\textsc{DiARC}: Distinguishing Positive and Negative Samples Helps Improving ARC-like Reasoning Ability of Large Language Models
Summary
DiARC is a method that improves the reasoning ability of large language models on ARC-like tasks by constructing preference pairs from positive and negative samples, outperforming baselines across multiple benchmarks.
View Cached Full Text
Cached at: 06/26/26, 05:18 AM
# DiARC: Distinguishing Positive and Negative Samples Helps Improving ARC-like Reasoning Ability of Large Language Models
Source: [https://arxiv.org/html/2606.26530](https://arxiv.org/html/2606.26530)
Yuxuan Yang∗, Feiyang Li∗, Yile Wang🖂 College of Computer Science and Software Engineering, Shenzhen University
###### Abstract
The Abstraction and Reasoning Corpus \(ARC;Chollet,[2019](https://arxiv.org/html/2606.26530#bib.bib9)\) 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 onlypositivesample supervision but also the ability to improve model reasoning by distinguishingnegativesamples\. To this end, we draw on the idea of preference alignment and proposeDiARC, 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 thatDiARCconsistently improves performance over baseline models\. The code is released at[https://github\.com/szu\-tera/DiARC](https://github.com/szu-tera/DiARC)\.
DiARC: Distinguishing Positive and Negative Samples Helps Improving ARC\-like Reasoning Ability of Large Language Models
Yuxuan Yang∗, Feiyang Li∗, Yile Wang🖂College of Computer Science and Software Engineering, Shenzhen University
11footnotetext:Equal contribution\.22footnotetext:Corresponding towangyile@szu\.edu\.cn\.## 1Introduction
Large language models \(LLMs\) have achieved impressive progress in natural language understanding\(Yanget al\.,[2024b](https://arxiv.org/html/2606.26530#bib.bib1)\)and mathematical reasoning\(Ahnet al\.,[2024](https://arxiv.org/html/2606.26530#bib.bib8)\)\. However, their ability for solving abstract symbolic tasks still have some limitationsGendronet al\.\([2023](https://arxiv.org/html/2606.26530#bib.bib65)\); Wanget al\.\([2024b](https://arxiv.org/html/2606.26530#bib.bib64)\)and recent work try improving LLMs to solve the typical Abstraction and Reasoning Corpus \(ARC;Chollet,[2019](https://arxiv.org/html/2606.26530#bib.bib9)\), where the model must infer a latent transformation rule from a small number of grid\-based input\-output examples and apply it to a new test input, as Figure[1](https://arxiv.org/html/2606.26530#S1.F1)\(a\) shows\.
Figure 1:Schematic of ARC tasks and our method\. \(a\) A visualized example of ARC problem\. \(b\) Conventional methods with correct target supervision only\. \(c\)DiARCleverage negative samples and optimize the model with preference alignment objectives\.Yet current results remain limited, and stronger frontier LLM\-based systems often require substantial inference\-time cost to achieve higher ARC performance, underscoring the difficulty of abstract rule induction in ARC\(ARC Prize,[2026](https://arxiv.org/html/2606.26530#bib.bib13); Vahdatiet al\.,[2026](https://arxiv.org/html/2606.26530#bib.bib7)\)\. Recent progress on improving ARC reasoning ability of LLMs has increasingly relied on expanding supervision beyond the original benchmark tasks\.Franzenet al\.\([2025](https://arxiv.org/html/2606.26530#bib.bib17)\)builds training pipeline around ARC\-specific data augmentation using the RE\-ARC dataset\(Hodel,[2024](https://arxiv.org/html/2606.26530#bib.bib10)\)\.Wang and Huang \([2025](https://arxiv.org/html/2606.26530#bib.bib29)\)propose program\-based data synthesis for ARC tasks to expose the model to substantially more ARC\-style supervision\. NVARCSorokin and Puget \([2025](https://arxiv.org/html/2606.26530#bib.bib14)\)pushes this trend further by introducing a multi\-stage synthetic data generation pipeline, where LLMs are used to produce puzzle summaries, compose more complex task descriptions, and generate corresponding input\-output grid logic in code, yielding a large synthetic corpus for ARC\-style training\. Although these approaches differ in how additional supervision is constructed, the overall pipelines still rely only on input\-output supervision over correct targets, as Figure[1](https://arxiv.org/html/2606.26530#S1.F1)\(b\) shows\.
