AI Advances Math Problem Solving and Reduces Hallucinations

Researchers are developing advanced AI systems to tackle complex challenges across various domains. In mathematics, a new set of research-level questions aims to assess AI's problem-solving capabilities, with answers to remain encrypted for a short period. For retrieval-augmented generation (RAG), new methods like DAKS and alignment graphs improve traceability and reduce hallucinations in Chinese Tibetan medicine question answering, outperforming baselines on cross-KB evidence coverage. To mitigate harmful fine-tuning in LLMs, a defense dubbed "Surgery" uses "sink divergence" to steer attention heads away from learning harmful patterns, showing significant performance gains on benchmarks like BeaverTails and HarmBench. The pursuit of universal time series foundation models is questioned, with a proposal for a Causal Control Agent paradigm that uses specialized solvers and lightweight adaptors, advocating for benchmarks that prioritize "Drift Adaptation Speed" over "Zero-Shot Accuracy." For embodied agents, a dynamic in-situ task generation method (TEA) creates realistic tasks in unseen environments, revealing that current models perform poorly on basic perception and 3D interaction tasks compared to humans.

In security and planning, a framework integrates LLMs into an iterative loop with consistency checks and external feedback (e.g., digital twins) to control hallucination risk in security management, reducing recovery times by up to 30%. LLM agents are enhanced for interactive environments with ProAct, which uses Grounded LookAhead Distillation and a Monte-Carlo Critic to improve planning accuracy, outperforming open-source baselines. A benchmark called PATHWAYS evaluates web agents' ability to discover and use hidden contextual information, revealing frequent hallucinations and failures in evidence integration. Automatic optimization methods are explored for Rocq proof-generation agents, with few-shot bootstrapping showing consistent effectiveness, though not matching state-of-the-art engineered agents. For negotiation, PieArena, a large-scale benchmark, shows frontier agents match or outperform business-school students, but reveals novel behavioral differences in deception and instruction compliance.

Medical AI is advancing with domain-specific LLMs evaluated for ophthalmic patient queries, where Meerkat-7B showed strong performance, and GPT-4-Turbo grading aligned well with clinicians. A multimodal dataset and framework for ameloblastoma diagnosis integrate radiological, histopathological, and clinical images, improving variant classification and abnormal tissue detection. In drug discovery, Phi-former uses a pairwise hierarchical approach to predict compound-protein interactions by modeling motifs. For efficient LLM deployment, SDFP uses Fisher Information Trace (FIT)-based pruning for training-free, plug-and-play acceleration, achieving 1.32x-1.5x speedup. ALIVE, a hands-free alignment framework, uses adversarial learning and instructive verbal evaluation to improve LLM reasoning, generalization, and self-correction without human supervision. For explainable recommendation, RGCF-XRec integrates reasoning-guided collaborative filtering into LLMs, improving recommendation accuracy and reducing the cold-start gap. Cross-lingual knowledge transfer methods like GETR show significant improvements for low-resource languages in tasks like POS tagging and NER. Reactive programming and asynchronous reasoning are combined in Resin and Reactive Circuits for efficient belief updates in dynamic environments, achieving orders of magnitude speedup in drone swarm simulations. Generative Ontology synthesizes structured knowledge with LLM creativity for generating artifacts like tabletop games, ensuring structural validity and novelty. A survey on graph-based agent memory provides a taxonomy and techniques for knowledge accumulation and iterative reasoning. GenLoRA replaces explicit basis vectors in LoRA with nonlinear basis vector generation using radial basis functions, achieving superior fine-tuning performance. Anchored Policy Optimization (APO) mitigates exploration collapse in RL by shifting from global shape matching to support coverage. Financial RAG systems use a Reinforcement Learning framework with Fine-grained Knowledge Verification (RLFKV) to mitigate hallucinations by decomposing responses into atomic knowledge units. LeakBoost, a perceptual-loss-based interrogation framework, enhances membership inference attacks by synthesizing interrogation images to expose hidden membership signals. STProtein predicts spatial protein expression from multi-omics data using graph neural networks and multi-task learning. Speech emotion recognition leverages Whisper representations with attentive pooling methods, achieving state-of-the-art results on Persian datasets. FiMI, a domain-specialized LLM for Indian finance, adapts Mistral architecture for improved finance reasoning and tool-calling. Boolean networks are learned using strategies for learned connections, compact convolutions, and adaptive discretization, outperforming prior methods in accuracy vs. computation. OmniVideo-R1 improves audio-visual reasoning with query intention and modality attention. BABE, a benchmark for biology, evaluates experimental reasoning capabilities by integrating experimental results with contextual knowledge. Hierarchical Seating Allocation Problem (HSAP) is addressed with a framework combining PRM, RRT, and integer programming for optimal team seating arrangements. A guide to LLMs in Modeling & Simulation emphasizes principled design choices and diagnostic strategies. Geographically-aware Transformer-based Traffic Forecasting (GATTF) exploits geographical relationships using mutual information for improved motorway traffic prediction. HugRAG designs hierarchical causal knowledge graphs for RAG, enabling scalable reasoning and suppressing spurious correlations. A discrete-event simulator learns shooter behavior from VR experiments to evaluate security interventions. DyTopo reconstructs dynamic communication graphs for multi-agent reasoning via semantic matching, outperforming fixed communication patterns. MINT, a neuro-symbolic tree, reasons about knowledge gaps and elicits human inputs for objective-driven planning. M$^2$-Miner, a multi-agent framework using MCTS, automates mobile GUI agent data mining for intent-trajectory pairs. TangramSR uses test-time self-refinement with ICL and reward loops to enhance geometric reasoning in VLMs. LLMs emulate aggregate human choice behavior and biases, reproducing biases with precision in conversational settings. Energy efficiency "sweet spots" for LLM inference are identified, revealing non-linear dependencies on sequence lengths. Agent UQ research shifts to reducible uncertainty modeling for interactive agents, highlighting interactivity of actions. GAMMS, a graph-based simulator, supports fast development and evaluation of agent behavior in graph-represented environments. Explainable AI integrates GRAD-CAM, LRP, and SHAP for comprehensive insights into brain tumour detection models. Recos, a new similarity metric, outperforms cosine similarity by capturing nonlinear relationships. AMR models path-specific multiple aspects for aspect-aware MOOC recommendation, outperforming GNN baselines. AgentXRay reconstructs interpretable workflows for agentic systems using search-based optimization. AI is characterized as "strange intelligence" with nonlinear patterns of ability and inability. DeepRead, a structure-aware agent, enhances long-document QA by operationalizing document priors. LLMs support exploration of established graph theory material but are limited in novel insight. RL-based framework optimizes multi-debris ADR mission planning with refueling and adaptive collision avoidance. VERA-MH evaluation supports clinical validity and reliability for AI safety in mental health. Domain-randomized PPO and MCTS are compared for adaptive mission planning in ADR, showing trade-offs between learned policies and search-based methods. A multi-evaluator framework assesses LLM reasoning in merchant risk assessment, revealing biases and alignment with human judgment. Democratic Preference Optimization (DemPO) applies algorithmic sortition to RLHF for more representative AI alignment. SocialVeil simulates communication barriers to probe LLM social intelligence, showing significant performance impairment. RaBiT, a residual binarization training framework, achieves state-of-the-art performance and speed-up for 2-bit LLMs. OPINN, a physics-informed neural framework, models opinion dynamics using a Diffusion-Convection-Reaction system. NEX, a label-free scoring framework, ranks LLM responses and merges checkpoints by analyzing neuron exploration-exploitation. TKG-Thinker uses agentic RL for dynamic reasoning over temporal knowledge graphs, achieving state-of-the-art performance. Agent2Agent threats in safety-critical LLM assistants are analyzed using a human-centric taxonomy and workflow reconstruction. CAST-CKT framework enables chaos-aware spatio-temporal and cross-city knowledge transfer for traffic flow prediction. Conditional diffusion guidance under hard constraints is studied using a stochastic analysis approach. RL-VLA$^3$ accelerates Vision-Language-Action models via full asynchronism, achieving significant throughput improvements. Quantum RL with Transformers is applied to the Capacitated Vehicle Routing Problem, showing potential for more robust routing solutions. AgenticPay, a multi-agent LLM negotiation system, benchmarks buyer-seller transactions.

