Boosting LLM reasoning performance through reinforcement learning with minimal-sufficient knowledge guidance, achieving state-of-the-art results at scale.

Domain-specific autoencoders significantly improve medical image super-resolution fidelity in diffusion models, outperforming generic VAEs.

A benchmark and evaluation framework for assessing the reasoning quality of large language models in coding tasks, offering improved accuracy and insights.

A memory-augmented reinforcement learning agent for sample-efficient molecular optimization in drug discovery.

A plug-and-play framework for reducing visual tokens in multimodal LLMs through class-adaptive layer fusion and dual-stage pruning, outperforming existing methods.

A prompt learning system that leverages LLMs to improve knowledge graph reasoning and predict missing information with high confidence.

A novel evaluation metric that assesses the quality of LLM reasoning by filtering for high-confidence traces, revealing deeper model capabilities.

A framework to extract LLM decision knowledge into an interpretable model, enabling calibrated probabilities and direct parameter editing.

A module that injects subcharacter compositional knowledge into Korean language models to enhance their understanding and generation capabilities.

GF-Score provides certified class-conditional robustness evaluation with fairness guarantees, decomposing aggregate scores into per-class profiles and eliminating the need for adversarial attacks.

An LLM-powered tool integrated into Google's code review system that diagnoses integration test failures with 90% accuracy, deployed across thousands of tests.

SEATrack is a simple, efficient, and adaptive multimodal tracker that uses AMG-LoRA and HMoE to improve cross-modal alignment and global relation modeling.

Demonstrates that agent identity creates attractor-like geometry in LLM activation space, enabling persistent cognitive cores.

Adapt autonomous driving models to different vehicle dynamics with a physics-conditioned framework for improved transferability.

A graph-based soft prompting framework for LLMs to perform multi-hop knowledge graph question answering by reasoning over subgraphs, reducing sensitivity to KG incompleteness.

KG-Reasoner is an end-to-end framework using Reinforcement Learning to enable LLMs to perform precise multi-hop reasoning over Knowledge Graphs.

MMA2A extends Agent-to-Agent networks with modality-native routing, significantly improving cross-modal reasoning accuracy by preserving richer context for downstream agents.

A novel agent design paradigm that grounds reasoning in deterministic computation before querying LLMs for spatial-aware benchmarks, achieving competitive accuracy.

DocSeeker is a multimodal LLM framework for understanding long documents by structuring analysis, localization, and reasoning with evidence grounding.

X-VC is a zero-shot, low-latency voice conversion tool that transforms voice characteristics without pre-training, ideal for real-time applications.

Optimize ad personalization for cold-start users using generative CTR models powered by hypernetworks.

A novel security paradigm for AI agents that separates reasoning from execution to prevent malicious actions, with a highly effective open-source implementation.

A hint-assisted reasoning framework that uses two small language models to improve mathematical problem-solving capabilities.

An end-to-end AI framework that unifies the detection and physical characterization of Fast Radio Transients, significantly reducing false positives and increasing speed.

This paper introduces an Opinion-Aware Retrieval-Augmented Generation system that addresses the bias of current RAG models towards factual content, enabling synthesis of diverse perspectives and improving retrieval diversity.

MemJack is a memory-augmented multi-agent framework that uses visual semantics to automate jailbreak attacks on VLMs, achieving high success rates and releasing a comprehensive benchmark dataset.

SIR-Bench is a benchmark for evaluating autonomous security incident response agents, measuring genuine investigation depth beyond simple alert parroting.

Develop a multimodal search agent for complex, long-horizon reasoning tasks.

A novel, annotation-free reward system for text-to-image models that leverages existing vision-language models to significantly improve prompt adherence and image quality.

A deterministic framework for enterprise-grade text categorization using LLMs, improving accuracy and reproducibility by focusing on high-value semantic features.

An offline distillation framework for large language models that significantly speeds up post-training for reasoning and code generation tasks without requiring a live teacher server.

A foundation model for relational data that natively processes connected tables, outperforming supervised methods for predictive tasks.

