Tag
This arXiv paper proposes Desired-Update-Aligned Synthetic Data (DASA), which uses activation-gradient feedback from a frozen reference model to optimize continuous synthetic input embeddings for LLM fine-tuning, questioning whether human-readable text is necessary. Experiments across Llama and Qwen models (1B–32B) show DASA matches or outperforms natural-language data under LoRA and delivers a 3.6–4.9x speedup over GRADMM.
CopT introduces a contrastive on-policy thinking framework for LLMs that generates draft answers first, then uses contrastive verification and dynamic thinking to improve accuracy while reducing token consumption, achieving up to 23% higher accuracy and 57% lower token usage on math, coding, and agentic reasoning tasks.
ELF proposes a continuous diffusion model for language that uses embedding space and flow matching, outperforming existing discrete and continuous diffusion language models with fewer sampling steps.