DARTS: Decoder-Aware Representation Tuning via Surgery for Model Merging

Forum topic · 小凯 · 2026-09-01

Summary

DARTS (Decoder-Aware Representation Tuning via Surgery) is a new method for reducing representation bias in merged decoder-only LLMs. Model merging combines multiple task-specific fine-tuned models into one without extra training, but the merged model's hidden states drift systematically from the source models' hidden states. Prior correction work targeted encoder-based vision models, and decoder models pose two unique challenges: causal attention masks cause bias to accumulate across token positions, requiring position-dependent correction, and high-entropy decision-critical positions matter far more than low-entropy ones. DARTS addresses these with an entropy-weighted L1 loss that prioritizes correction at high-entropy positions and a per-position additive bias that captures positional errors without over-parameterization. Evaluated on Llama-2-7B across code generation (HumanEval), math reasoning (GSM8K), and instruction following (AlpacaEval), DARTS achieves significant improvements over standard surgery methods while adding only about 0.1% additional parameters. Paper: arXiv 2608.28547.

Paper Overview

Field: ML Authors: Aaryan Ajay Sharma, Sai Nishanth Padala, Seganrasan Subramanian Published: 2026-08-28 arXiv: 2608.28547

Abstract

Model merging combines multiple task-specific fine-tuned LLMs into a single multi-task model without additional training. However, merged models are known to suffer from representation bias: systematic drift between the merged model's hidden states and those of each individual source model. Prior work (Yang et al., 2024a) studied and mitigated this bias for encoder-based vision models using a lightweight correction module trained with L1 loss. However, such bias had not been studied for decoder models due to their autoregressive nature.

The authors analyze representation bias in decoder models and identify two challenges absent in encoders:

1. The causal attention mask causes bias to accumulate across token positions, requiring position-dependent correction. 2. Not all token positions are equally important — high-entropy (decision-critical) positions are far more important than low-entropy ones.

Method: DARTS

DARTS (Decoder-Aware Representation Tuning via Surgery) tackles these challenges with:

A novel entropy-weighted L1 loss that weights the correction at high-entropy positions, which most affect generation quality.
A per-position additive bias that captures position-dependent errors without over-parameterization.

Results

Extensive evaluation on Llama-2-7B across three domains:

Code generation: HumanEval
Math reasoning: GSM8K
Instruction following: AlpacaEval

DARTS achieves significant improvements over standard surgery methods while adding only negligible parameters (0.1% of total parameters).

--- *Auto-collected on 2026-09-01*