A human-centric topic modeling system that uses LLMs and optimal transport to generate goal-oriented and interpretable topics.

A language-agnostic method for LLM safety that anchors alignment in the model's semantic bottleneck, significantly reducing attack success rates across languages.

A Gulf-focused multimodal dataset and tool-augmented agent framework for climate decision support, improving LLM reliability through domain fine-tuning and geospatial integration.

BEAM is a bi-level evolutionary algorithm that designs high-level algorithmic structures for LLM-powered heuristic design, significantly outperforming existing methods in optimization problems and designing a novel solver for Maximum Independent Set.

Socrates Loss unifies classification and confidence calibration for deep neural networks by leveraging an auxiliary unknown class, improving stability and accuracy-performance trade-offs.

A lightweight neural network access control method that conditions activations on secret keys, enabling gated inference without encrypting weights.

A challenge and benchmark for cross-domain few-shot object detection, evaluating novel methods for unseen target domains.

A zero-shot pipeline for frugal knowledge graph construction using local LLMs, achieving competitive results with consumer hardware.

An LLM-driven multi-agent framework that enables domain scientists to autonomously build high-quality deep learning surrogate models for complex simulations using natural language.

A novel attack infers communication topologies in LLM-based multi-agent systems from black-box queries, revealing significant privacy risks.

A minimalist conversational memory framework using only retrieval and generation, addressing signal sparsity and redundancy for robust long-term dialogue management.

An end-to-end LLM agent autonomously reproduces, critiques, and extends computational physics research, demonstrating a novel mini research loop capable of identifying significant scientific findings.

A multi-agent deliberation system for LLM cascades that uses debate to improve accuracy and reduce costs in uncertain queries.

A new benchmark for self-evolving AI agents in real-world engineering tasks, focusing on iterative optimization and feasibility constraints.

A new framework and dataset for autonomous driving that significantly improves navigation understanding by integrating global and local planning, achieving state-of-the-art results without auxiliary losses.

A challenge and dataset for multimodal large language models to perform professional image quality assessment and provide expert-level reasoning.

An open-source, efficient hybrid Mamba-Transformer model for agentic reasoning that significantly outperforms existing models in inference throughput.

StsPatient simulates cognitively impaired standardized patients with fine-grained control over impairment severity and domain specificity using steering vectors and stochastic token modulation.

A novel strategy to detect and correct errors in medical image segmentation training data, improving model performance under challenging conditions.

Enabling smaller LLMs to predict their own performance limitations to improve efficiency and self-awareness in AI systems.

A novel reinforcement learning paradigm for LLMs that bifurcates policy into normal and high-entropy modes to improve exploration without sacrificing accuracy.

WiseOWL is a methodology and Streamlit app that scores ontologies on descriptiveness and semantic correctness to guide reuse and recommendations.

A novel training recipe and dataset for LLMs that generates constructive scientific paper feedback, significantly improving its validity and actionability as validated by expert authors.

A multimodal chunking pipeline that leverages document structure and vision to significantly improve RAG performance on long industrial documents.

A diagnostic framework for LLMs that moves beyond single scores to assess fine-grained abilities across scientific domains, enabling targeted model improvement and selection.

A suite of LLM co-pilots and a novel dataset for translating natural language into formal models for optimization and satisfaction problems, with an interactive editor.

A two-stage LLM framework that verifies and refines AI explanations for accuracy and accessibility, making XAI systems more trustworthy.

An evolutionary algorithm enhanced with generative replay and meta-learning for faster and more robust optimization of streaming data.

An intelligent framework automatically identifies optimal regions for vehicle counting in traffic videos, achieving high accuracy and up to 4x faster processing.

A unified model for advanced industrial anomaly detection to boost manufacturing efficiency.

TRUST Agents is a multi-agent framework for explainable fake news detection and claim reasoning, offering improved interpretability and evidence transparency.

Precision-Allocated Sparse Attention (PASA) accelerates video generation by dynamically allocating computation to critical transitions and reducing visual flickering.