Further research explores energy efficiency "sweet spots" in LLM inference, identifying optimal input/output sequence lengths for reduced energy usage. Agent UQ research is shifting towards reducible uncertainty modeling for interactive agents, emphasizing the interactivity of actions. GAMMS, a graph-based simulator, supports fast development and evaluation of agent behavior in graph-represented environments. Explainable AI integrates multiple techniques (GRAD-CAM, LRP, SHAP) for comprehensive insights into brain tumour detection models. Recos, a new similarity metric, outperforms cosine similarity by capturing nonlinear relationships in semantic spaces. AMR models path-specific multiple aspects for aspect-aware MOOC recommendation, outperforming GNN baselines. AgentXRay reconstructs interpretable workflows for agentic systems using search-based optimization. AI is characterized as "strange intelligence" with nonlinear patterns of ability and inability, challenging linear models of progress. DeepRead, a structure-aware agent, enhances long-document QA by operationalizing document priors like hierarchical organization. LLMs can support exploration of established graph theory material but are limited in tasks requiring novel mathematical insight. A reinforcement learning-based framework optimizes multi-debris Active Debris Removal (ADR) mission planning with refueling and adaptive collision avoidance. VERA-MH evaluation supports the clinical validity and reliability of AI safety assessments in mental health. Domain-randomized PPO and MCTS are compared for adaptive mission planning in ADR, highlighting trade-offs between learned policies and search-based methods. A multi-evaluator framework assesses LLM reasoning in merchant risk assessment, revealing biases and alignment with human judgment. Democratic Preference Optimization (DemPO) applies algorithmic sortition to RLHF for more representative AI alignment. SocialVeil simulates communication barriers to probe LLM social intelligence, demonstrating significant performance impairment. RaBiT, a residual binarization training framework, achieves state-of-the-art performance and speed-up for 2-bit LLMs. OPINN, a physics-informed neural framework, models opinion dynamics using a Diffusion-Convection-Reaction system. NEX, a label-free scoring framework, ranks LLM responses and merges checkpoints by analyzing neuron exploration-exploitation. TKG-Thinker uses agentic RL for dynamic reasoning over temporal knowledge graphs, achieving state-of-the-art performance. Agent2Agent threats in safety-critical LLM assistants are analyzed using a human-centric taxonomy and workflow reconstruction. CAST-CKT framework enables chaos-aware spatio-temporal and cross-city knowledge transfer for traffic flow prediction. Conditional diffusion guidance under hard constraints is studied using a stochastic analysis approach. RL-VLA$^3$ accelerates Vision-Language-Action models via full asynchronism, achieving significant throughput improvements. Quantum RL with Transformers is applied to the Capacitated Vehicle Routing Problem, showing potential for more robust routing solutions. AgenticPay, a multi-agent LLM negotiation system, benchmarks buyer-seller transactions.