CATArena Advances AI Agent Testing While Denario Simplifies Financial Research

Recent research advances AI across diverse domains, with GSFC introducing sparse fuzzy clustering and TLAPS-Bench proving 31/70 TLA+ properties. Think with Skills boosts success to 62%, while Hybrid systems show no single dominant manager. SelfImprove-Bench reveals inconsistent generalization, and Chain of Faith proposes reasoning rewards. Parametric CAD Bench v3 yields 61.03% reward for Claude Code Opus 5.5, and Collaborative Intelligence Benchmark finds models struggle as teammates with 14% conflict deals. Neuro-symbolic Zebra solving outperforms pure neural CoT, and WebHarbor provides 153 dockerized sites. FedReady automates site selection with 1.00 recall, and Grid Agents reduce reinforcement costs by 33-67%. Demography AI simplifies proofs, and LEDGER improves Shapley determination to 0.797. ScienceCUA boosts task success by 17.12 points, and Knowledge-centric architectures address retention. LUCA shows agents fail to use evidence correctly with 7x diagnosis variance, and Theorem interestingness combines structural features and LLM judges. smBGC improves neurodegenerative diagnosis, and EEFA assesses ethical dimensions. EVOLVE achieves +22% F1 improvement, and AttentionIndex recovers 17.0 top-k recall points. MAGAR-Bench improves 3D assembly by 3.89 points, and TIGER generates preferred research ideas with 56.6%-68.4% success. ZeroGraph enables compositional training, and RRE establishes tensor decomposition theorems. DPDM boosts accuracy to 84.92%, and Virtual patient simulations train psychiatrists. Task 'depth' predicts difficulty better than turn count, and Reflective LLM feedback improves biological predictors. CoGSeq-LTV handles missing identifiers, and I-PTC reduces token usage. Zep Agent Skills increases pass rates by 4.5-18.8 points, and PURE retains 87.9% of gains. Fake citations rise 47.8 points with prompts, and RecGroupMemory increases transactions by 1.58-1.78%. Built environment AI faces interoperability challenges, and MemHarmBench reveals stealthy prompt injections. SkillCDG improves compliance detection, and Leap-VLA achieves 95.2% accuracy. Pulse scores 71.9% on π-Bench, and CLDA ensures network safety. Behavioral profiles predict human outcomes, and OpenPerov achieves 27.05% efficiency. Multi-Agentic AI analyzes thermoelectric data, and XLA verification resolves 142+ bugs. Trail enables 85.3% robust smart home control, and SAE steering redirects robot goals. System One models show demographic bias, and LLMs memorize medical benchmarks. Human-AI collaboration requires impressionability accounting, and Reward hacking monitors struggle. Scientific VQA improves with retrieval, and NURESA enables zero-shot memory transfer. Egocentric video synthesis remains challenging, and LAND improves compute efficiency. BlackSwan boosts success rates, and Tabular models excel on relational data. Political LLMs show left-leaning bias, and Trajectory/state judges detect 94% of harms. TriPhysBench accuracy is 50.38%, and CONFIRM reduces false support to 3%. Clinical LLM serving prioritizes fleet composition, and CivicBench shows positive score differences. Crime apps rely on perceived value with 75% variance, and Safety4DM reaches 89.83% road risk accuracy. RubricWorld raises grading accuracy to 46.70%, and Diffusion/flow matching unified theoretically. BridgeDock improves molecular docking, and Recursive self-training sharpens confidence. SHAP fails on high-cardinality features, and AdKnob controls ad intensity 500x faster. VCoT-SQL helps small models outperform large ones, and Multi-engine analysis improves accuracy by 20-23 points. IJev eliminates order dependence, and Appointed-Time control ensures UAV convergence. EMG-Hybrid scales to 100k features, and Compile as Hypothesis prevents seed collapse. HiGenTrans improves recommendations by 33.7%, and Robust Reward ML learns from noisy labels. OntoDoc improves ontology completeness by 12%, and Concept-Guided Prompting reaches 47.0% recipe diversity. FOCUS reduces VLM costs by 57%, and SE2 achieves 26% higher accuracy. LLMs underrepresent malice, and Uni-LaSpec improves decoding throughput 5.11x. ChatDify reaches 0.96 success rate, and BioAgent-DF audits biomedical agents. Xiaomi Auto World Model 2.0 uses causal world models, and Post-training routing improves accuracy by 27.1 points. Chain-of-thought oversight requires separate evidence, and MemUse finds most memory entries irrelevant. Route-Guided CBF-QP enhances safe VLA, and JitSkill improves complex task completion. New proposal distinguishes emotion concealment, and Fairness-aware design, P2P inference markets, self-evolving workflows, and community alignment are explored. Hybrid models use linear attention, and Geometric morphology learning instills geometry. Neural auction mechanisms audited, and Hippocampal-style consolidation enabled. Student clickstreams interpreted, and Coding agent trajectories guided. LARK ensures GPU kernel correctness, and EEG protocols decode inner speech. ParetoSaddler optimizes agent harnesses with ClawMark at 81.35%, and BGC-Bench highlights data scarcity. Financial QA detectors miss 5.3-46.8% errors, and Self-evaluation shows inverse expressiveness/calibration. RLiD achieves 99.3% success in 12.7 hours, and SGR-OPSD improves RL success by 7.4 points. LeanGuard reduces prompt injection to 5.9%, and Medical agents exhibit silent PHI access. Value localization enables inference steering, and THReD agent completes 92.7% tasks. MV-Thinker improves recommendations, and FORGE-Bench shows LLMs fail interlocking assemblies. AI trajectories create ambiguity, and Balloon effect shifts hidden traits. BasisKV achieves 1.42x speedup, and LoFIT detects diffusion membership. SearchInject reduces injection exposure by 83%, and InterFuse improves payment risk ranking. LLMs use variable names as priors, and Shilling detectors collapse under mimicry. UPSHIFT improves Elo by 200+ points, and GeoGuesser Bench shows weak world knowledge correlation. Human-oriented world models anticipate user evolution, and Formal proof search uses Monte Carlo. Human-supervised AI for power systems, and AffToM and RxnMiner benchmarks introduced. MURFormer denoises PPG, and CARE enables recursive self-improvement. LLM psychological profiles studied, and Adaptive reasoning evaluation needed. Narrative paradoxes formally modeled, and Evidence-grounded career support provided. Predictive safety guardrails limited.