SLATE is a new benchmark for evaluating tool-integrated LLM agents, and EGB is an uncertainty-aware search algorithm that improves task success and efficiency in large tool spaces.

A benchmark and methodology to evaluate deflection and hallucination in Vision-Language Models, ensuring reliability by filtering for retrieval-dependent samples and assessing behavior under conflicting evidence.

A resilient framework for autonomous cars that uses sensor fusion and an object distance correction algorithm to improve perception robustness in real-time.

A training-free framework that mitigates multimodal LLM hallucinations by dynamically perturbing text to stabilize visual grounding.

A meta-learning framework that generates synthetic tasks to improve black-box optimization from small offline datasets.

BID-LoRA, a parameter-efficient framework for Continual Learning and Unlearning that precisely deletes unwanted knowledge and efficiently integrates new knowledge while minimizing leakage.

This research analyzes how well AI agents follow plans, revealing that explicit plan guidance and reminders improve performance, suggesting a need for models that adaptively follow instructions rather than memorizing workflows.

A training-free framework for multi-style image transfer that uses mask-guided attention to prevent artifacts and preserve structure.

A cognition-inspired dual-stream model that enhances dynamic emotion recognition by integrating semantic and contextual knowledge with facial dynamics, achieving state-of-the-art performance.

A framework for improving multimodal AI by using instance-specific rubrics to generate preference data, outperforming existing methods on benchmarks.

A hardware-efficient framework for LLM quantization that suppresses activation outliers to maintain accuracy and achieve significant speedups on modern AI accelerators.

MISID, a multimodal dataset and FRACTAM framework for complex intent recognition in strategic deception games, addressing deficiencies in current MLLMs for long-context discourse and cross-modal synergy.

A prompting schema that integrates expert system heuristics to improve structured reasoning and efficiency in large language models for complex problem-solving.

A diagnostic benchmark and methodology for analyzing and attributing failures in LLM agents on long-horizon tasks.

A new benchmark and dataset for evaluating Video-LLMs in high-velocity esports environments, revealing significant performance gaps and guiding future development for virtual egocentric applications.

A physics-inspired neural operator synthesizes high-resolution multispectral images from panchromatic and low-resolution data with improved efficiency and accuracy.

A novel approach that simultaneously constrains model weights and activations to prevent safety degradation during large language model fine-tuning.

Improving LLM reasoning accuracy and calibration by optimizing for uncertainty-aware advantage estimation, reducing overconfidence and hallucinations.

A novel LLM-guided framework for time series forecasting that asynchronously fuses semantic and temporal features to achieve state-of-the-art performance and strong generalization.

A case-based learning framework for LLM agents that transfers prior task experience to new, complex real-world settings, improving structured analysis and performance.

An AI system that automates clinical research workflows, from planning to report generation, without requiring coding or direct patient data access.

A framework for building longitudinal health agents that maintain user engagement and adapt to evolving goals over time.

A framework that uses LLMs to generate interpretable, high-level DNA sequence features for improved classification accuracy.

FastGrasp is a learning-based framework for fast, dexterous grasping with mobile manipulators, integrating grasp guidance, whole-body control, and tactile feedback to achieve robust manipulation across diverse scenarios.

A framework for LLM agents that uses graph simulation to enforce policies by detecting violations hidden in action outcomes.

Enabling privacy-preserving LLM inference by integrating fully homomorphic encryption with Llama 3 for secure text generation.

A new benchmark and fine-tuning approach for cross-lingual table understanding in Bahasa Indonesia documents, significantly improving VLM performance on complex tables and low-resource languages.

RePAIR offers interactive, prompt-based machine unlearning for language models, allowing users to autonomously control data erasure at inference time.

An efficient reinforcement learning algorithm for human-robot task planning and allocation in manufacturing, considering spatial awareness and dynamic environments.

A novel method to accelerate vision foundation model training by creating a chain of models that transfer knowledge sequentially, reducing computational cost without performance loss.

