Multi-agent LLM coordination framework using entropy-regularized equilibrium selection to improve stability and performance beyond single-model sampling.
Framework for improving LLM reasoning via reinforcement learning by integrating token-level confidence signals from model log-probabilities into RLVR training.
Protein stability prediction using multimodal models combining protein language models with inverse-folding representations, addressing out-of-distribution robustness.
Log anomaly detection system using GPT-Neo (1.3B) fine-tuned with reinforcement learning (PPO) and position-aware rewards for real-time infrastructure monitoring.
Deep learning approach using transformers for capacitance matrix extraction in chip design, addressing limitations of rule-based pattern matching at advanced nodes.
Theoretical framework explaining empirical scaling laws for model performance on multi-domain data mixtures, extending neural scaling law perspectives to multi-domain settings.
Neural field tokenization method with hierarchical and spatial locality priors for efficient feed-forward encoding across modalities.
Platform combining algorithm selection guidance with decision-support systems to help non-experts solve machine learning problems.
Theoretical analysis of compositional Gaussian processes showing depth thresholds and non-Gaussian limits in deep Bayesian models.
Token-conditioned GPT2 model for OLED molecular generation with targeted optical properties in low-data regime.
LLM-based approach for automated discovery of thermodynamically-consistent constitutive models in manufacturing materials.
Differentially private synthetic tabular data generation extending Private Evolution framework to handle high-order correlations.
Method addressing spurious correlations in LLM-based time series forecasting through causal semantic alignment of temporal patterns.
Framework using counterfactual analysis to identify root causes of LLM agent failures by replaying causal decision chains.
System converting natural language queries to execution graphs for multi-tool cross-application data queries using LLM reasoning and deterministic planning.
Graph foundation models for network dynamics with inductive cross-network generalization, demonstrated on super-spreader identification.
Generalization prediction method for deep neural networks using Fourier fractal dimension analysis without validation data.
Wavelet analysis of score-based diffusion models examining how architectural choices (CNNs, U-Nets, Transformers) impact generative behavior.
Framework for predicting sparse autoencoder steering side effects in language models before intervention using feature statistics.
Information-theoretic framework defining open-endedness in AI systems using bit-equivalents to quantify exploration in open-ended environments.
All-atom generative flow model for rapid protein-ligand structure prediction, reducing computational cost of iterative diffusion rollouts.
RiskNet: Large-scale dataset of 5,000+ AI risk incidents from news with multi-dimensional annotations for tracking real-world AI harms and failures.
STAR-KV framework for compressing KV cache in LLMs using adaptive low-rank projection with soft thresholding for fine-grained rank control.
Analyzes optimization bias in Muon optimizer that replaces matrix gradients with polar factors, proving entropy-maximizing properties under alignment assumptions.
Studies identifiability of neural interaction discoveries in time-series models, examining whether discovered interactions reflect data properties or model artifacts.
Sparse rollout optimization for efficient long-context reinforcement learning of LLMs with verifiable rewards balancing stability and speed.
Sleep-inspired replay mechanism for neural networks to prevent catastrophic forgetting when learning sequential tasks without immediate retraining.
Control-theoretic framework for continual learning addressing catastrophic forgetting in nonstationary sequential data streams using drift-plus-penalty.
Method for controlling LLM behavior at inference time using invertible latent transformations beyond linear activation steering approaches.
Study of adversarial attacks targeting confidence calibration in graph neural networks through structural perturbations.
Adaptive loss balancing method for reinforcement learning in generative recommendation systems to handle noisy reward signals from biased rankers.
Benchmark generation system for evaluating text-to-Cypher systems on enterprise property graphs with automated executable query-pair creation.
Meta reinforcement learning approach for autonomous aerial vehicles to perform versatile payload manipulation across variable flight dynamics.
Theoretical analysis of memorization and overfitting in stochastic interpolation models with closed-form solutions for optimal velocity fields.
Tensor network compression method for LLMs that discovers low-rank structures to reduce memory and computational costs through adaptive tensorization.
Speech emotion recognition method using test-time memory mechanisms with pretrained speech-language models for conversational context.
Data pruning framework using dynamic sample selection to reduce training data volume while maintaining performance and unbiased gradient estimation.
Research on fine-tuning large time series models addressing non-convex loss landscape challenges and overfitting during adaptation.
Launcher suite enhancing Codex App with workspace management and workflow control. Related to AI but product-specific, limited technical depth.
VirusTotal partnering with Knostic to analyze VS Code extensions for security threats using AI specialists. AI security application for developer tools.
Reference resource with VRAM tables, GPU tier filtering, and tool-call ratings for LLMs. Useful developer resource with technical specifications.
AI agent that researches LinkedIn profiles, aggregates web data, and generates personalized outreach messages. LLM application for sales automation.
Research paper on configuring agentic AI coding tools. Academic study of AI agent configuration and best practices.
Apple integrates Google's Gemini models into native development framework and Xcode for iOS app development.
Standard for managing API credentials and secrets lifecycle across services. Developer tool addressing credential management complexity.
Graph memory library for AI agents providing structured persistent knowledge storage with vocabulary-driven entity management. Open-source framework for agent memory systems.
Local proxy preventing AI coding agents from leaking secrets via multiple redaction layers. Security tool for LLM applications and agents.
Open-source multi-agent red team collaboration platform for security testing with agent orchestration and runtime model. AI agents for security research.
OpenAI planning ChatGPT overhaul to add coding tools and AI agents, targeting higher revenue before IPO.
Context Ledger technique extends Claude Code agent runtime 28x by compacting git commit history into structured summaries.