AI Agents Advance Reasoning While New Frameworks Enhance Evaluation

Recent advancements in AI are pushing the boundaries of agentic computing, with new frameworks enabling more sophisticated interactions across diverse domains. For instance, the "Real-Time AI Service Economy" paper proposes a hybrid architecture to manage complex service dependencies, reducing price volatility by up to 75% and ensuring decentralized markets can match centralized allocation quality. In the realm of agent benchmarks, "The World Won't Stay Still" introduces ProEvolve, a graph-based framework for programmable environment evolution, allowing for scalable and controllable evaluation of agent adaptability. For product concept evaluation, an LLM-based multi-agent system described in "An Interactive Multi-Agent System for Evaluation of New Product Concepts" uses specialized agents to gather evidence and validate concepts, aligning with expert evaluations. Task planning for autonomous systems is explored in "Agentic LLM Planning via Step-Wise PDDL Simulation," where PyPDDLEngine enables LLMs to act as interactive search policies, though agentic gains depend on environmental feedback. Medical AI is advanced by MACRO, a self-evolving agent from "Evolving Medical Imaging Agents via Experience-driven Self-skill Discovery," which autonomously discovers composite tools to improve orchestration accuracy and generalization.

The development of more robust and reliable AI agents is a central theme, with research focusing on improving reasoning, memory, and safety. "Talk Freely, Execute Strictly" proposes schema-gated orchestration to balance deterministic execution with conversational flexibility in scientific workflows, while "Agentic LLM Planning via Step-Wise PDDL Simulation" highlights that agentic gains in planning depend on the nature of environmental feedback. "Reasoning Models Struggle to Control their Chains of Thought" investigates the controllability of LLM chains of thought, finding it significantly lower than output controllability, suggesting current CoT monitoring is cautiously safe. "SAHOO: Safeguarded Alignment for High-Order Optimization Objectives" introduces a framework to monitor and control alignment drift during recursive self-improvement, yielding substantial quality gains while preserving constraints. "DeepFact: Co-Evolving Benchmarks and Agents for Deep Research Factuality" presents an evolving benchmarking approach to improve factuality verification in deep research reports. "The EpisTwin: A Knowledge Graph-Grounded Neuro-Symbolic Architecture for Personal AI" grounds generative reasoning in a user-centric Personal Knowledge Graph for trustworthy Personal AI. "Memory for Autonomous LLM Agents" surveys mechanisms, evaluation, and frontiers in agent memory, formalizing it as a write-manage-read loop. "SoK: Agentic Retrieval-Augmented Generation (RAG)" provides a unified framework for understanding agentic RAG systems, identifying systemic risks and research directions.

AI's application in specialized fields and its inherent challenges are also highlighted. In climate adaptation, "Artificial Intelligence for Climate Adaptation" uses reinforcement learning for long-term flood adaptation planning, discovering coordinated adaptation pathways. For materials discovery, "Offline Materials Optimization with CliqueFlowmer" offers an alternative technique based on offline model-based optimization that fuses direct optimization into generation, outperforming generative baselines. "Symmetry-Constrained Language-Guided Program Synthesis" introduces SymLang for discovering governing equations from noisy data, achieving an 83.7% exact structural recovery rate under 10% noise. "AutoControl Arena" synthesizes executable test environments for AI risk evaluation, revealing that alignment can degrade under pressure. "LieCraft: A Multi-Agent Framework for Evaluating Deceptive Capabilities" measures LLM deception in grounded scenarios, finding all tested models willing to lie. "FinToolBench" introduces a benchmark for evaluating financial tool learning agents in a realistic ecosystem of executable financial tools. "OfficeQA Pro" presents a benchmark for grounded, multi-document reasoning over a large corpus of U.S. Treasury Bulletins, where frontier agents struggle. "FinSheet-Bench" evaluates LLMs on financial spreadsheets, revealing limitations in extracting and reasoning over structured tabular data. "CDRRM: Contrast-Driven Rubric Generation for Reliable and Interpretable Reward Modeling" proposes a framework for high-quality rubric generation to enhance reward modeling transparency and efficiency. "S2S-FDD: Bridging Industrial Time Series and Natural Language" offers a framework for explainable zero-shot fault diagnosis by converting sensor signals into natural language summaries. "VisualScratchpad" provides an interactive interface for visual concept analysis during inference to debug vision-language models. "The Yerkes-Dodson Law for AI Agents" studies stress-performance relationships in LLM multi-agent systems, finding an inverted-U curve for cooperation. "RetroAgent" introduces an online RL framework for agents to evolve through retrospective dual intrinsic feedback, outperforming existing methods. "Trust via Reputation of Conviction" explores trust through a mathematical formulation of claims and sources, grounding reputation in conviction. "CoCo: Code as CoT for Text-to-Image Preview and Rare Concept Generation" uses executable code as a reasoning paradigm for precise text-to-image generation. "Ares: Adaptive Reasoning Effort Selection" dynamically selects reasoning effort per step for LLM agents, reducing