Tag
This paper introduces GSM-SEM, a framework for generating semantically diverse benchmark variants to mitigate memorization in mathematical reasoning evaluations. The authors demonstrate that this approach reveals significant performance drops in current SOTA LLMs compared to static benchmarks.
Proposes TAP, a tabular augmentation policy that couples diffusion inpainting with a learner-conditioned policy to improve downstream model performance under data scarcity, outperforming strong baselines on real-world datasets.
This paper investigates how informal text (slang, emoji, Gen-Z filler tokens) degrades NLI accuracy in ELECTRA-small and RoBERTa-large models, identifying two distinct failure mechanisms—tokenization failure (emoji mapped to [UNK]) and distribution shift (out-of-domain noise tokens)—and proposes targeted mitigations that recover accuracy without harming clean-text performance.
OpenAI presents a simple data augmentation technique that enables autoregressive language models to perform fill-in-the-middle (FIM) text generation without harming left-to-right performance, with extensive ablations and best practices provided for training such models.
A paper reviewing Kornia's differentiable data augmentation module for 2D and 3D tensors, integrating with PyTorch's autograd and optim, along with benchmarks comparing data augmentation frameworks.