Training-time information allocation view of implicit bias in neural networks; explains how optimization forms writing patterns for error signals during training.
PeTeR: post-training robustification of probabilistic circuits via distributionally-robust optimization to improve generalization under noise and distribution shift.
Omni-Sleep: foundation model for sleep staging using hierarchical contrastive learning on multimodal polysomnography signals (EEG, ECG, respiration).
Jet-Long: efficient long-context extension for LLMs using dynamic bifocal RoPE, enables zero-shot context extension 10x beyond pretraining window.
Budget-aware test-time selection for routing LLM queries across multiple models; resampling strategy improves cost-quality tradeoff vs single-router approach.
Fourier analytic method for learning Gaussian mixture models with theoretical guarantees on separating mixture components under distance constraints.
Ruby: ML-based tool to detect unsafe Rust regions in stripped binaries without source code access; addresses memory safety analysis gap.
Contrastive learning approach for multimodal EHR analysis combining structured codes and unstructured clinical notes via deep learning.
arXiv paper on accelerated first-order methods for bilevel and minimax optimization.
arXiv paper on universal approximation theorems for operators in Banach spaces using projections.
arXiv paper on accelerating LLM inference via self-supervised early exits at intermediate layers.
arXiv paper on inference-time intervention for multi-attribute steering of LLMs without parameter updates.
Transformer-based actor-critic reinforcement learning for optimizing Service Function Chain partitioning in 6G network virtualization.
Analysis of evasion mechanisms against white-box LLM monitors and principled defenses to prevent circumvention of safety auditing tools.
Research on how human beliefs about agent capabilities affect preference generation in RLHF, with descriptive and normative analysis.
M4V uses Mamba architecture for efficient text-to-video generation with linear-time sequence modeling instead of transformer quadratic complexity.
PCS-guided neighbor embedding method for extracting smooth, low-dimensional representations from high-dimensional single-cell sequencing data.
Interpretable role-based clustering approach for multi-layer financial networks to identify functional positions of institutions across market segments.
Framework for evaluating data efficiency of quantum machine learning models using semi-artificial datasets tailored to quantum kernel methods.
Algorithm for estimating unknown Lipschitz functions from noisy observations using delta-convex functions and max-affine methods.
AutoGraphAD uses Variational Graph Autoencoders for unsupervised network anomaly detection without requiring labeled datasets.
Pipeline for converting Zoom recordings into speaker-attributed transcripts to enable LLM-based simulations of civic deliberation with stable participant behavior modeling.
Practical guide for generating high-quality synthetic data using differential privacy to address data scarcity and representativeness issues in AI.
ReinforceGen system combining task decomposition, data generation, imitation learning, and reinforcement learning for long-horizon robotic manipulation.
Transition Matching Distillation framework for distilling video diffusion models into efficient few-step generators for real-time applications.
RELISH: lightweight architecture for text regression using frozen LLM representations with iterative latent state refinement.
HiPO: hierarchical preference optimization for LLMs enabling fine-grained feedback on multi-step reasoning tasks.
Concept-based explainability evaluation of multimodal LLMs for in-context learning visual classification tasks.
Kernel of Partition Paths: unified geometric representation for tree ensemble models enabling oracle inequalities.
Memory-Managed Long-Context Attention: novel mechanism for bounded memory in LLMs with lifecycle control and sparse fallback.
Sheaf: Open-source ML framework for Clojure using code-as-data paradigm with composable functional models, automatic GPU compilation, and no parameter registration.
Open-source voice AI agent for Asterisk/FreePBX with modular STT/LLM/TTS pipeline. Production-ready baselines with admin UI.
Live2D desktop pet UI for Hermes AI agents with voice lip-sync and HTTP API. Agent interface tool using Tauri and pixi-live2d.
Baton is a local command center for managing multiple AI coding agents in flight. Monitors agent status and prioritizes which needs attention.
Framework analyzing self-improving agent systems as compound systems where improvements come from modifying prompts, tools, workflows, memory, and evaluators rather than weight rewriting.
Study showing people cannot reliably identify AI-generated poetry but rate it lower when told the source is AI.
System for building user personas from behavior data that improve over time by scoring predictions. Built for Claude Code, uses markdown files as agent and skill definitions.
Autonomous pipeline using Claude to programmatically control After Effects via WebSocket, building and rendering video clips with self-correction through rendering and revision cycles.
Embedcache deduplicates and caches embedding API requests, reducing costs. Pure Go proxy compatible with vLLM, Ollama, OpenAI.
Optimized C/CUDA inference engine for Qwen 3.6 35B on RTX 5090, claims faster performance than llama.cpp.
peek-cli browser extension enables AI agents to iterate on frontend designs via screenshots. Works with Claude, Copilot, Codex.
OneDev integrates AI agents as teammates in development workflow: issues, PRs, reviews, CI/CD.
Lyzr startup used its own SivaClaw AI agent to manage $100M Series B fundraising from 130+ investors.
Argo CD UI extension with AI assistant for Kubernetes resources. Supports OpenAI-compatible backends like Ollama, vLLM.
Benchmark comparing LLM performance on PR security vulnerability detection, finding GPT5.6 and Grok 4.6 outperform Anthropic models on this workload.
Prescryb MCP server for CVE remediation and configuration management. Natural language interface for security fixes via Claude/MCP clients.
Compares LLM chat interfaces to graphics pipelines, argues for more flexible LLM interfaces.
Context-sharing memory platform building knowledge graphs from team tools to provide agents and AI systems with relevant institutional context.
Teaching repository: training 113M parameter earthquake-domain LLM from scratch using nanoGPT, covers entire lifecycle on 2× A30 GPUs.
Almanac auto-generates and updates wikis from uploaded files. Agent-native design for Claude/Codex integration.