Online sequential optimization for driver subsidies in ride-hailing platforms balancing supply-demand dynamics with multiple constraints.
LLM-based framework for predicting missing survey responses in repeated cross-sectional studies using question/respondent embeddings.
Learning-to-Defer framework allocating extractive QA queries to specialized experts with theoretical guarantees for resource-constrained LLM deployment.
Privacy-preserving language modeling approach to prevent memorization and exposure of sensitive personal information in fine-tuned models.
Framework converting LLM-simulated survey responses into confidence sets with uncertainty quantification for population parameters.
Plugin reasoner module for frozen LLMs that enables reasoning without retraining, compatible across different model architectures.
Study of how instruction-tuned multimodal LLMs align with brain activity patterns under naturalistic stimuli.
arXiv paper on control and optimization methods for neural PDEs in supervised learning contexts.
arXiv paper on machine-learned force fields for lattice dynamics at coupled-cluster accuracy levels.
arXiv paper on efficient long-context LLM inference using retrospective sparse attention to optimize KV cache.
Presents TimeRewarder, a method for learning dense reward signals from unlabeled videos by estimating frame-wise task progress for reinforcement learning.
Analyzes adversarial robustness in learning-to-defer systems where inputs are routed to predictors or experts, extending prior two-stage analyses to one-stage joint training.
Performance study and optimization strategies for multi-node distributed LLM inference across GPU clusters and communication bottlenecks.
Interpretable regression framework combining random Fourier features with localized additive models for heterogeneous data.
Agentic physical AI framework for nuclear reactor control as domain-specific foundation model alternative to general-purpose models.
Sequential knowledge distillation methodology for training compact image compression autoencoders with reduced computational requirements.
Graph imitation learning approach addressing optimization and representation challenges in neural graph generation tasks.
Framework for learning incentive mechanisms enabling cooperative resilience in multi-agent systems under social dilemma conditions.
Pix2Fact benchmark combining fine-grained visual grounding and web knowledge for evaluating vision-language models on high-resolution scenes.
Framework for explaining behavioral shifts in LLMs across scaling, fine-tuning, and RLHF interventions beyond static explainability methods.
Query-aware runtime memory routing system for LLM agents operating across extended context windows with cost-performance tradeoffs.
UG-Separation technique reducing compute costs in large recommendation models by decoupling user and group feature interactions.
Context-aware explainability method using layer-wise integrated gradients to interpret transformer model predictions across network depth.
FT-Dojo benchmark environment enabling autonomous LLM fine-tuning through language agents, automating data curation and model configuration.
Research on timing side-channel vulnerabilities in LLM prefix caching optimization and mitigation strategies for multi-tenant systems.
Framework for routing inputs to experts with dynamic information selection, applicable to LLM systems with retrieved documents and tool outputs.
Praxium system diagnoses cloud microservice anomalies using AI-based telemetry analysis and dependency mapping for CI/CD environments.
Tutor-Student RL framework dynamically optimizes training curriculum for deepfake detection using agent-guided sample prioritization.
Analyzes why self-distillation degrades LLM reasoning in math tasks by suppressing epistemic verbalization of uncertainty.
Profiles energy consumption, performance, and privacy trade-offs for LLM deployment on edge devices with constrained memory and battery.
Demonstrates facial recognition systems exhibit demographic performance disparities, arguing aggregate accuracy inadequate for fairness evaluation in law enforcement.
Evaluates reliability and risk of AI systems in medication decision support, showing performance gaps between standard metrics and real-world deployment.
Proposes improved learning-to-defer systems that move beyond augmented-action surrogates for multi-expert decision routing.
JoyAI-Image unified multimodal foundation model combines spatially-enhanced MLLM with diffusion transformer for visual understanding and image generation.
Benchmarks JPEG decoder performance across multiple CPU architectures for ML data loaders, revealing limitations of single-thread microbenchmarks.
Comprehensive review of LLM applications in hardware design automation and electronic design, covering RTL code generation, testbenches, and security vulnerabilities.
First online learning-to-defer algorithm for multiclass classification with dynamically varying expert pools and streaming data.
GAP method improves visual reasoning in multimodal LLMs by addressing feature-space mismatches in latent reasoning without external tools.
MeMo framework encodes new knowledge into dedicated memory models while keeping LLMs frozen, enabling efficient incorporation of domain-specific information without retraining.
Research on visuomotor policies for robotic manipulation using chunked action distributions and flow matching to improve trajectory coherence.
Charon: unified simulator for large-scale LLM training and inference with modular architecture for performance prediction.
SafeRun provides replay debugging and inline prevention for AI agents. SDK available in Python and TypeScript with sub-50ms API latency. Currently onboarding design partners.
Agent Chat Bridge enables async callbacks for AI IDE agents. Register timers, shell commands, or webhooks to resume agent sessions. Supports VS Code Copilot Chat and Windsurf Cascade.
TBN Protocol provides runtime governance infrastructure for AI agents. Includes automated guardrails and cost-control mechanisms with 14-step flow visualization.
Analysis of production issues in AI-built SaaS. Case study of tenant isolation bug missed in code review due to identical patterns.
SafeRun provides replay debugging and inline prevention for AI agents. SDK available in Python and TypeScript with sub-50ms API latency. Currently onboarding design partners.
Graph-based semantic memory system for AI coding agents. Local-first persistent memory using SQLite, llama.cpp, and BGE-M3. Compatible with Claude, Codex, ChatGPT.
AI system generates expert-level empirical software for scientific computing, addressing software bottlenecks in discovery cycles.
InferenceBench benchmark evaluates AI agents optimizing LLM serving workloads under compute constraints. Agents outperform baselines but underperform tuned hyperparameters.
Doubleword optimizes Mixture-of-Experts inference throughput by 15% via request reordering to reduce memory-bandwidth bottlenecks in batch processing.