MABGrader Advances LLM Efficiency by 26.4% While GUI-JEPA Accelerates Agent Execution 2.8–4.0x

Recent advances in LLM efficiency and agentic systems demonstrate significant performance leaps, with MABGrader outperforming SOTA proof grading by 26.4% and GUI-JEPA accelerating agents 2.8–4.0x. Wind farm control via Noise Steering reduces power loss to 2% versus 17% for industry baselines, while Timestral introduces four time series foundation models achieving SOTA on GIFT-Eval and TIME. In reasoning and retrieval, SpectRAG improves multi-hop reasoning via spectral graph augmentation, and Guided-RLVR expands LLM exploration through steering tokens. WorkFactory scales professional task generation with 112,686 tasks, improving Qwen models, while Batching GEPA optimization halves wall-clock time and boosts LiveBench-Math scores from 68.9% to 72.1%.

Evaluation frameworks and safety mechanisms are critical for reliable deployment, as OpenGray reveals hidden failures in radiotherapy planning agents requiring decision-focused evaluation beyond dose metrics. LiveSearch exposes benchmark staleness with 70% of financial questions changing daily, whereas ReCoF predicts self-recovery for 725 of 1403 agent errors. Safety protocols like Concordat enforce constraints against user manipulation, and ADTGuard enables safe streaming inference with 16M parameters. BiPA critiques skill optimizers for adding redundant rules, proposing instead to charge for turns, while TeamGym reveals a 22.9-point communication gap in team decision-making that RL training helps mitigate.

Domain-specific applications show tailored solutions for medicine, science, and security. ProactMem improves event-driven agent effectiveness by 21.4%, and CohortFlow reduces clinical cohort curation latency by 64.8%. Spiking neural networks prove significantly harder to steal than ANNs, with 10–20% lower agreement in surrogate models. In healthcare, DelSA addresses evidence coherence traps in fake news, and an ontology-guided framework improves differential diagnosis transparency. Quantum synthesis discovers efficient CNOT decompositions, while KPPax reduces biogenic emission errors by 2.4–5x in atmospheric chemistry modeling.

Memory and reasoning architectures continue to evolve, with DiscernMem refining memory quality in multi-agent teams and Dwell enabling VLMs to learn in latent worlds. Memory compression fails under rare heuristics, losing 13–24 points, whereas Rénnyi-regularized flow inference reduces mode collapse from 50% to 25%. The Trinity framework uses recursive self-improvement to organize tasks, improving mean performance by 34 points. Natural Hard-Negative Mining repurposes positives as hard negatives, matching larger retrievers, while FoG trains KG agents to retrieve evidence triples, enabling smaller models to outperform larger baselines in auditability.