Reproduce “Delta Attention Residuals for Cross-Layer Information Flow in Indic Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Vaibhav Verma, Chintakrindi Kumar Kartik, Sanyam Jain
Delta Block adds learned, token-dependent reads from earlier block-level differences to a Transformer’s usual residual pathway, preserving the running carrier while routing over grouped sources. In Qwen3-0.6B fine-tuned for 10,000 steps on Hindi FineWeb-2, it reaches a 29.968% downstream macro-average versus 29.266% for standard fine-tuning; this single-model, Hindi-only result does not establish broader gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.