Researchers introduce GoldiMask to improve diffusion language model fine-tuning
A new method called GoldiMask optimizes how discrete diffusion language models are fine-tuned by strategically selecting which tokens to reveal as context. Unlike random masking, GoldiMask uses a submodular objective to balance the benefit of visible tokens against their value as prediction targets. It then adjusts the weights of remaining targets based on their learning potential and context…
Key points
- GoldiMask selects tokens to reveal by maximizing a submodular objective balancing context benefit and target value
- Method improves accuracy in reasoning and code generation across three backbones and datasets
- Reduces decoding iterations on GSM8K and MATH-500 without sacrificing accuracy at higher thresholds
The approach, detailed in a paper on arXiv, shows improved accuracy across three model backbones and datasets, particularly in reasoning and code generation tasks. Ablation studies confirm both context selection and target weighting contribute to these gains. GoldiMask also reduces decoding iterations on GSM8K and MATH-500 while keeping accuracy stable at higher confidence thresholds.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Researchers release ArgGYM benchmark for testing defeasible reasoning in AI models · 1 src
- Researchers introduce GFlowNets for diverse synthetic expert conversations · 1 src
- arXiv study finds AI tutors’ guidance varies by impasse type · 1 src
- CARAT study finds materials LLMs often recite rather than reason about crystal structures · 1 src
- Researchers find AI agents can radicalize each other in simulated conversations · 1 src
Comments
via GitHub Discussions