DASA
An LLM fine-tuning method that optimizes continuous synthetic input embeddings and uses activation-gradient feedback from a frozen reference model to guide data generation. By bypassing human-readable text constraints, it aligns desired model update directions directly in embedding space.
Timeline (last 90 days)
An arXiv paper (arXiv:2609.35868) proposes a method called DASA (Desired-Update-Aligned Synthetic Data) for bypassing human-readable text constraints during LLM fine-tuning.
DASA optimizes a set of continuous synthetic input embeddings and uses them directly for downstream fine-tuning, rather than using human-readable text.
DASA leverages activation gradient feedback from a frozen reference model to guide the optimization of continuous synthetic input embeddings.
In the DASA pipeline, projecting continuous embeddings back to discrete tokens is used only for qualitative inspection, not as training input.
DASA's experiments were validated on 6 models from the Llama and Qwen model families, covering parameter scales from 1B to 32B.
DASA's evaluation benchmark covers six aspects including knowledge tasks, mathematical reasoning, code generation, and commonsense reasoning.
DASA experiments are conducted under a unified LoRA adaptation setting.
DASA reaches a level comparable to source natural language data and surpasses the original text data in multiple configurations.
DASA outperforms the baseline method GRADMM in most comparisons.
Under the synthetic settings evaluated, DASA provides a 3.6 to 4.9× speedup compared to GRADMM, with comparable peak GPU memory usage.
2 more timeline events
All Facts (12)
The conclusions of the DASA paper are limited to the synthetic settings evaluated, and its generalizability remains to be confirmed through broader replication.
50%UnverifiedDASA optimizes a set of continuous synthetic input embeddings and uses them directly for downstream fine-tuning, rather than using human-readable text.
50%UnverifiedDASA leverages activation gradient feedback from a frozen reference model to guide the optimization of continuous synthetic input embeddings.
50%UnverifiedIn the DASA pipeline, projecting continuous embeddings back to discrete tokens is used only for qualitative inspection, not as training input.
50%UnverifiedDASA's experiments were validated on 6 models from the Llama and Qwen model families, covering parameter scales from 1B to 32B.
50%UnverifiedDASA's evaluation benchmark covers six aspects including knowledge tasks, mathematical reasoning, code generation, and commonsense reasoning.
50%UnverifiedDASA experiments are conducted under a unified LoRA adaptation setting.
50%UnverifiedDASA reaches a level comparable to source natural language data and surpasses the original text data in multiple configurations.
50%UnverifiedDASA outperforms the baseline method GRADMM in most comparisons.
50%UnverifiedUnder the synthetic settings evaluated, DASA provides a 3.6 to 4.9× speedup compared to GRADMM, with comparable peak GPU memory usage.
50%UnverifiedDASA synthetic data lacks interpretability, and projecting back to tokens is only used for qualitative inspection, which may pose governance challenges in scenarios requiring auditability and controllability.
50%UnverifiedAn arXiv paper (arXiv:2609.35868) proposes a method called DASA (Desired-Update-Aligned Synthetic Data) for bypassing human-readable text constraints during LLM fine-tuning.
50%