New Research Shows Agentic Systems Advance LLM Reasoning and Memory

New research explores advanced agent architectures and memory systems for LLMs, aiming to improve reasoning, task execution, and interaction coherence. ArcDeck introduces a multi-agent framework for paper-to-slide generation by reconstructing narratives, while Memory Worth (MW) provides a principled metric for agent memory governance, tracking memory co-occurrence with successful outcomes. Identity as Attractor demonstrates that agent identities induce attractor-like geometry in LLM activation space, with paraphrases converging to tighter clusters than controls. WiseOWL offers a methodology for evaluating ontological descriptiveness and semantic correctness for reuse and recommendations, scoring documentation, label-definition alignment, interconnectedness, and hierarchical balance. Memory as Metabolism proposes a companion-specific governance profile for LLM memory, focusing on mirroring users and compensating for epistemic failures through operations like Triage and Decay. Human-Inspired Context-Selective Multimodal Memory enhances social robots by capturing and retrieving textual and visual episodic traces, prioritizing emotional salience or novelty. GAM presents a hierarchical Graph-based Agentic Memory framework that decouples encoding from consolidation for robust knowledge retention and context perception. LIFE is an energy-efficient, incremental, flexible, and agent-centric AI framework for managing frontier systems like HPCs, combining an orchestrator, Agentic Context Engineering, a novel memory system, and information lattice learning.

Several papers address enhancing LLM reasoning and task performance through structured approaches and specialized benchmarks. Spatial Atlas instantiates compute-grounded reasoning (CGR) for spatial-aware agents, resolving sub-problems via computation before LLM generation. The A-R Behavioral Space characterizes tool-using LLM agents by Action Rate and Refusal Signal, offering a deployment-oriented lens. EMBER presents a hybrid cognitive architecture with a spiking neural network (SNN) that can trigger and shape LLM actions autonomously. HintMR improves SLM mathematical reasoning by providing context-aware hints generated by another SLM. KnowRL uses Reinforcement Learning with minimal-sufficient knowledge guidance, decomposing guidance into atomic knowledge points for robust subset curation. Frontier-Eng benchmarks self-evolving agents on real-world engineering tasks via generative optimization, using simulators and verifiers for iterative design. DocSeeker employs a structured Analysis, Localization, and Reasoning workflow for long document understanding, improving retrieval precision and QA scores. QuarkMedSearch focuses on Chinese medical deep search with a full-pipeline approach for agentic foundation models, including data construction, training strategies, and benchmarks. A longitudinal health agent framework is proposed to operationalize adaptation, coherence, continuity, and agency across repeated interactions for health tasks. BEAM uses a Bi-level Memory-adaptive Algorithmic Evolution for LLM-powered heuristic design, evolving high-level algorithmic structures and realizing placeholders via MCTS. LIFE is an energy-efficient, incremental, flexible, and agent-centric AI framework for managing frontier systems like HPCs, combining an orchestrator, Agentic Context Engineering, a novel memory system, and information lattice learning.

Research also delves into improving LLM evaluation, safety, and specific reasoning capabilities. AISafetyBenchExplorer catalogues 195 AI safety benchmarks, revealing fragmented measurement and weak governance. The Non-Optimality of Scientific Knowledge argues that scientific knowledge represents a local optimum shaped by historical contingency, path dependence, and institutional lock-in. REL introduces a benchmark framework that varies Relational Complexity (RC) to evaluate LLM relational reasoning, showing performance degrades monotonically as RC increases. Policy-Invisible Violations highlights that LLM agents can violate policy due to hidden facts, proposing PhantomPolicy benchmark and Sentinel enforcement framework. Beyond Scores proposes a cognitive diagnostic framework to estimate model abilities across fine-grained dimensions, generalizing across scientific domains. The DeepTest Tool Competition benchmarks LLM-based automotive assistants for identifying failures in car manual information retrieval. MISID is a multimodal, multi-turn dataset for complex intent recognition in strategic deception games, revealing deficiencies in current MLLMs. The IDEA framework extracts LLM decision knowledge into an interpretable parametric model for calibrated probabilities and quantitative human-AI collaboration. StsPatient simulates cognitively impaired standardized patients with fine-grained control over impairment severity via stochastic steering. A Two-Stage LLM Framework uses meta-verification for accessible and verified XAI explanations, filtering unreliable narratives. Cross-Cultural Simulation uses LLM agents to model citizen emotional responses to bureaucratic red tape, finding limited alignment and weaker performance in Eastern cultures. Heuristic Classification of Thoughts prompting (HCoT) integrates expert system heuristics for structured