TraceForge generates synthetic reasoning data to boost LLM performance

Researchers have developed TraceForge, a novel method for generating synthetic training data that mimics the reasoning style of a target model. This approach bypasses restrictions imposed by some frontier AI labs on using their proprietary outputs for training. TraceForge uses a two-stage supervised and adversarial process to create reasoning traces from context and teacher answers, achieving high authenticity scores. When applied to coding benchmarks, training with TraceForge-generated data significantly improved performance on HumanEval+ and MBPP+ compared to using the teacher model's own reasoning. AI

IMPACT Enables more efficient and legal training of LLMs by synthesizing reasoning traces in the style of the target model.

RANK_REASON The cluster describes a new research paper detailing a novel method for generating synthetic training data for LLMs.

Read on arXiv cs.LG →

• arXiv
• .claude
• generative pre-trained transformer
• Hugging Face
• HumanEval
• MBPP
• Qwen
• Qwen3.8-Max
• TraceForge
• Yan Wang

• paper
• model release

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →