SGLang implements optimized pipeline parallelism for million-token context inference using chunked parallelism, asynchronous communication, and dynamic chunking mechanisms.
MCP server that compresses tool outputs for Claude Code by processing them in sandboxes and returning summaries, reducing 315 KB context to 5.4 KB and extending session time from 30 min to 3 hours.
Platform where AI agents make predictions on factual claims scored by temporal accuracy rather than self-reporting, using Temporal Knowledge Graph.
vLLM and Nvidia achieved +38% throughput and +13% latency improvements on GPT-OSS-120B using Blackwell GPUs via FlashInfer and torch.compile optimizations.
Local AI agent with persistent memory using River Algorithm to store facts, timelines, and confidence levels without cloud dependency.
AgentPass provides cryptographic identity and authentication infrastructure for autonomous AI agents. Identity layer enabling agent autonomy.
Relay product for efficient agent context management via ephemeral/durable classification. Token optimization for autonomous agents.
Standard for tracking human-AI creative control attribution in code generation. Developer tool for documentation and collaboration.
vLLM optimization achieving 26.2K prefill and 10.1K decode throughput on NVIDIA Blackwell for DeepSeek MoE models. Production LLM serving research.
Benchmark for evaluating LLM vision and tool-use on cursor control task based on Neuralink's Webgrid test. Measures multimodal agent capabilities.
Multi-agent autonomous workflow built Game Boy GBA emulator in 48 hours using TypeScript. Demonstrates agent orchestration and tool use.
DataClaw tool exports Claude Code conversation history to HuggingFace datasets. Data extraction tool with open source intent.
News article about Anthropic accusing Chinese AI firms of model distillation for IP theft. Industry news, not technical content.
Semantic parallelism optimization for efficient MoE model inference across multiple devices.
Diffusion-based recommendation system using continuous tokens with LLM integration.
Evaluation of LLM accuracy for health advice across languages and contextual factors.
Multi-agent system using LLMs to detect vulnerabilities in hardware RTL design specifications.
Multimodal LLM combining vision, audio, and sensor data for embodied AI agents in smart homes.
Benchmark for evaluating multimodal LLM capabilities in humanities and social sciences requiring interdisciplinary reasoning.
Comprehensive benchmark for evaluating multimodal LLMs on front-end code generation from visual designs using modern frameworks.
Framework addressing dependency, asynchrony, and missing values in multivariate time series forecasting from real-world data.
LLM-based framework for evaluating children's language function through phonetic transcription and automated speech assessment.
Characterization and comparison of State Space Models and hybrid architectures versus Transformers for long-context processing on edge devices.
Monte Carlo tree diffusion with multiple experts for protein sequence design combining masked diffusion models with tree search.
Framework using LLMs as spatio-temporal predictors with hierarchical temporal tokenization for human mobility and trajectory prediction.
Reinforcement learning fine-tuning approach using polychromic objectives to prevent policy collapse and maintain behavioral diversity.
Learnable dynamic routing mechanism for mixture-of-experts with LoRA adapters enabling efficient LLM task adaptation without fixed expert assignment.
Policy optimization method for text-to-image models addressing credit assignment instability in reinforcement learning fine-tuning.
Evaluation of Vision-Language-Action model robustness against multi-modal perturbations across 17 adversarial conditions.
LLM-based agent framework for recommendation systems that leverages commonsense reasoning to capture item relationships and user intent.
Study on how LLM-generated rationales influence human plausibility judgments in commonsense reasoning tasks using 3,000 human and 13,600 LLM judgments.
Latent-Augmented Discrete Diffusion model improving fast language generation by modeling cross-token dependencies.
Method for scalable AI oversight via partitioned human supervision across multiple domain experts for complex multi-domain tasks.
Security evaluation framework for LLM backbone models in AI agents, addressing vulnerabilities unique to agent architectures.
Survey examining terminology, definitions, and taxonomy of data agents—autonomous systems orchestrating data and AI for complex data tasks.
LLM-based search agents trained on synthetic entity-centric data using improved reward mechanisms to capture informative near-miss samples.
OckBench: Benchmark measuring LLM reasoning efficiency via token usage, revealing up to 5x differences in token length across models.
Data-efficient fine-tuning strategy for adding controllable parameters to text-to-video diffusion models using synthetic data.
Refusal Steering: Inference-time method for fine-grained control over LLM refusal behavior on sensitive topics without retraining.
HiGR: Generative slate recommendation system using hierarchical planning and multi-objective preference alignment for ranked lists.
CogFlow: Multimodal LLM system for visual math problem solving improving visual perception integration and reasoning.
Comprehensive empirical study evaluating factors affecting safety alignment in LLMs and LRMs across 32 recent models.
Fast-ThinkAct: Efficient Vision-Language-Action framework reducing inference latency through verbalizable latent planning.
CLiMB: Domain-informed clustering framework for novelty detection in scientific discovery with application to galactic archaeology.
Molmo2: Open-weight vision-language model with video understanding, grounding, and disclosed training data and recipe.
FROST: Attention-aware pruning method for efficient LLM reasoning by identifying and removing reasoning outliers while preserving capacity.
Persona Brainstorm Audit method for detecting bias and fairness issues in open-ended creative outputs from LLMs.
Study showing SGD with sparsity outperforms Adam for RL from verifiable reward in LLM training, challenging standard optimization practices.
AceGRPO: Reinforcement learning agent for autonomous ML engineering using adaptive curriculum and group relative policy optimization to overcome parameter freezing.
Study on paraphrase generation and detection as mechanisms for language understanding and modeling in neural networks.