A proactive design approach for autonomous vehicles integrating redundancy, diversity, and anomaly detection to ensure security and resilience against cyberattacks.

This paper introduces the Platonic Representation Hypothesis for tables, advocating for permutation-invariant representations to improve robustness in table reasoning and retrieval systems, and proposes a novel structure-aware encoder.

The A-R Behavioral Space provides an execution-layer profiling method for tool-using LLM agents, characterizing their behavior across different autonomy levels and risk contexts.

A survey and taxonomy of hallucinations in Video LLMs, identifying dynamic distortion and content fabrication as key issues.

A parameter-efficient fine-tuning method that significantly reduces the computational cost of adversarial training for Vision Transformers by focusing on robustness-critical parameters.

A clustering-enhanced domain adaptation method for industrial control systems that significantly improves unknown attack detection and reduces performance degradation.

A novel LLM memory system that uses keyword bookmarks and a recall tool to significantly improve long-conversation answer quality, outperforming existing methods.

A lightweight sequential unlearning framework for LLMs that enables privacy-aligned deployment by suppressing sensitive data with minimal impact on performance.

LLM agents with dual-trace memory encoding significantly improve cross-session recall and temporal reasoning by pairing facts with contextual scene traces.

A novel deepfake detection method using Curvelet transforms with wedge-level attention and scale-aware spatial masking to enhance frequency-domain features for improved robustness.

VISTA is a self-distillation framework that improves model robustness and generalization by enforcing consistency along the optimization trajectory using validation-informed anchors.

VFA is a hardware-friendly method to optimize FlashAttention by reducing vector operations, achieving significant speedups without performance loss.

Revolutionize bug detection with LLM-based test generation for software projects.

Investigating the dynamics and mechanisms of on-policy distillation for large language models, proposing strategies to improve its success.

Reveals that instruction-tuned LLMs are fragile to simple lexical constraints, leading to significant response collapse, and proposes a two-pass generation method to recover performance.

A gold-supervision-free framework for training search agents using cycle-consistency to reconstruct original questions from search trajectories.

Empirical assessment of LLM-generated code for construction safety reveals significant silent failure rates, highlighting the need for deterministic AI wrappers and governance.

PR-MaGIC is a training-free framework that refines prompts for in-context segmentation using gradient flow from SAM's mask decoder, improving segmentation quality without additional training.

A fuzzing framework to find vulnerabilities in LLM chat templates for jailbreaking and red teaming.

Evaluates LLM and back-translation data augmentation for low-resource African languages, revealing task-specific effectiveness for NLP tasks like NER and POS tagging.

A multi-agent framework that reconstructs the narrative flow of academic papers to generate coherent and logically structured presentations.

Extracting low-resource language data from LLMs using strategic prompting, with code and corpora released.

LLMs for molecular optimization that preserve scaffold integrity and improve properties, trained on chemistry-grounded preference data.

A novel neural compression technique for temporal lightmaps, enabling high-quality dynamic global illumination in real-time rendering with significantly reduced storage.

QuarkMedSearch is a long-horizon deep search agent for Chinese medical intelligence, achieving state-of-the-art performance through novel data construction and training strategies.

A lightweight convolutional network for satellite image restoration, demonstrating competitive quality and significant latency reduction for onboard AI applications.

A physics-bounded deep learning model for solar forecasting in autonomous microgrids that eliminates nocturnal generation and achieves zero-lag synchronization during weather transients.

ARGOS is a benchmark and framework for agentic multi-camera person search, requiring agents to reason, question, and use tools to identify individuals under information asymmetry.

MODIX is a training-free framework that scales positional indices in Vision-Language Models based on information density, improving multimodal reasoning.

Analyzing architecture-dependent sample forgetting in fine-tuned image classifiers to inform data pruning and curriculum design.

Round-trip translation reveals limitations in current multilingual benchmarks and proposes a more realistic evaluation method with the Lost in Translation benchmark.

