Formal framework using Temporal Behavior Trees to repair suboptimal trajectories from imperfect demonstrations before downstream imitation and reinforcement learning.
Framework and benchmark for converting web elements into autonomous agents as foundational primitives for the Agentic Web, enabling automated agent generation from digital assets.
Analysis of Mixture-of-Experts token routing across training phases using congestion game modeling, tracking three-phase trajectory in OLMoE and OpenMoE models.
Systematic audit of probability calibration in multimodal deep learning models combining histopathology images and genomic data for cancer survival prediction.
Federated reinforcement learning from human feedback method for aligning LLMs with diverse human preferences while preserving privacy and achieving fair reward aggregation.
Interrogator-based framework for behavioral trust monitoring in autonomous underwater vehicles and IoT sensor networks with decentralized coordination.
Two preregistered experiments (N=2,012) measuring how LLM agents embed commercial persuasion into conversational recommendations compared to traditional search engines.
Framework combining blockchain-enforced oversight with AI agents for wildfire monitoring, ensuring human control and cryptographic verification in safety-critical autonomous systems.
Study on Claude Opus 4.6's ability to preserve poisoned identifier names during JavaScript deobfuscation across 192 inference runs, revealing consistent persistence patterns.
Integration of persistent homology into deep learning architectures for 3D point cloud analysis capturing multi-scale topological structure invariants.
HighFM foundation model for learning representations from high-frequency satellite earth observation data for climate disaster monitoring and early warning.
Distributional reinforcement learning approach for optimization in robotics and healthcare addressing decision-making under uncertainty with heterogeneous groups.
GroundedKG-RAG system using knowledge graph indexing for retrieval-augmented generation in long-document question answering with LLMs.
Techniques for reducing computational cost of extreme learning machine classifiers using integer-only operations without accuracy loss at test time.
Research identifying sparse routing mechanisms in alignment-trained language models, localizing gate attention and amplifier heads controlling policy behavior.
Method for stable language model alignment using relative density ratio optimization with statistical consistency guarantees beyond Bradley-Terry assumptions.
Study revealing gaps between internal representations and responses in vision language models for visual document understanding tasks.
Metric framework formalizing error verifiability to measure whether LLM justifications help users distinguish correct from incorrect answers.
Investigation of prompt selection necessity in task-free online continual learning with non-stationary data streams and no task boundaries.
Training approach for transformers using discrete cosine transform domain parameterization with reduced coefficients for efficient weight matrix representation.
Framework for constrained LLM generation using ontological definitions to enable modular and explainable control in conversational agents.
Method for differentially private LLM compression via on-policy distillation, balancing privacy guarantees with model deployment efficiency.
Research on federated time series foundation models using discrete prototypical memories to align time-series data with LLM latent space.
Paper presenting SLaB framework for LLM compression via sparse-lowrank-binary decomposition, maintaining performance at high compression ratios.
Research paper on controllable LLMs with multi-objective alignment to varying human preferences, extending beyond fixed reward RLHF approaches.
GAIN method uses multiplicative modulation for domain adaptation in LLMs, re-emphasizing existing features instead of injecting new directions.
Reproducibility study on detecting and fixing spurious correlations, shortcut learning, and group-distributional non-robustness in DNNs.
ENCRUST pipeline for safe C-to-Rust translation using agentic LLM refinement with whole-program reasoning on live scaffolds.
Study showing evaluation language choice inverts agent-as-judge rankings across five languages on 55 development tasks, revealing backbone sensitivity.
StableTTA training-free test-time adaptation method improving ensemble prediction stability and computational efficiency on ImageNet.
Systematic taxonomy from 10,000 trials identifying which system prompt features trigger LLM agents to exploit security vulnerabilities across models.
Paper Espresso open-source platform automatically discovers, summarizes and analyzes trending arXiv papers using LLMs with structured labeling.
PassiveQA framework for calibrated question answering that handles incomplete or ambiguous queries through three-action decision awareness.
Contrastive hypothesis retrieval method for medical RAG systems that suppresses clinically distinct but semantically similar negatives.
Legal analysis of how EU AI Act regulates autonomous AI agents across enterprise functions including customer service and clinical decision support.
ROSClaw framework integrates LLMs with embodied agents to bridge semantic understanding and physical execution for multi-agent robot collaboration tasks.
Implementation of LLM-based AI teaching assistant using RAG for a Master's program in motion picture engineering.
Research on multimodal fact-checking showing visual evidence doesn't universally improve performance in automated fact-checking systems.
Quantization method for LLMs using mixed-to-uniform precision and low-rank decomposition for efficient on-device deployment.
Bilingual corpus of Bangla-English sentences annotated for syntactic structure and tense for low-resource multilingual NLP.
Analysis of what characterizes effective reasoning in multilingual large reasoning models, challenging assumptions that English reasoning patterns transfer.
Study analyzing combined effects of English as second language and typographical errors on LLM performance in multilingual contexts.
Computational audit examining whether LLMs conduct culture-aware reasoning or merely translate between cultures in creative writing tasks.
Reinforcement learning approach for automatically discovering failure modes in vision-language models beyond manual evaluation.
Sampling parallelism method for efficient Bayesian neural networks and uncertainty quantification in risk-sensitive domains.
Geometric dynamical systems framework explaining LLM hallucinations as arising from basin structure in latent space with task-dependent separability.
Protocol enabling two AI agents to carry out secret conversations while producing transcripts indistinguishable from honest interactions to passive auditors.
Real-world safety evaluation of OpenClaw personal AI agent analyzing attack surface and vulnerabilities in local system access and service integrations.
Method enabling LLMs to learn from hard reasoning problems through adaptive task reformulation with reinforcement learning from verifiable rewards.
Framework for automatically constructing plug-and-play skill knowledge bases for LLM agents to improve learning efficiency and generalization.