English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Tools annual

Global: Zendy Free Plan

Global: Zendy Tools monthly

Global: Zendy Plus monthly

Adapting large language models to multiple tasks can cause cross-skillinterference, where improvements for one skill degrade another. While methodssuch as LoRA impose orthogonality constraints at the weight level, they do notfully address interference in hidden-state representations. We proposeCompositional Subspace Representation Fine-tuning (CS-ReFT), a novelrepresentation-based approach that learns multiple orthonormal subspacetransformations, each specializing in a distinct skill, and composes them via alightweight router. By isolating these subspace edits in the hidden state,rather than weight matrices, CS-ReFT prevents cross-task conflicts moreeffectively. On the AlpacaEval benchmark, applying CS-ReFT to Llama-2-7Bachieves a 93.94% win rate, surpassing GPT-3.5 Turbo (86.30%) while requiringonly 0.0098% of model parameters. These findings show that specializedrepresentation edits, composed via a simple router, significantly enhancemulti-task instruction following with minimal overhead.

Compositional Subspace Representation Fine-tuning for Adaptive Large Language Models