reasoning into LLMs, outperforming existing approaches in accuracy and token efficiency. RePAIR offers Interactive Machine Unlearning for users to instruct LLMs to forget targeted knowledge via natural language at inference time. Operationalising the Right to be Forgotten introduces a lightweight sequential unlearning framework for privacy-aligned deployment in sensitive environments. Cycle-Consistent Search trains search agents without gold supervision by using question reconstructability as a proxy reward. The Long-Horizon Task Mirage? diagnostic benchmark and LLM-as-a-Judge pipeline analyze long-horizon failure behaviors in LLM-based agents. Drawing on Memory introduces dual-trace memory encoding, pairing facts with narrative reconstructions, to improve cross-session recall in LLM agents. Modality-Native Routing in Agent-to-Agent Networks improves task accuracy by preserving multimodal signals across agent boundaries. TRUST Agents is a collaborative multi-agent framework for fake news detection, explainable verification, and logic-aware claim reasoning. LLM-HYPER uses LLMs as hypernetworks to generate CTR estimator parameters for cold-start ad personalization. MultiDocFusion is a multimodal chunking pipeline for RAG on long industrial documents, integrating vision-based parsing and hierarchical chunking. Transferable Expertise for Autonomous Agents uses case-based learning to convert experience into reusable knowledge assets for agents. RPRA predicts an LLM-judge for efficient inference, allowing smaller models to defer to larger ones when uncertain. PAL (Personal Adaptive Learner) transforms lecture videos into interactive learning experiences with dynamic question generation and personalized summaries. Bilevel Late Acceptance Hill Climbing addresses the Electric Capacitated Vehicle Routing Problem. A hierarchical spatial-aware algorithm with efficient reinforcement learning is proposed for human-robot task planning and allocation in production. Mathematics Teachers Interactions with a Multi-Agent System for Personalized Problem Generation reveals teachers and students wanted to modify fine-grained personalized elements of problem contexts. The Platonic Representation Hypothesis for tables posits that a semantically robust latent space must be Permutation Invariant (PI), and proposes a structure-aware encoder. Beyond Factual Grounding argues for Opinion-Aware Retrieval-Augmented Generation to handle subjective content and avoid echo chambers. Developing, Evaluating, and Deploying a Multi-Agent System for Thoracic Tumor Board automates patient case summarization for clinical practice. Human-Inspired Context-Selective Multimodal Memory enhances social robots by capturing and retrieving textual and visual episodic traces, prioritizing emotional salience or novelty. Spatial Atlas instantiates compute-grounded reasoning (CGR) for spatial-aware agents, resolving sub-problems via computation before LLM generation. The A-R Behavioral Space characterizes tool-using LLM agents by Action Rate and Refusal Signal, offering a deployment-oriented lens. EMBER presents a hybrid cognitive architecture with a spiking neural network (SNN) that can trigger and shape LLM actions autonomously. HintMR improves SLM mathematical reasoning by providing context-aware hints generated by another SLM. KnowRL uses Reinforcement Learning with minimal-sufficient knowledge guidance, decomposing guidance into atomic knowledge points for robust subset curation. Frontier-Eng benchmarks self-evolving agents on real-world engineering tasks via generative optimization, using simulators and verifiers for iterative design. DocSeeker employs a structured Analysis, Localization, and Reasoning workflow for long document understanding, improving retrieval precision and QA scores. QuarkMedSearch focuses on Chinese medical deep search with a full-pipeline approach for agentic foundation models, including data construction, training strategies, and benchmarks. A longitudinal health agent framework is proposed to operationalize adaptation, coherence, continuity, and agency across repeated interactions for health tasks. BEAM uses a Bi-level Memory-adaptive Algorithmic Evolution for LLM-powered heuristic design, evolving high-level algorithmic structures and realizing placeholders via MCTS. LIFE is an energy-efficient, incremental, flexible, and agent-centric AI framework for managing frontier systems like HPCs, combining an orchestrator, Agentic Context Engineering, a novel memory system, and information lattice learning. The Non-Optimality of Scientific Knowledge argues that scientific knowledge represents a local optimum shaped by historical contingency, path dependence, and institutional lock-in. REL introduces a benchmark framework that varies Relational Complexity (RC) to evaluate LLM relational reasoning, showing performance degrades monotonically as RC increases. Policy-Invisible Violations highlights that LLM agents can violate policy due to hidden facts, proposing PhantomPolicy benchmark and Sentinel enforcement framework. Beyond Scores proposes a cognitive diagnostic framework to estimate model abilities across fine-grained dimensions, generalizing across scientific domains. The DeepTest Tool Competition benchmarks LLM-based automotive assistants for identifying failures in car manual information retrieval. MISID is a multimodal, multi-turn