Satellite logistics optimize via offline value iteration, and speculative decoding accelerates by 3.56x. Domain adaptation improves saline-alkali land reclamation planning to 95.83% accuracy, while test-time reasoning boosts MATH-500 scores by over 15 points. LANCE enhances image fusion metrics by 11.7%, and JABER reduces QC loss by 10.05%. PRIME aids scientific ideation, and effect-aware evaluation reveals LLM function calling risks. Transformers struggle with soluble state tracking tasks, whereas avatar intelligence amplifies human agency. Internet of Agentics frameworks ensure trustworthy multi-agent systems, and Agentic AI tools like CoworkOS improve post-training scores by 9.76 points. CUA-Master achieves 80.2 on OSWorld, and ComplianceAtlas identifies regulatory gaps with 0.831 F1 accuracy. SCOPE gains 18.6% in web exploration, and Axion 1.0 trains open agents without DPO. Auto-Research Agents increase DockQ by 0.147, and DeliverBench shows multi-stage delivery coherence loss. PROMOTE eliminates false financial promotions, and Benchmark auditing reveals 22-point scoring distortions. The SIMPLE Law predicts LLM energy costs with 96.9% R², and WorkRB exposes semantic gaps in skill extraction. OLGA enables label-free continual learning, and Maia-2 recovers chess behavior via fine-tuning. CLEO achieves 83.8% pass rates with 100x fewer parameters, and Strategist reduces inference costs by 39%. DiabAgent outperforms OpenAPS for diabetes control, and Four Bridges detects 73% deception in multi-agent games. LLM judges struggle with incomplete amendment records, and a posterior-sampling agent optimizes retail bidding. Early stopping audits reveal reliability risks, and PatchScout discovers 102 new social science entities. Surveys categorize Text-to-SQL paradigms, and OmniPiano benchmarks show SOTA struggles with dexterous control. Manifold-Aware Diffusion edits spatial transcriptomics, and MemU-Bench reveals state-of-the-art GUI agents fail without structured memory. BrainGFM boosts stroke subtyping accuracy by 14%, and Taurus automates CT annotation at half the cost. Atomic Creator Tokens improve recommendation recall by 19.8 points, and local and global fidelity are validated as one statistic. Automated theorem generation uses triangular contradictions, and Fourier circuits exhibit nonconformal distortion. CalVLA improves failure discrimination AUROC to 0.7392, and Dependence-aware risk assessment lowers costs by 39.6%. Exo reduces model costs by 12–53%, and Streaming quantile matching calibrates LLMs. System-1 models match GPT-5.6 reasoning at 1/16th cost, and Biological sleep laws were derived from 1,106 nights. Skill++ resolves lifelong routing interference, and State-aware activation recomputation reduces step time by 18.16%. FastAgent accelerates CFD screening by 1.53–15.36x, and TopoQuant predicts quantization drift with R² up to 0.85. Jev retrieves demonstrations via knowledge distillation, and graph-free RAG uses AI/human feedback. Hierarchical neuro-symbolic systems ensure safety, and frozen tape modules aid small models. Subgroup offset baselines improve GRPO, and parallelism-aligned pruning supports on-device agents. Alignment training limitations in instilling norms are noted, and X-World 1.6 enhances driving simulators. WAR-CRAFT refines web API recommendations, and Agentic Context Management closes 55–84% of context pollution gaps. ALPS employs hierarchical planning for student simulations, and TAILOR certifies robust LLM heuristics. INT8 quantization is found lossless, and NPN Classification scales to 16 inputs via graph neural networks. ParallACT shows parallel execution fails on contradictory tasks, and Decision-Centric Research achieves 89.4% accuracy on scientific agent decisions.

Research continues to address safety, efficiency, and domain-specific challenges across AI applications. Recent studies highlight the need for robust evaluation metrics and the integration of human feedback in automated systems. Innovations in medical diagnostics, financial analysis, and autonomous control demonstrate significant performance gains. However, issues like prompt injections, data scarcity, and model biases remain critical areas requiring further investigation. The convergence of neuro-symbolic approaches and large language models offers promising pathways for solving complex reasoning tasks. Future work must focus on improving generalization capabilities and ensuring ethical deployment of AI technologies in real-world scenarios. Collaborative frameworks and multi-agent systems show potential but require better conflict resolution mechanisms. Energy efficiency and cost reduction strategies are increasingly vital as models scale to handle larger datasets. Theoretical advancements in diffusion models and quantization techniques enable faster and more accurate inference. Overall, the field is moving towards more specialized, efficient, and trustworthy AI solutions tailored to diverse industrial and scientific needs.