Reproduce “Feed-Forward Steering in Transformer Residual Dynamics” 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
Timur Mudarisov, Mikhail Burtsev, Radu State
This research extends continuous dynamical theories of Transformer residual streams by formally integrating the Feed-Forward Network (FFN) component, modeling it as a local steering field that complements attention's non-local aggregation. Experiments demonstrate that FFNs primarily contribute to residual dynamics through tangential steering, counteract attention-induced diversity collapse, and offer a practical diagnostic for parallelization optimizations.
Key insights
- The FFN's primary contribution to modifying residual directions is mediated by its tangential steering component, rather than radial magnitude adjustments, confirming its role as a directional modifier.
- FFNs play a critical role in maintaining the spherical diversity of token representations by counteracting the aggregative effects of attention, preventing over-concentration.
- The "sequential-to-parallel defect" and FFN input-sensitivity (`C_ℓ`) can effectively predict which Transformer layers are suitable for parallelization, offering a practical pathway for architectural optimization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.