dataset for complex intent recognition in strategic deception games, revealing deficiencies in current MLLMs. The IDEA framework extracts LLM decision knowledge into an interpretable parametric model for calibrated probabilities and quantitative human-AI collaboration. StsPatient simulates cognitively impaired standardized patients with fine-grained control over impairment severity via stochastic steering. A Two-Stage LLM Framework uses meta-verification for accessible and verified XAI explanations, filtering unreliable narratives. Cross-Cultural Simulation uses LLM agents to model citizen emotional responses to bureaucratic red tape, finding limited alignment and weaker performance in Eastern cultures. Heuristic Classification of Thoughts prompting (HCoT) integrates expert system heuristics for structured reasoning into LLMs, outperforming existing approaches in accuracy and token efficiency. RePAIR offers Interactive Machine Unlearning for users to instruct LLMs to forget targeted knowledge via natural language at inference time. Operationalising the Right to be Forgotten introduces a lightweight sequential unlearning framework for privacy-aligned deployment in sensitive environments. Cycle-Consistent Search trains search agents without gold supervision by using question reconstructability as a proxy reward. The Long-Horizon Task Mirage? diagnostic benchmark and LLM-as-a-Judge pipeline analyze long-horizon failure behaviors in LLM-based agents. Drawing on Memory introduces dual-trace memory encoding, pairing facts with narrative reconstructions, to improve cross-session recall in LLM agents. Modality-Native Routing in Agent-to-Agent Networks improves task accuracy by preserving multimodal signals across agent boundaries. TRUST Agents is a collaborative multi-agent framework for fake news detection, explainable verification, and logic-aware claim reasoning. LLM-HYPER uses LLMs as hypernetworks to generate CTR estimator parameters for cold-start ad personalization. MultiDocFusion is a multimodal chunking pipeline for RAG on long industrial documents, integrating vision-based parsing and hierarchical chunking. Transferable Expertise for Autonomous Agents uses case-based learning to convert experience into reusable knowledge assets for agents. RPRA predicts an LLM-judge for efficient inference, allowing smaller models to defer to larger ones when uncertain. PAL (Personal Adaptive Learner) transforms lecture videos into interactive learning experiences with dynamic question generation and personalized summaries. Bilevel Late Acceptance Hill Climbing addresses the Electric Capacitated Vehicle Routing Problem. A hierarchical spatial-aware algorithm with efficient reinforcement learning is proposed for human-robot task planning and allocation in production. Mathematics Teachers Interactions with a Multi-Agent System for Personalized Problem Generation reveals teachers and students wanted to modify fine-grained personalized elements of problem contexts. The Platonic Representation Hypothesis for tables posits that a semantically robust latent space must be Permutation Invariant (PI), and proposes a structure-aware encoder. Beyond Factual Grounding argues for Opinion-Aware Retrieval-Augmented Generation to handle subjective content and avoid echo chambers. Developing, Evaluating, and Deploying a Multi-Agent System for Thoracic Tumor Board automates patient case summarization for clinical practice. Human-Inspired Context-Selective Multimodal Memory enhances social robots by capturing and retrieving textual and visual episodic traces, prioritizing emotional salience or novelty. Spatial Atlas instantiates compute-grounded reasoning (CGR) for spatial-aware agents, resolving sub-problems via computation before LLM generation. The A-R Behavioral Space characterizes tool-using LLM agents by Action Rate and Refusal Signal, offering a deployment-oriented lens. EMBER presents a hybrid cognitive architecture with a spiking neural network (SNN) that can trigger and shape LLM actions autonomously. HintMR improves SLM mathematical reasoning by providing context-aware hints generated by another SLM. KnowRL uses Reinforcement Learning with minimal-sufficient knowledge guidance, decomposing guidance into atomic knowledge points for robust subset curation. Frontier-Eng benchmarks self-evolving agents on real-world engineering tasks via generative optimization, using simulators and verifiers for iterative design. DocSeeker employs a structured Analysis, Localization, and Reasoning workflow for long document understanding, improving retrieval precision and QA scores. QuarkMedSearch focuses on Chinese medical deep search with a full-pipeline approach for agentic foundation models, including data construction, training strategies, and benchmarks. A longitudinal health agent framework is proposed to operationalize adaptation, coherence, continuity, and agency across repeated interactions for health tasks. BEAM uses a Bi-level Memory-adaptive Algorithmic Evolution for LLM-powered heuristic design, evolving high-level algorithmic structures and realizing placeholders via MCTS. LIFE is an energy-efficient, incremental, flexible, and agent-centric AI framework for managing frontier systems like HPCs, combining an orchestrator, Agentic Context Engineering, a novel memory system, and information lattice learning.

