\textsc{DiARC}: Distinguishing Positive and Negative Samples Helps Improving ARC-like Reasoning Ability of Large Language Models

2026-06-24

AI Summary

This work addresses the limited performance of existing open-source large language models on ARC-style abstract reasoning tasks and the high cost of proprietary alternatives, while noting that current approaches predominantly focus on positive example augmentation and largely neglect discriminative modeling of negative samples. To bridge this gap, the study introduces preference alignment into ARC reasoning for the first time, proposing three structured strategies for constructing negative examples—output-level visual transformations, DSL rule inversion, and task-specific rule editing—thereby generating positive-negative sample pairs without altering the original demonstrations. The authors employ a combination of program synthesis and grid transformation techniques to perform supervised fine-tuning on open-source large language models. Experimental results demonstrate that the proposed method significantly outperforms existing baselines across multiple ARC benchmarks, effectively enhancing both discriminative capability and generalization.

Abstract

The Abstraction and Reasoning Corpus (ARC;~\citealp{chollet2019measure}) contains tasks that require summarizing patterns from limited grid samples and predicting output grids. Recently, many large language model based approaches have attempted to transform it into a text-based reasoning task. However, methods based on open-source models have generally yielded unsatisfactory results, while those relying on closed-source models are too costly. Current efforts mainly focus on data augmentation, constructing ARC-like data for more comprehensive supervised fine-tuning. In this work, we argue that solving ARC-like problems requires not only \textit{positive} sample supervision but also the ability to improve model reasoning by distinguishing \textit{negative} samples. To this end, we draw on the idea of preference alignment and propose \textsc{DiARC}, a method that constructs preference pairs to enable the model to distinguish between them. Specifically, we propose three ways to construct negative samples, including output-level visual transformations, DSL-level rule inversion, and task-specific rule editing. The resulting negative samples provide informative near-miss alternatives while keeping the observed demonstrations unchanged. Experimental results across multiple ARC-like benchmarks show that \textsc{DiARC} consistently improves performance over baseline models. The code is released at https://github.com/szu-tera/DiARC.

Problem

• ARC
• negative samples
• reasoning ability
• large language models
• abstraction

Innovation

• preference alignment
• negative sampling
• ARC-like reasoning
• large language models
• abstraction and reasoning