From the perspective of learning and cognitive science, learning on difficult problems does not arise only from correct examples, but also from identifying and correcting informative errorsMetcalfe \([2017](https://arxiv.org/html/2606.26530#bib.bib18)\); Dieterichet al\.\([2025](https://arxiv.org/html/2606.26530#bib.bib2)\)\. This idea is also useful for training LLMs\. For example, learning from step\-level preference signals and erroneous traces can improve long\-chain mathematical reasoning performance\(Laiet al\.,[2024](https://arxiv.org/html/2606.26530#bib.bib3); Luet al\.,[2025](https://arxiv.org/html/2606.26530#bib.bib4)\)\.LeGriset al\.\([2025](https://arxiv.org/html/2606.26530#bib.bib20)\)also shows that human solvers in ARC can improve across attempts using feedback\. Inspired by them, we proposeDiARC, a preference alignment framework for improving ARC reasoning ability through encourage LLMs to distinguish the correct output from rejected alternatives, as an example shown in Figure[1](https://arxiv.org/html/2606.26530#S1.F1)\(c\)\.
Specifically,DiARCconstructs preference pairs in which the chosen response is the answer to the query input and the rejected response is a negative derived from that task\. The negative samples are constructed from three levels: \(1\) visual modifications applied directly to the output grid to produce candidates that remain plausible while violating the underlying rule; \(2\) recurring domain\-specific language patterns extracted from task\-specific transformation programs, yielding reusable program\-level rewrites that produce rule\-level negatives across tasks; and \(3\) LLM\-assisted edits to each task’s transformation rule, constructing task\-specific negatives that are as close as possible to the semantic opposite of the original rule\. Training on these preference pairs allows to optimize model not only for correctness, but also for discrimination between the correct rule and plausible near\-miss alternatives\.
We evaluate on six ARC benchmarks including: ARC\-AGI\-1\(Chollet,[2019](https://arxiv.org/html/2606.26530#bib.bib9)\), ARC\-AGI\-2\(Cholletet al\.,[2025](https://arxiv.org/html/2606.26530#bib.bib16)\), MiniARC\(Kimet al\.,[2022](https://arxiv.org/html/2606.26530#bib.bib6)\), ConceptARC\(Moskvichevet al\.,[2023](https://arxiv.org/html/2606.26530#bib.bib27)\), 1D\-ARC\(Xuet al\.,[2024](https://arxiv.org/html/2606.26530#bib.bib32)\), and ARCcommunity\. Experimental results using three open\-source LLMs show thatDiARCyields consistent gains across all ARC datasets\. By using Qwen3 model, it achieves over 96% accuracy on ARC\-AGI\-1, MiniARC, and ConceptARC, outperforming various ARC\-specialized models and strong closed\-source LLMs\.
Figure 2:Examples of negative construction strategies forDiARC\. \(a\) Visualization of ARC task 1f0c79e5\. \(b\) Output\-level visual transformation \(§\\S[3\.2](https://arxiv.org/html/2606.26530#S3.SS2)\)\. \(c\) DSL\-level rule inversion \(§\\S[3\.3](https://arxiv.org/html/2606.26530#S3.SS3)\)\. \(d\) Task\-specific rule editing \(§\\S[3\.4](https://arxiv.org/html/2606.26530#S3.SS4)\)\.
## 2Preliminaries and Motivations
#### ARC Task Formulation\.
An ARC\-like taskτ\\tauconsists of a support set𝒮τ=\{\(Ii,Oi\)\}i=1nτ\\mathcal\{S\}\_\{\\tau\}=\\\{\(I\_\{i\},O\_\{i\}\)\\\}\_\{i=1\}^\{n\_\{\\tau\}\}, a query input gridqτq\_\{\\tau\}, and a target output gridOτ∗O\_\{\\tau\}^\{\*\}\. Each grid is transferred to numerical entries in\{0,…,9\}\\\{0,\\dots,9\\\}, where each integer denotes a grid with corresponding color\. Each task is governed by a latent task\-specific transformationTτT\_\{\\tau\}such thatOi=Tτ\(Ii\)O\_\{i\}=T\_\{\\tau\}\(I\_\{i\}\)fori=1,…,nτi=1,\\dots,n\_\{\\tau\}, andOτ∗=Tτ\(qτ\)O\_\{\\tau\}^\{\*\}=T\_\{\\tau\}\(q\_\{\\tau\}\)\.