Advancements in AI safety and evaluation include AISafetyBenchExplorer's catalogue of benchmarks, highlighting fragmented measurement and weak governance. The Non-Optimality of Scientific Knowledge posits that scientific progress may be trapped in local optima due to historical contingency and institutional lock-in. REL benchmarks LLM relational reasoning by varying Relational Complexity, revealing performance degradation with higher arity binding. Policy-Invisible Violations identifies LLM agent failures due to hidden facts, proposing solutions like the Sentinel enforcement framework. Beyond Scores offers a cognitive diagnostic framework for fine-grained ability assessment across scientific domains. The DeepTest Tool Competition evaluated LLM-based automotive assistants for identifying failures in car manuals. MISID, a multimodal dataset, reveals deficiencies in current MLLMs for complex intent recognition in strategic games. IDEA provides an interpretable framework for LLM decision-making, enabling calibrated probabilities and human-AI collaboration. StsPatient simulates cognitively impaired patients with fine-grained control. A Two-Stage LLM Framework uses meta-verification for reliable XAI explanations. Cross-Cultural Simulation shows LLM agents have limited alignment with human emotional responses to red tape, especially in Eastern cultures. Heuristic Classification of Thoughts (HCoT) integrates expert heuristics for structured LLM reasoning, improving accuracy and efficiency. RePAIR enables Interactive Machine Unlearning for users to forget targeted knowledge via natural language. Operationalising the Right to be Forgotten introduces a lightweight framework for privacy-aligned LLM deployment. Cycle-Consistent Search trains search agents without gold supervision using question reconstructability as a proxy reward. The Long-Horizon Task Mirage? benchmark analyzes agent failures in extended tasks. Drawing on Memory introduces dual-trace encoding to improve LLM agent cross-session recall. Modality-Native Routing enhances agent-to-agent network accuracy by preserving multimodal signals. TRUST Agents provides a multi-agent framework for fake news detection and explainable verification. LLM-HYPER uses LLMs as hypernetworks for cold-start ad personalization. MultiDocFusion improves RAG on long industrial documents via multimodal chunking. Transferable Expertise for Autonomous Agents utilizes case-based learning for reusable knowledge assets. RPRA predicts LLM judge scores for efficient inference. PAL transforms lecture videos into interactive learning experiences. Bilevel Late Acceptance Hill Climbing addresses the Electric Capacitated Vehicle Routing Problem. A hierarchical spatial-aware algorithm with efficient reinforcement learning is proposed for human-robot task planning. Mathematics Teachers Interactions with a Multi-Agent System for Personalized Problem Generation highlights user desire to modify personalized problem elements. The Platonic Representation Hypothesis for tables advocates for Permutation Invariant representations. Beyond Factual Grounding argues for Opinion-Aware Retrieval-Augmented Generation. Developing, Evaluating, and Deploying a Multi-Agent System for Thoracic Tumor Board automates patient case summarization. Human-Inspired Context-Selective Multimodal Memory enhances social robots with multimodal memory. Spatial Atlas instantiates compute-grounded reasoning for spatial-aware agents. The A-R Behavioral Space characterizes tool-using LLM agents. EMBER presents a hybrid cognitive architecture with an SNN for autonomous actions. HintMR improves SLM mathematical reasoning with hints. KnowRL uses RL with minimal-sufficient knowledge guidance. Frontier-Eng benchmarks generative optimization for engineering tasks. DocSeeker improves long document understanding with structured reasoning. QuarkMedSearch focuses on Chinese medical deep search. A longitudinal health agent framework supports long-term health tasks. BEAM uses Bi-level Memory-adaptive Algorithmic Evolution for heuristic design. LIFE is an energy-efficient AI framework for HPC management.