Tag: #data-synthesis5 results
OSCToM introduces a novel approach for modeling nested belief conflicts in LLMs, significantly improving performance on complex social reasoning tasks. By combining reinforcement learning with compositional surrogate models, it achieves superior accuracy on information-asymmetric benchmarks like FANToM.
MagicLab hosted GEIS in Silicon Valley, launching MagicBot X1 robot, MagicHand H01 hand, and Magic-Mix world model for embodied AI. Magic-Mix features WAM for physical reasoning and Creator for data synthesis in a closed-loop system. Company eyes $14B revenue by 2036 via nine scenarios.
LOGIGEN is a framework that synthesizes verifiable training data for agentic LLMs using logic-driven methods and triple-agent orchestration. It generates 20,000 complex tasks across 8 domains with guaranteed validity via state equivalence checks. Models trained with SFT and RL achieve 79.5% success on τ²-Bench, far surpassing baselines.
MMKG-RDS is a flexible framework leveraging multimodal knowledge graphs for high-quality reasoning data synthesis, overcoming limitations in coverage, verification, and interpretability. It features fine-grained extraction, customizable sampling, and quality scoring, validated on MMKG-RDS-Bench (5 domains, 17 tasks, 14,950 samples). Fine-tuning Qwen3 models yields 9.2% reasoning accuracy gains; code open-sourced on GitHub.
Feature Activation Coverage (FAC) measures diversity in LLM feature space using sparse autoencoders. FAC Synthesis generates samples targeting missing features from seed data. Boosts diversity and performance on instruction, toxicity, reward, and steering tasks.