MAPUS: LLM-based multi-agent framework for personalized and fair participatory urban sensing modeling participants as autonomous agents with preferences.
ELITE framework for self-improving embodied agents using vision-language models with experiential learning and intent-aware transfer to bridge vision-action gap.
Enhanced Mycelium of Thought (EMoT): bio-inspired hierarchical reasoning architecture for LLMs with four-level hierarchy, strategic dormancy, and mnemonic encoding.
AutoProf: multi-agent orchestration framework for autonomous AI research with persistent world model, gap analysis, and inter-agent verification mechanisms.
Multi-agent framework with specialist agents for medical multiple-choice question answering, improving calibration and confidence scoring through verification.
Markovian framework for auditing reliability and oversight costs in agentic AI systems operating as stochastic policies with sequential decisions and tool calls.
Analysis of many-shot jailbreaking technique exploiting long context windows; probes effectiveness and develops mitigation strategies for LLM safety.
Novel methodology quantitatively evaluating metacognitive abilities in LLMs, testing self-awareness without relying on model self-reports.
Computerized Adaptive Testing framework grounded in Item Response Theory for cost-effective and scalable evaluation of LLMs in medical benchmarking.
Deletion-Insertion Diffusion language models replacing masking paradigm with discrete diffusion processes for improved computational efficiency and generation flexibility.
Internal Safety Collapse (ISC) failure mode identified in frontier LLMs where models generate harmful content under certain task conditions; TVD framework presented to trigger and study ISC.
Evaluation of visuospatial perspective-taking abilities in multimodal language models using adapted tasks from human studies (Director Task, Rotating F task).
DISCO benchmark suite for evaluating OCR pipelines and vision-language models on document parsing and QA across diverse document types including handwritten and multilingual text.
S-Path-RAG framework for multi-hop question answering over knowledge graphs using semantic-aware shortest-path retrieval with differentiable path scoring.
Berta: open-source modular platform for AI-enabled clinical documentation with institutional data governance and workflow integration, deployed at Alberta Health Services.
DepthCharge framework for measuring how deeply LLMs sustain accurate responses in domain-specific topics through adaptive probing across arbitrary domains.
Privacy-preserving synthetic clinical data trains LLM for medical coding automation, improving ICD-10-CM and CPT code assignment from clinical documentation.
Memory Sparse Attention enables end-to-end LLM scaling to 100M tokens for long-term memory tasks, extending effective context beyond 1M token limits.
Position paper proposes mechanism-aware evaluation combining symbolic rules and mechanistic interpretability to distinguish genuine generalization from shortcuts.
Cluster-R1 reframes instruction-following clustering as generative task, enabling reasoning models to autonomously infer corpus structure while respecting user instructions.
MedMT-Bench stress-tests LLMs on long-context memory, interference robustness, and safety in multi-turn medical conversations with realistic clinical scenarios.
Lightweight LLM framework captures and scales physician expertise for clinical decision-making agents using individualized diagnostic methodologies.
Chitrakshara multimodal dataset provides multi-image and Indian language coverage for training Vision-Language Models beyond English-centric datasets.
Qworld framework generates question-specific evaluation criteria for LLMs on open-ended tasks, capturing context-dependent response quality requirements.
ConceptMap tool enables scalable exploratory discovery of human-interpretable concepts in sparse autoencoders trained on LLM activations.
Konkani-Instruct-100k synthetic dataset and benchmarks address LLM performance gaps for low-resource Indian language across multiple scripts via instruction tuning.
Cognitive psychology-inspired study reveals LLMs drop formatting instruction compliance by 2-21% under concurrent task load, identifying prospective memory vulnerabilities.
Fine-tuned lightweight LLM generates hierarchical JSON representations of scientific sentences preserving semantic meaning for structured knowledge extraction.
MDKeyChunker pipeline enables structure-aware chunking of Markdown documents and single-call LLM enrichment with metadata extraction for improved RAG accuracy.
Mixture of Demonstrations approach improves GraphRAG performance for domain-specific QA by selecting high-quality demonstrations to reduce irrelevant retrieved information.
Native GUI agent framework ReCAP adds CAPTCHA-solving capability to vision-language models using self-corrective training and automated reasoning-action data generation.
Synthetic Mixed Training combines synthetic QAs and documents to improve LLM knowledge acquisition beyond RAG performance in data-constrained domains.
Safe reinforcement learning approach using preference-based constraint inference for learning complex, subjective safety constraints with minimal expert demonstrations.
AI agent optimizes operator performance on Huawei Ascend NPUs by addressing knowledge bottleneck through episodic learning for tiling and kernel programs.
StateLinFormer: linear-attention navigation model with persistent memory for long-term navigation tasks, combining flexibility with efficiency.
Dual-Criterion Curriculum Learning proposes a meta-learning approach using dual criteria for difficulty assessment in temporal data training.
APreQEL proposes adaptive mixed precision quantization to reduce memory and computational costs of LLMs for edge device deployment while maintaining performance.
Time-LLM model for predicting wafer-level spatial etch depth distributions in plasma etching process monitoring.
LLMORPH automated testing tool for LLMs using metamorphic testing to detect NLP task failures without human-labeled oracles.
LLMLOOP framework automating iterative refinement of LLM-generated code and test cases through automated feedback loops.
Theory of LLM information susceptibility analyzing fundamental limits of LLM-mediated optimization in agentic systems.
Swiss-Bench SBP-002: trilingual benchmark of 395 expert-crafted regulatory compliance tasks across FINMA, Legal-CH, and EFK domains.
Probing study revealing how LLMs internally represent different ethical frameworks with asymmetric transfer patterns across model sizes.
Training-free out-of-distribution detection using multi-layer prototype fusion approach for robust deep learning deployment.
Privacy-preserving LLM system for disambiguating clinical acronyms in healthcare without transmitting data to external servers.
Measurement methodology for identifying assessment items where LLMs perform differently than humans using theory-grounded evaluation.
Analysis of early-exit decoding in modern LLMs showing reduced efficiency gains due to improved architectures with lower layer redundancy.
Study of filtered vector search algorithms in PostgreSQL for semantic search and GenAI applications, evaluating real-world database performance.
Self-paced curriculum learning for RL using closed-form Gaussian updates to improve efficiency in high-dimensional contexts.
Intent-Based Networking using AI to translate high-level natural language intents into network policies with automated compliance assurance.