Emad

Arbor: A Tree Search Cognition Layer Enhances Autonomous Agent Performance in LLM Inference Optimization

Arbor is a novel multi-agent framework that introduces structured tree search as a cognition layer for autonomous agents, specifically targeting large, stateful action spaces like LLM inference optimization. It maintains a shared search tree as working memory, adapting exploration based on successes and failures. Its architecture, featuring Orchestrator and Critic agents with distinct hard and soft skills, enables autonomous, multi-day optimization campaigns.

Efficient Diffusion LLMs via Temporal-Spatial Parallel Decoding and Confidence Extrapolation

PR2: Predictive Routing Replay for MoE-Based LLM Reinforcement Learning

Pause and Think: A Dataset and Benchmark for Video-Grounded Assistive Action Suggestion

KG-FairDiff: Knowledge Graph-Guided Prompt Refinement for Demographically Fair Text-to-Image Generation

Beyond the Target: From Imitation to Collaboration in Speculative Decoding

A Unified KLD Framework for Duplexity and Deployment Paradigms in Cell-Free mMIMO-ISAC

Voltage and Frequency Stability Analysis of Transmission Power Grids with EV Charging Stations

Approximate mass spectra of the heavy mesons under a Coulomb plus logarithmic spin-dependent potential function

AgentKernelArena: Generalization-Aware Benchmarking of GPU Kernel Optimization Agents

Approximate mass spectra of the heavy mesons under a Coulomb plus logarithmic spin-dependent potential function

GradShield: Alignment Preserving Finetuning

Ego-InBetween: Generating Object State Transitions in Ego-Centric Videos

ARFBench: Benchmarking Time Series Question Answering Ability for Software Incident Response

CCL-Bench 1.0: A Trace-Based Benchmark for LLM Infrastructure

PARD-2: Target-Aligned Parallel Draft Model for Dual-Mode Speculative Decoding