AI Reasoning Doubled with Just One Example, 1-bit Model Achieving 4-bit Precision
Expanding Applications from Multimodal Generation Across Video and Speech Spaces to AI Evaluation Framework Reform

This week''s META-X AI paper review covers reasoning optimization, multimodal advances, and AI evaluation reform.

AI Reasoning and Efficiency Optimization: 1-Shot RLVR demonstrates that RL with just one training example dramatically improves reasoning ability — boosting MATH500 benchmark from 36% to 73.6% using a single math problem example, matching results from thousands of examples; entropy loss addition during training is key. ReasonIR develops specialized information retrievers supporting complex reasoning processes. Skywork R1V2 advances multimodal model reasoning capabilities. BitNet v2 maximizes memory and computation efficiency by quantizing 1-bit language model activations to 4-bit precision.

Multimodal Perception and Generation: "Camera Motions" enables deep understanding and analysis of dynamic camera movements in video. UniversalRAG develops next-generation RAG technology integrating diverse information sources (text, image, video) with varying granularity. TesserAct learns 4D world models capturing temporal changes in 3D space. In-Context Edit enables precise image editing following natural language instructions. KeySync realistically generates mouth shapes in video synchronized to audio. Spatial Speech Translation translates multi-speaker speech while preserving spatial audio information in real-time.

AI Evaluation and Reliability: "Leaderboard Illusion" critically analyzes potential biases and structural problems in widely used AI performance leaderboards, proposing improvement directions. Sadeed develops specialized lightweight models for Arabic diacritization while building new benchmark datasets — exemplifying language-specific AI specialization approach.