Pith. sign in

REVIEW 1 cited by

Compositional Subspace Representation Fine-tuning for Adaptive Large Language Models

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2503.10617 v3 pith:ZC3GBRZQ submitted 2025-03-13 cs.CL cs.AI

Compositional Subspace Representation Fine-tuning for Adaptive Large Language Models

classification cs.CL cs.AI
keywords subspacecs-reftrepresentationcompositionaleditsfine-tuninginterferencelanguage
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Adapting large language models to multiple tasks can cause cross-skill interference, where improvements for one skill degrade another. While methods such as LoRA impose orthogonality constraints at the weight level, they do not fully address interference in hidden-state representations. We propose Compositional Subspace Representation Fine-tuning (CS-ReFT), a novel representation-based approach that learns multiple orthonormal subspace transformations, each specializing in a distinct skill, and composes them via a lightweight router. By isolating these subspace edits in the hidden state, rather than weight matrices, CS-ReFT prevents cross-task conflicts more effectively. On the AlpacaEval benchmark, applying CS-ReFT to Llama-2-7B achieves a 93.94% win rate, surpassing GPT-3.5 Turbo (86.30%) while requiring only 0.0098% of model parameters. These findings show that specialized representation edits, composed via a simple router, significantly enhance multi-task instruction following with minimal overhead.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Gate-and-Merge: Zero-shot Compositional Personalization of Vision Language Models

    cs.CV 2026-05 unverdicted novelty 6.0

    Gate-and-Merge enables zero-shot compositional personalization of VLMs by independently learning concept-specific LoRA adapters and merging them in weight space with cue-based gating to suppress interference.