The support set𝒮τ\\mathcal\{S\}\_\{\\tau\}and queryqτq\_\{\\tau\}are serialized to the input promptxτx\_\{\\tau\}, and the target outputOτ∗O\_\{\\tau\}^\{\*\}is serialized to the outputyτy\_\{\\tau\}for LLMs\. Thus, each ARC task is converted into a standard sequence\-to\-sequence example\(xτ,yτ\)\(x\_\{\\tau\},y\_\{\\tau\}\)and can be treated as next grid prediction in a language modeling style\.
#### Supervised Fine\-Tuning\.
Given a autoregressive LLMsπθ\\pi\_\{\\theta\}, let𝒟SFT=\{\(xτ,yτ\)\}\\mathcal\{D\}\_\{\\mathrm\{SFT\}\}=\\\{\(x\_\{\\tau\},y\_\{\\tau\}\)\\\}denote the training set constructed from serialized ARC tasks, supervised fine\-tuning \(SFT\) optimizes the model by maximizing the conditional likelihood of each token in target sequence given the input prompt:
ℒSFTθ=−𝔼\(xτ,yτ\)∼𝒟SFT\[∑t=1\|yτ\|logπθ\(yτt∣xτ,yτ<t\)\]\.\\mathcal\{L\}\_\{\\mathrm\{SFT\}\}^\{\\theta\}=\-\\mathbb\{E\}\_\{\(x\_\{\\tau\},y\_\{\\tau\}\)\\sim\\mathcal\{D\}\_\{\\mathrm\{SFT\}\}\}\\left\[\\sum\_\{t=1\}^\{\|y\_\{\\tau\}\|\}\\log\\pi\_\{\\theta\}\(y\_\{\\tau\_\{t\}\}\\mid x\_\{\\tau\},y\_\{\\tau\_\{<t\}\}\)\\right\]\.
\(1\)
This method uses all positive samples, i\.e\., correct input\-output pairs, and requires the model to accurately predict the integer corresponding to the next grid at each prediction step during training, as Figure[1](https://arxiv.org/html/2606.26530#S1.F1)\(b\) shows\.
#### Our Motivations\.
Our motivation consists of three parts: 1\) The output sequences of ARC tasks are often very long, making it difficult to accurately predict at each step from a generative perspective\. 2\) SFT\-based methods focus on local information, whereas ARC tasks are designed with global grid characteristics in mind\. 3\) SFT\-based methods only rely on a large number of positive samples and lack guidance from incorrect outputs\. Compared with generation, holistically judging which output is more reasonable is a relatively easy skill to acquire, yet it is a crucial ability for ARC reasoning\. Thereby, inspired by preference alignment post\-trainingOuyanget al\.\([2022](https://arxiv.org/html/2606.26530#bib.bib11)\), we consider using preference optimization to further enhance the model’s ability to solve ARC tasks\.
## 3Method
### 3\.1Overview
Given the task with support set𝒮τ\\mathcal\{S\}\_\{\\tau\}, queryqτq\_\{\\tau\}, targetOτ∗O\_\{\\tau\}^\{\*\}, the serialized input and output\(x∗,y∗\)\(x^\{\*\},y^\{\*\}\), we denote the chosen output byy\+=y∗y^\{\+\}=y^\{\*\}\. We aim to optimize the model not only to prefer the correct output, but also to reject plausible yet rule\-incorrect alternatives during post\-training\.
To this end, we use anegative samples generator\(𝒩𝒢\\mathcal\{NG\}in short\) to construct a rejected output for the query while keeping the support demonstrations unchanged\. We denote negative sampley−=𝒩𝒢\(𝒮τ,qτ,y\+\)y^\{\-\}=\\mathcal\{NG\}\(\\mathcal\{S\}\_\{\\tau\},q\_\{\\tau\},y^\{\+\}\), where the implementation of𝒩𝒢\\mathcal\{NG\}depends on the negative\-construction strategy\. We then obtain a preference dataset𝒟pref=\{\(x,y\+,y−\)\}\\mathcal\{D\}\_\{\\text\{pref\}\}=\\\{\(x,y^\{\+\},y^\{\-\}\)\\\}, wherey\+y^\{\+\}is the chosen output andy−y^\{\-\}is the rejected output\. To construct informative rejected alternatives, we generate negatives at three levels ranging from visual modifications of the output grid to program\-level rule reversal, including output\-level visual transformation \(§\\S[3\.2](https://arxiv.org/html/2606.26530#S3.SS2)\), DSL\-level rule inversion \(§\\S[3\.3](https://arxiv.org/html/2606.26530#S3.SS3)\), task\-specific rule editing \(§\\S[3\.4](https://arxiv.org/html/2606.26530#S3.SS4)\), and we introduce them as follows\.
