Engineering
Researchers Propose “Graph Engineering” as a New Paradigm for Multi-Agent Systems
A new paper argues that individual AI agents hit a ceiling on tasks requiring parallel execution, specialized expertise, and persistent…
Engineering
ParaTempo Cuts LLM Reasoning Latency by Up to 32% Using “Temporal Confidence”
Researchers introduced ParaTempo, a framework that reduces the computational cost of parallel reasoning by measuring “temporal confidence” — how consistently…
Engineering
New Training-Free Sparse Attention Method Targets Video Generation and World Models
Researchers introduced SparsePR, a training-free block-sparse attention method for video transformers and world models that combines Response-Coupled Partitioning with Probe-Fitted…
Engineering
AgentMercury Generates Thousands of Business Environments to Train Enterprise AI Agents
Researchers introduced AgentMercury, a framework that synthesizes persistent, executable business environments — complete with entities, services, and cross-service rules —…
Engineering
Llama-Mobile Squeezes an 11B Vision-Language Model Down to 3.7GB for Phones
Researchers developed a quantization framework for deploying large vision-language models on memory-constrained mobile hardware, combining a self-generated training data pipeline…
Engineering
CLEAR Cuts Harmful LLM Responses From 32% to 0.5% Without Hurting Performance
Researchers developed CLEAR, a technique that uses a lightweight hidden-state gate to continuously control the activation strength of a safety…
Engineering
FACET Improves How Terminal AI Agents Are Trained by Keeping Tasks Internally Consistent
Researchers introduced FACET, a framework for generating training tasks for terminal/command-line agents that keeps instructions, solutions, and verifiers aligned by…
Engineering
Study Finds LLMs Cost Up to 1,431x More Than Embedding Models for Similar Accuracy
A cost-aware evaluation compared large language models against specialized embedding models across 37 text-understanding tasks and found the best LLM…
Engineering
New Post-Training Method Teaches LLMs Document Knowledge Without Retrieval at Inference Time
Researchers developed Inject, Align, and Recover (IAR), a three-stage post-training framework that lets language models answer questions about specific document…
Engineering
PolicyGuide Nearly Doubles Policy-Compliance Rates for Customer-Service AI Agents
Researchers built PolicyGuide, a system that compiles organizational policies into workflow graphs and uses a proactive verifier at user-turn boundaries…
Engineering
Daedalus-150M Is a Language Model Built From Scratch for Fast CPU Inference
Researchers built Daedalus-150M, a 150-million-parameter language model architected specifically for efficient CPU inference rather than adapted from a GPU-designed model.…
Engineering
Chain-of-Experience Lets LLMs Improve at Test Time Through Self-Feedback
Researchers introduced Chain-of-Experience (CoE), a test-time learning approach that allows large language models to improve iteratively using self-feedback and environmental…