token usage by up to 52.7%. "OSExpert: Computer-Use Agents Learning Professional Skills" introduces a framework for agents to learn professional computer skills through exploration and compositionality. "PIRA-Bench" is a benchmark for evaluating proactive intent recommendation agents on continuous, weakly-supervised visual inputs. "Rel-MOSS" addresses class imbalance in relational deep learning by oversampling minority entities. "Advancing Automated Algorithm Design" introduces EvoStage, an evolutionary paradigm that decomposes algorithm design into sequential stages with LLM integration. "CMMR-VLN" enables vision-and-language navigation agents to recall and use relevant prior experiences through multimodal memory retrieval. "SMGI: A Structural Theory of General Artificial Intelligence" recasts learning from optimization of hypotheses to controlled evolution of the learning interface. "EveryQuery" achieves zero-shot clinical prediction via task-conditioned pre-training over EHRs. "AutoControl Arena" synthesizes executable test environments for AI risk evaluation, revealing alignment illusions under pressure. "COOL-MC" verifies and explains RL policies for bridge network maintenance, providing formal safety guarantees. "Rigidity in LLM Bandits" tests LLMs for robust decision biases, finding amplified positional order into one-arm policies. "Visualizing Coalition Formation" proposes image segmentation as a testbed for coalition formation in hedonic games. "Reinforcing the World's Edge" frames continual RL as an agent-world boundary problem in decentralized MARL. "M$^3$-ACE" rectifies visual perception in multimodal math reasoning via multi-agentic context engineering. "IronEngine" presents a general AI assistant platform with a unified orchestration core. "Efficient Policy Learning with Hybrid Evaluation-Based Genetic Programming" addresses uncertain satellite scheduling. "LEAD: Breaking the No-Recovery Bottleneck" proposes a method for stable long-horizon execution by incorporating future validation. "The Third Ambition" articulates the use of LLMs as scientific instruments for studying human behavior. "CORE-Acu" integrates structured CoT with KG safety verification for acupuncture CDS. "Intentional Deception as Controllable Capability" studies deception as an engineered capability in LLM agents. "Hospitality-VQA" introduces a hospitality-specific VQA dataset to evaluate decision-oriented informativeness. "A Lightweight Traffic Map for Efficient Anytime LaCAM*" proposes a dynamic traffic map for multi-agent path finding. "The Boiling Frog Threshold" studies anomaly detection boundaries under gradual drift. "Agentic Neurosymbolic Collaboration for Mathematical Discovery" demonstrates AI-human collaboration in combinatorial design. "A Hierarchical Error-Corrective Graph Framework" incorporates multi-dimensional transferable strategy and error matrix classification. "The Struggle Between Continuation and Refusal" mechanistically analyzes continuation-triggered jailbreaks in LLMs. "Towards a more efficient bias detection in financial language models" explores cross-model-guided bias detection. "Self-Supervised Multi-Modal World Model with 4D Space-Time Embedding" presents a novel planetary-scale 4D space-time positional encoder. "CoTJudger" quantifies reasoning efficiency by extracting the shortest effective path from CoTs. "Vision Language Models Cannot Reason About Physical Transformation" shows VLMs fail to maintain transformation-invariant representations. "$\textbf{Re}^2$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving" enables LLMs to abandon unproductive reasoning paths and restart. "VisualDeltas" learns preferences from visual quality perturbations. "Ares: Adaptive Reasoning Effort Selection" dynamically selects reasoning effort per step for LLM agents. "OfficeQA Pro" presents a benchmark for grounded, multi-document reasoning. "FinToolBench" introduces a benchmark for evaluating financial tool learning agents. "The Third Ambition" articulates the use of LLMs as scientific instruments for studying human behavior. "CORE-Acu" integrates structured CoT with KG safety verification for acupuncture CDS. "Intentional Deception as Controllable Capability" studies deception as an engineered capability in LLM agents. "Hospitality-VQA" introduces a hospitality-specific VQA dataset to evaluate decision-oriented informativeness. "A Lightweight Traffic Map for Efficient Anytime LaCAM*" proposes a dynamic traffic map for multi-agent path finding. "The Boiling Frog Threshold" studies anomaly detection boundaries under gradual drift. "Agentic Neurosymbolic Collaboration for Mathematical Discovery" demonstrates AI-human collaboration in combinatorial design. "A Hierarchical Error-Corrective Graph Framework" incorporates multi-dimensional transferable strategy and error matrix classification. "The Struggle Between Continuation and Refusal" mechanistically analyzes continuation-triggered jailbreaks in LLMs. "Towards a more efficient bias detection in financial language models" explores cross-model-guided bias detection. "Self-Supervised Multi-Modal World Model with 4D Space-Time Embedding" presents a novel planetary-scale 4D space-time positional encoder. "CoTJudger" quantifies reasoning efficiency by extracting the shortest effective path from CoTs. "Vision Language Models Cannot Reason About Physical Transformation" shows VLMs fail to maintain transformation-invariant representations. "$\textbf{Re}^2$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving" enables LLMs to abandon unproductive reasoning paths and restart. "VisualDeltas" learns preferences from visual quality perturbations.

