CATArena Advances AI Agent Testing While Denario Simplifies Financial Research

Recent advances enhance LLM reliability through code-based harnesses, synthetic calibration, and trajectory audits, while domain-specific applications span formalized math archives, embodied event planning, and biomedical tabular optimization. Enterprise agents utilize latent equivalence learning for routing, and attention mechanisms are extracted as live graphs from single forward passes. Social reasoning improvements via ReAdapt boost warm-introduction accuracy from 37% to 51%, and the AI Neuroscientist enables natural-language fNIRS analysis. Retrieval throughput increases 1.28–4.52× via Orthrus heterogeneous batching, though dialogue studies reveal 'early posterior collapse' hindering correction. Chain-of-thought load-bearingness is driven 98.8% by task difficulty, complicating safety monitoring, while RAG-NAROK exploits retrieval transparency to poison knowledge bases more effectively than static attacks.

Privacy-preserving multi-agent workflows and robust knowledge distillation from cross-model failures address security gaps, alongside attentional impacts of false-positive CADe prompts in colonoscopy. Automated theorem proving utilizes search-aware loss functions, yet debate mechanisms risk erasing necessary disagreement without improving accuracy. Smaller LLMs benefit from self-evolving reasoning curricula, and specialized agents like ChatT2 assist natural product research. ArticleMiner constructs ontology-guided knowledge graphs, whereas coding agents struggle to generate exploit primitives despite finding vulnerabilities. Traditional Chinese Medicine knowledge graphs utilize confidence-aware querying, and neurosymbolic models successfully learn action models under partial observability. Post-training erases cross-cultural variance via 'consensus collapse,' while OmniFysics-Nano-V2 improves physical-world understanding across modalities.

Medical imaging advances include DW-MRI temporal deep learning predicting breast cancer response with 0.90 AUC after one cycle, and multimodal pretreatment models achieving 0.86 AUC using diffusion and contrast-enhanced MRI. FusionMMT unifies multimodal multitask learning for nuclear fusion diagnostics, outperforming prior methods on the EAST-VTD640 dataset. CLAIM generates K-12 assessment items with 97.8% expert pass rates, revealing asymmetries where fill-in-the-blank generation lags behind multiple-choice. Online sequential testing identifies LLMs via probe design, framing model selection as a weighted set cover problem. TREND-10K introduces a 10,000-video dataset for quality assessment, while EADC evaluates LLM compliance using legal knowledge graphs to detect deep regulatory blind spots.

SWE-Serve benchmarks agentic engineering, revealing a gap where 45.9% of locally passing patches fail end-to-end production tests. Transformer-based models improve real-time hand gesture recognition via OpenXR, and Null-Basis LoRA preserves reasoning in post-RL LLMs during fine-tuning. Rollout efficiency is critical for reasoning RL, requiring new taxonomies, while ZeroGate uses fast paths for AI agents but shows prepared admission adds latency. A layered framework aligns offline proxies with online A/B tests, achieving 81.1% F1, and decision models follow option names over rubrics, causing errors despite type safety. Quantum search reduces active device detection complexity in energy-harvesting networks, and JEV-as-a-Judge offers economical evaluation retaining 99% accuracy via cascading.

Verdict substitution harms verifiers by -11.2pp but benefits them in other configs by +24.2pp, driven by state-specific accuracy. Exact-support generation trades front-loaded coordination for online memory, with costs scaling by conservation rank. ChainUQ improves LLM uncertainty via reasoning consistency, gaining 3.1% AUROC and 45% ECE reduction. Runtime policies (FIRE) boost agent reliability on Terminal-Bench 2.1, converting reachable solutions into dependable delivery. VideoX-Qwen uses 1.2M+ data records to achieve top performance on 9/11 video editing metrics, while trait interference in LLM simulators is mitigated by PQA. Contrastive Epistemic Decoding neutralizes swarm consensus, recovering up to 30.75% accuracy, and audits find queer representation is 0-1.4% in speech datasets.

LSTM ensembles with PSO predict Ajichay River runoff with $R^2$ 74.95–91.42%, and multiplayer bandit algorithms using KL bounds improve regret guarantees by at least two. Dual-Frontier formalizes world-model failure attribution, admitting decisions only when advantage exceeds certified error bounds. FISSION generates labels by splitting account activity, outperforming prior bot detection methods, while ENTOP audits evidential learning showing implicit training lacks count signal. REFLEX uses a typed decision layer to reduce strong-model calls by 72.7% while maintaining 95% success. AIDE$^2$ achieves recursive self-improvement, discovering seven code changes that reduce reward hacking from 55% to 32%. Adaptive regulation burden varies by disturbance source, with internal noise causing largest exposure, and neutral-atom quantum computing solves NP-hard NOMA resource allocation via MIS reformulation.

Hybrid AI frameworks for academic advising use Stacking Ensemble models (RMSE 2.35) on 416,558 University of Birjand records. MAC-RRG, an iterative multi-agent framework for X-ray report generation, fuses multimodal knowledge graphs to reduce hallucinations. MaP, a masked trajectory prediction framework, resolves gradient conflicts in multi-task GUI navigation. Task state enforcement improves LLM agent reliability on state-decidable failures but harms performance when the gate is incorrect, while self-written ledgers often outperform shown checklists. MedGate-Fusion integrates narratives and biomarkers for stroke risk stratification using 102,736 Canadian patient records with redacted terms. FairMon, a runtime monitoring tool using extended RTLola and real-time visualization, analyzes algorithmic fairness in high-stakes systems.