Research
GRASP and StochGRASP Cut AI Fine-Tuning Costs for Edge Devices
Two new frameworks slash trainable parameters and boost robustness, enabling smarter AI on limited hardware.
Youtu-Agent Sets New Standards for LLM Agent Frameworks
Youtu-Agent cuts configuration costs and boosts adaptability with automated generation and continuous evolution.
BatteryAgent Advances Lithium-Ion Battery Fault Diagnosis with AI
BatteryAgent blends physical insights and large language models to deliver safer, clearer lithium-ion battery diagnostics.
CREST: Boosting LLM Accuracy and Efficiency Without Retraining
CREST improves large language model reasoning by steering cognitive attention heads, raising accuracy by up to 17.5% and cutting token use by 37.6%.
CogRec: Merging Cognitive Architecture and LLMs to Fix Recommendation Systems
CogRec combines Large Language Models with Soar architecture to deliver clearer, more accurate recommendations.
New System Cuts 3D Mesh Generation to Under One Second for Real-Time Robotics
A breakthrough speeds up 3D mesh creation from a single RGB-D image, enabling robots to perceive and plan in real time with better environmental context.
LongCat ZigZag Attention Boosts AI Efficiency with Sparse Models
LoZA enables AI models to process up to 1 million tokens efficiently, cutting computational costs drastically.
Recursive Language Models: Extending Context Windows Without the Cost
Recursive Language Models (RLMs) break long prompts into chunks, enabling large language models to process far more context efficiently and affordably.
CogRec: Making Recommendation Systems Transparent and Accurate
CogRec blends Large Language Models with Soar to tackle AI’s black-box problem, boosting recommendation clarity and precision.
SPARK: Advancing Personalized Search with Persona-Based LLM Agents
SPARK uses persona-driven agents and multi-agent coordination to deliver sharper, more personalized search results.
CEC-Zero Cuts Chinese Spelling Errors Without Supervision
CEC-Zero uses zero-supervision reinforcement learning to beat traditional Chinese spelling correction methods.
New STED Metric Boosts Consistency in LLM Structured Outputs
Researchers unveil STED, a metric that sharpens consistency in structured outputs from large language models, with Claude-3.7-Sonnet leading performance tests.
Quantum Computing’s Next Target: Error Correction by 2026
Microsoft, Atom Computing, and QuEra push for error-corrected quantum machines using neutral atoms within four years.
SocialVeil Benchmark Reveals LLM Failures in Real-World Social Communication
New research shows large language models stumble when faced with vagueness and emotional noise, exposing gaps in their social understanding.
New TEA Framework Reveals AI Failures in Real-World 3D Tasks
The TEA framework dynamically generates tasks in unseen 3D environments, exposing AI models' struggles with basic perception and interaction beyond standard benchmarks.