Further research explores enhancing AI's reasoning, safety, and applicability across domains. "The Third Ambition" frames LLMs as scientific instruments for studying human behavior, while "Agentic Neurosymbolic Collaboration for Mathematical Discovery" demonstrates AI-human partnership in producing novel mathematical results. "CORE-Acu" integrates structured reasoning with knowledge graph verification for safe acupuncture clinical decision support, achieving zero safety violations. "Intentional Deception as Controllable Capability" studies deception as an engineered trait in LLM agents, finding misdirection is the primary attack vector. "Hospitality-VQA" evaluates VLMs on decision-oriented informativeness for hospitality, showing domain-specific fine-tuning is crucial. "A Lightweight Traffic Map for Efficient Anytime LaCAM*" improves multi-agent path finding with a dynamic traffic map. "The Boiling Frog Threshold" analyzes anomaly detection boundaries under gradual drift, identifying a sharp detection threshold. "A Hierarchical Error-Corrective Graph Framework" enhances autonomous agents with multi-dimensional strategy and error classification. "The Struggle Between Continuation and Refusal" mechanistically analyzes jailbreaks, attributing them to a conflict between continuation drive and safety alignment. "Towards a more efficient bias detection in financial language models" explores cross-model-guided bias detection to reduce costs. "Self-Supervised Multi-Modal World Model with 4D Space-Time Embedding" introduces a novel planetary-scale 4D space-time encoder. "CoTJudger" quantifies reasoning efficiency by extracting the shortest effective path from CoTs, revealing pervasive redundancy. "Vision Language Models Cannot Reason About Physical Transformation" demonstrates VLMs' failure to maintain transformation-invariant representations. "$\textbf{Re}^2$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving" enables LLMs to abandon unproductive reasoning paths and restart. "VisualDeltas" learns preferences from visual quality perturbations. "IronEngine" presents a general AI assistant platform with a unified orchestration core. "AutoControl Arena" synthesizes executable test environments for AI risk evaluation, revealing alignment illusions under pressure. "COOL-MC" verifies and explains RL policies for bridge network maintenance, providing formal safety guarantees. "Rigidity in LLM Bandits" tests LLMs for robust decision biases, finding amplified positional order into one-arm policies. "Visualizing Coalition Formation" proposes image segmentation as a testbed for coalition formation. "Reinforcing the World's Edge" frames continual RL as an agent-world boundary problem. "M$^3$-ACE" rectifies visual perception in multimodal math reasoning via multi-agentic context engineering. "FinSheet-Bench" evaluates LLMs on financial spreadsheets, revealing limitations in extracting and reasoning over structured tabular data. "The Third Ambition" articulates the use of LLMs as scientific instruments for studying human behavior. "CORE-Acu" integrates structured CoT with KG safety verification for acupuncture CDS. "Intentional Deception as Controllable Capability" studies deception as an engineered capability in LLM agents. "Hospitality-VQA" introduces a hospitality-specific VQA dataset to evaluate decision-oriented informativeness. "A Lightweight Traffic Map for Efficient Anytime LaCAM*" proposes a dynamic traffic map for multi-agent path finding. "The Boiling Frog Threshold" studies anomaly detection boundaries under gradual drift. "Agentic Neurosymbolic Collaboration for Mathematical Discovery" demonstrates AI-human collaboration in combinatorial design. "A Hierarchical Error-Corrective Graph Framework" incorporates multi-dimensional transferable strategy and error matrix classification. "The Struggle Between Continuation and Refusal" mechanistically analyzes continuation-triggered jailbreaks in LLMs. "Towards a more efficient bias detection in financial language models" explores cross-model-guided bias detection. "Self-Supervised Multi-Modal World Model with 4D Space-Time Embedding" presents a novel planetary-scale 4D space-time positional encoder. "CoTJudger" quantifies reasoning efficiency by extracting the shortest effective path from CoTs. "Vision Language Models Cannot Reason About Physical Transformation" shows VLMs fail to maintain transformation-invariant representations. "$\textbf{Re}^2$: Unlocking LLM Reasoning via Reinforcement Learning with Re-solving" enables LLMs to abandon unproductive reasoning paths and restart. "VisualDeltas" learns preferences from visual quality perturbations.