A reproducible workflow for building and validating a legally screened chemistry corpus from S2ORC, enriched with embeddings and annotations for downstream text mining.

A human-inspired multimodal memory system for social robots that selectively stores and retrieves emotional and novel experiences for personalized interactions.

A bidirectional attention classifier that improves LLMs' ability to comprehend abstract meanings in text.

SpecBound accelerates LLM inference by adaptively bounding self-speculation length and calibrating confidence layer-wise, achieving up to 2.33x speedup without modifying base LLM parameters.

Reports on the first LLM Testing competition benchmarking an LLM-based automotive assistant for car manual information retrieval failures.

A new benchmark framework to evaluate relational reasoning in LLMs by systematically varying the complexity of entity binding.

A survey of AI methods for modeling mixed automated and human traffic simulation, proposing a taxonomy and identifying gaps in existing tools.

An information-theoretic framework for task-adapted compressed sensing MRI that enables probabilistic inference for uncertainty prediction and adaptive sampling for clinical tasks.

This research explores multi-token prediction as a superior objective for training language models to perform complex reasoning tasks like planning, outperforming standard next-token prediction.

Introduces LogicEval, a framework and dataset for evaluating automated repair techniques for logical vulnerabilities in software, highlighting challenges for LLM-based approaches.

A measurement study of seven tactics to reduce cloud LLM token usage for coding agents, offering workload-dependent strategies for significant cost savings.

This research proposes a framework to infer and balance a latent prognostic factor from observational survival data, aiming to reduce unobserved confounding and improve treatment-effect estimation.

A new benchmark dataset and challenge for AI-accelerated photon and proton dose calculation in radiotherapy, enabling the development of faster and more accurate dose prediction methods.

Leveraging large-scale travel surveys and graph neural networks, this research uncovers latent patterns of urban social mixing across five global cities, revealing how mobility shapes social interactions.

A multi-agent system that uses structured reflection notes to improve the factuality and coverage of image captions generated by large vision-language models.

A novel design for companion knowledge systems that addresses entrenchment and drift in personal LLM wikis by implementing a metabolism-like process for memory updates.

A bilevel optimization algorithm for the Electric Capacitated Vehicle Routing Problem that separates routing and charging decisions to accelerate convergence and achieve near-optimal solutions.

Designing reliable LLM-assisted rubric scoring for constructed responses in STEM exams, focusing on rubric design and LLM configurations.

CycloneMAE is a scalable multi-task learning model for global tropical cyclone probabilistic forecasting, outperforming NWP systems in pressure, wind, and track forecasting.

A personalized learning tool that adapts educational content to individual student needs.

A framework that uses affect-reinforced generative augmentation to create realistic dynamic facial expressions for improved emotion recognition.

A novel Bayesian optimization approach that selectively uses related-task information to improve sample efficiency.

Automated prompt evolution using LLM agents to significantly improve password guessing rates for security auditing.

A hierarchical graph-based memory framework for LLM agents that separates memory encoding and consolidation to improve long-term coherence.

Develops semantic image communication pipelines for edge traffic monitoring that drastically reduce data transmission costs by replacing full images with compact representations.

An AI system automates patient summary generation for thoracic tumor boards, improving efficiency and accuracy in clinical practice.

Introduces SOAR, a novel post-training method for diffusion models that corrects exposure bias and improves alignment without requiring reward models.

Evaluating protein sequence representations for Parkinson's disease classification shows limited discriminative power, requiring more informative biological features.

Investigating representation geometry in vision-language models for CT enterography to improve disease assessment and report generation.

An interactive interface for simulating citizen emotional responses to bureaucratic red tape across cultures to improve policymaking.

Improving audio source separation in reverberant environments using beta-divergence based nonnegative factorization.

An evaluation of AI tools' capability to upgrade low-demand math tasks, finding moderate success rates and highlighting the distinct capability of task modification versus classification.

AISafetyBenchExplorer is a catalogue of AI safety benchmarks that reveals fragmentation in measurement and weak governance, providing a structured analysis of the current landscape.

