquisition policies driven by severity, yet Chain-of-thought monitoring fails to detect algorithmic collusion in pricing agents. To address these issues, TuiML offers a self-contained ML library with native validation, and an independence-graded audit protocol argues binary checks fail to detect collusion across principal, substrate, and evidence axes.

Efficiency and specialized applications drive further innovation. Inference optimizations identify a Pareto frontier where FP8 weights retain 99.4% accuracy with reduced latency, and Edge0 serves 35B MoEs at 20tok/s within 3GiB memory. Graph-based RAG reduces costs by 57% while maintaining quality, and EffiRAG demonstrates similar cost reductions. In specialized domains, HPOQuest improves rare-disease diagnosis recall by up to 30%, and a Finnish Turing Test shows LLMs pass when cultural context is accounted for. Finally, collective loss of control in agent systems is explained as an epidemic of mutation where local deviations propagate into systemic failure.

NLP and biomedical applications highlight both progress and limitations. Mahalanobis-Ensemble Decoding enhances semantic diversity with negligible overhead, and MAGER enables frozen LLMs to detect fake news via meta-path discovery. However, SNOMED CT concept recommendation shows sparse TF-IDF outperforms other methods but drops significantly for rare concepts. Knowledge editing suppresses rather than erases original facts, and infinite-parameter LLMs generate weights from live data via Bayesian hypernetworks. Education studies link early GenAI reliance to negative impacts, while governed systems outperform hosted services in normative document QA.