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

2026-06-25 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

DiARC is a novel method designed to enhance the Abstraction and Reasoning Corpus (ARC)-like reasoning capabilities of large language models (LLMs). Addressing the limitations of current LLM approaches, which are either unsatisfactory for open-source models or costly for closed-source ones, DiARC moves beyond traditional data augmentation. It posits that improving ARC-like problem-solving requires not only positive sample supervision but also the ability to distinguish negative samples. Drawing inspiration from preference alignment, DiARC constructs preference pairs. It introduces three specific techniques for generating negative samples: output-level visual transformations, DSL-level rule inversion, and task-specific rule editing. These methods create informative "near-miss" alternatives while preserving original demonstrations. Experimental results demonstrate that DiARC consistently improves performance across various ARC-like benchmarks. The project's code is publicly available at https://github.com/szu-tera/DiARC.

Key takeaway

For Research Scientists developing LLMs for complex reasoning tasks like those in the Abstraction and Reasoning Corpus, you should integrate negative sample distinction into your training methodology. Current data augmentation alone is insufficient. Instead, explore preference alignment techniques to enable your models to differentiate between correct and "near-miss" incorrect outputs. Implement output-level visual transformations, DSL-level rule inversion, or task-specific rule editing to construct informative negative samples. This approach consistently improves model performance.

Key insights

Distinguishing negative samples via preference alignment improves LLM reasoning on ARC-like tasks.

Principles

• Reasoning improvement requires distinguishing negative samples.
• Preference alignment can enhance model reasoning.

Method

DiARC constructs preference pairs by generating negative samples through output-level visual transformations, DSL-level rule inversion, or task-specific rule editing.

In practice

• Apply preference alignment to improve LLM reasoning.
• Generate near-miss negative samples for complex tasks.

Topics

• Large Language Models
• ARC Reasoning
• Preference Alignment
• Negative Sampling
• Data Augmentation

Code references

• szu-tera/DiARC

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Mind the Gap: Structure-Aware Consistency in Preference Learning — Standard preference learning surrogates are theoretically inconsistent for LLM alignment, requiring structure-aware margin adaptation.
• Aligning Implied Statements for Implicit Hate Speech Generalizability with Context-Bounded Semi-hard Negative Mining — Aligning implied statements with context-bounded semi-hard negative mining improves implicit hate speech detection generalizability.
• Large Language Models for Market Research: A Data-augmentation Approach — A statistical data augmentation method debiases LLM-generated data with real human data for accurate market research.
• Human Label Variation as Stable Signal: Learning Annotator-Specific Explanation Behavior via Cross-Annotator Preference Optimization — Large language models can learn and reproduce individual annotator explanation styles from human label variation.
• The Hidden Bias of Process Reward Models:PRISM for Rewarding the Right Reasoning — Process Reward Models (PRMs) suffer from hidden bias; PRISM offers a contrastive learning framework to reward correct reasoning and reduce false positives.
• Towards Spec Learning: Inference-Time Alignment from Preference Pairs — Spec learning aligns LLMs at inference time using natural-language specifications derived from user instructions and preference pairs, avoiding costly fine-tuning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.