Research Papers
Moonshot AI Unveils Kimi K2.5: Open-Source Visual Agentic Intelligence with Swarm Capabilities
Moonshot AI introduces Kimi K2.5, a native multimodal open-source model with advanced coding and vision, featuring a self-directed...
Moonshot AI Unveils WorldVQA Benchmark to Test Visual World Knowledge in Multimodal LLMs
Moonshot AI releases WorldVQA, a benchmark with 3,500 image-question pairs designed to measure factual visual knowledge in multimo...
Kimi Launches Agent Swarm: 100 AI Agents Self-Organize to Tackle Complex Tasks
Kimi (Moonshot) unveils Agent Swarm, a research preview that lets K2.5 deploy up to 100 parallel sub-agents that self-organize int...
Moonshot AI Launches PerceptionBench to Isolate and Measure Atomic Visual Perception in MLLMs
Moonshot AI releases PerceptionBench, a new benchmark that evaluates multimodal large language models on ten atomic visual percept...
MMDiff: New Framework Lets Researchers Isolate and Control Features in Multimodal AI Models
Researchers introduce MMDiff, a framework using multimodal sparse autoencoders to identify and control specific features in multim...
Hugging Face Study: Optimal Data Repetition Scales Mildly with LLM Size
A new paper from Hugging Face reveals that under proportional scaling of model size and training tokens, the optimal repetition of...
Hugging Face Paper: Claim-Level Verification Boosts Reasoning Efficiency
A new training-free method, Claim-Level Reliability Assessment (CLR), improves LLM reasoning accuracy by verifying critical claims...
Second Thought: Parallel Reasoning During Agent Idle Time Cuts Sequential Decoding by Up to 43%
A new training-free framework from Hugging Face, Second Thought, runs auxiliary reasoning branches in parallel while LLM agents wa...
LLMs Develop Brain-Like Modular Architecture, Study Finds
A new preprint shows that large language models spontaneously develop modular neural architectures mirroring the human brain's spe...
PRM-as-a-Judge 1.5: New Toolkit for Fine-Grained Robot Process Evaluation
Hugging Face researchers introduce PRM-as-a-Judge 1.5, a toolkit that evaluates embodied robotic models beyond binary success rate...
Hugging Face Researchers Unveil LOPD: A New Self-Distillation Method That Learns Its Own Teaching Context
Latent On-Policy Self-Distillation (LOPD) makes the teacher's privileged context learnable end-to-end, outperforming existing meth...
DFM Mimir v1: Open 1B Model Sets Danish SOTA Using Only Permissible Data
Hugging Face unveils Mimir v1, a 1-billion-parameter Hierarchical Reasoning Model trained solely on permissible data, delivering c...