Optimizing deep learning models with elastic net regularization and Gabor dictionaries for accurate heart sound signal classification.

This research investigates the Identifiable Victim Effect in large language models, revealing how alignment and reasoning training impact their ethical decision-making.

A hybrid cognitive architecture that uses a spiking neural network to modulate LLM behavior and enable autonomous actions.

Develops a contextual multi-task reinforcement learning approach to enable autonomous underwater vehicles to monitor marine ecosystems more effectively.

A new metric for evaluating Natural Language to SQL systems that focuses on answering the user's intent rather than just matching ground truth SQL.

Introduces Memory Worth, a lightweight primitive for governing agent memory quality by tracking co-occurrence with successful outcomes.

A hierarchical optimization framework using LLM-driven evolutionary search to efficiently tune GPU kernels across various dimensions.

A framework to transfer LLM knowledge into interpretable Tsetlin Machines for text classification, improving accuracy and transparency without runtime LLM calls.

A novel method for continual machine unlearning that uses SVD-guided orthogonal subspace projection to prevent parameter collision and interference between tasks.

A theoretical framework to reduce redundancy in explainable clustering by formally characterizing and removing duplicate pattern representations.

This research explores how memory length impacts cooperative behavior in LLM-based agents within a social particle swarm model, using Gemini and Gemma to analyze emergent social dynamics.

Continuous Knowledge Metabolism (CKM) is a framework for generating scientific hypotheses by incrementally processing evolving literature and updating a structured knowledge base.

A safe reinforcement learning approach for human-robot task planning and allocation that predicts and filters based on worker fatigue.

Introduces Residual Bottleneck Models (ResBM) for efficient low-bandwidth decentralized training of transformer architectures, achieving significant activation compression.

Aethon introduces a reference-based replication primitive for near-constant-time instantiation of stateful AI agents, shifting instantiation from duplication to reference for improved scalability and governance.

A novel framework for energy-efficient continual learning in HPC systems, focusing on agentic AI and brain-inspired architectures for self-evolving network management.

Reinforcement learning refines robotic motion plans to ensure physical feasibility, bridging the gap between planning and real-world execution.

Guidelines and a formal model to maintain academic integrity in computing colleges amidst the rise of AI tools like ChatGPT.

A tutorial bridging operations research with deep learning for sequential decision-making under uncertainty.

This paper investigates methods for training behavioral dispositions into small language models, ultimately reporting negative results across multiple experimental arcs.

This paper demonstrates that self-monitoring capabilities in RL agents only provide benefits when structurally integrated into the decision pathway, not as auxiliary add-ons.

Developing social learning strategies for virtual soft robots to accelerate brain optimization through peer knowledge sharing.

This paper analyzes resource sharing and knowledge transfer within underground forums for synthetic non-consensual intimate imagery creation, identifying intervention points for deterrence.

A biologically inspired digital twin of the visual hierarchy that generates contour sketches and refines them with semantic representations, mirroring human visual cortex for symbol invention.

This research highlights the significant risk of false positives when using current AI methods for extraterrestrial life detection due to their susceptibility to out-of-distribution samples.

This paper provides theoretical guarantees for the convergence and robustness of Dense Associative Memory, a generalization of Hopfield networks, with potential applications in memory systems.

A theoretical generalization of conditional syntax splitting for nonmonotonic reasoning from belief bases, broadening applicability.

A theoretical framework for measuring similarity in First-Order Logic arguments, accounting for structured content and contextual weights.

This paper analyzes the efficiency of proportional mechanisms in online advertising auctions, focusing on theoretical bounds for pure Nash equilibria.

A review of Large Language Model-based pedagogical agents in educational settings, analyzing design dimensions and emerging trends.

Deepfakes are wrong because they usurp our authority over the permissible uses of our image and identity by exploiting biometric features as a generative resource.

Investigating latent planning abilities in LLMs, showing that these capabilities emerge with scale and can be measured through internal planning representations.

Theoretical analysis of loop corrections to training and generalization errors in random feature models.