### 3\.2Output\-Level Visual Transformations
The first strategy𝒩𝒢Vis\\mathcal\{NG\}\_\{\\mathrm\{Vis\}\}constructs negatives directly in the output space through visual transformations of the output grid, motivated by the strongly visual nature of ARC tasks\(Huet al\.,[2025](https://arxiv.org/html/2606.26530#bib.bib12)\), as illustrated in Figure[2](https://arxiv.org/html/2606.26530#S1.F2)\(b\)\. The key idea is to generate candidate outputs that remain visually and structurally plausible while deviating from the correct transformation rule\. We use four families of atomic operations, yielding 28 transformations in total\.
The first family,grid\-blockoperators, contains three transformations:pooldownup, which applies2×22\\times 2mode pooling followed by upsampling;modecenter, which replaces only the center cell of each3×33\\times 3block with the block mode; andremovesmall, which removes connected components smaller than a preset size threshold\. These operations preserve coarse layout while damaging fine\-grained structure\.
The second family,rigid shifts, contains 16 transformations\. We shift the full grid content in eight directions\{𝒰,𝒟,ℒ,ℛ,𝒰ℒ,𝒰ℛ,𝒟ℒ,𝒟ℛ\}\\\{\\mathcal\{U\},\\mathcal\{D\},\\mathcal\{L\},\\mathcal\{R\},\\mathcal\{UL\},\\mathcal\{UR\},\\mathcal\{DL\},\\mathcal\{DR\}\\\}and use two magnitudes, 10% and 30%, with newly exposed cells filled by the background color\. This produces spatially coherent but misaligned outputs\.
The third family consists of seven morphological and geometry\-related operators:erosion,dilation,skeleton,edge,boundary,convexhull, andfillholes\. Intuitively, these operators thin, expand, outline, simplify, or complete objects while keeping the overall spatial layout largely unchanged\.
The fourth family consists of two local perturbation operators:noise5, which injects 5% salt\-and\-pepper noise, andswap10, which swaps 10% of boundary pixels with neighboring pixels of different colors\. These operations introduce localized appearance and boundary inconsistencies while largely preserving the global arrangement of objects in the grid\.
### 3\.3DSL\-Level Rule Inversion
ARC tasks are represented in a higher\-level domain\-specific language \(DSL\) as executable transformation programs built from predefined primitives\. Prior works represent solutions as executable programs and derives additional supervision through program mutation or decomposition\(Buttet al\.,[2024](https://arxiv.org/html/2606.26530#bib.bib23); Simpson and Banerjee,[2026](https://arxiv.org/html/2606.26530#bib.bib24)\)\. Based on them, the second strategy𝒩𝒢DSL\\mathcal\{NG\}\_\{\\mathrm\{DSL\}\}construct negatives at the program level by inverting recurring local rule patterns in task\-specific transformation programs, as shown in Figure[2](https://arxiv.org/html/2606.26530#S1.F2)\(c\)\.
We first analyze a pool of task\-specific programs that map input grids to output grids, identify recurring local DSL patterns across tasks, and retain those with relatively stable semantics\. To obtain a reusable pattern set, we apply a greedy coverage criterion: at each step, we select the candidate pattern whose rewrite covers the largest number of previously uncovered tasks\. For each retained patternmm, we then define a unified program\-level rewriteϕm\\phi\_\{m\}that acts as a semantic reversal of the original local rule\.
Given a task programppand query inputqq, the original program produces the chosen outputy\+=p\(q\)y^\{\+\}=p\(q\), while the rewritten program produces the rejected outputy−=ϕm\(p\)\(q\)y^\{\-\}=\\phi\_\{m\}\(p\)\(q\)\. The resulting negative is therefore rule\-incorrect due to a localized semantic inversion, while remaining close to the original task logic\.
### 3\.4Task\-Specific Rule Editing
Motivated by program refinement and explicit rule\-guided code editing based worksPourcelet al\.\([2025](https://arxiv.org/html/2606.26530#bib.bib25)\); Liet al\.\([2025a](https://arxiv.org/html/2606.26530#bib.bib26)\), the third strategy𝒩𝒢Edit\\mathcal\{NG\}\_\{\\mathrm\{Edit\}\}constructs negatives at the level of individual tasks by using LLM\-assisted edits to each task’s transformation rule, as Figure[2](https://arxiv.org/html/2606.26530#S1.F2)\(d\) shows\.
Concretely, for a task\-specific transformation program, we use GPT\-5\.4OpenAI \([2026](https://arxiv.org/html/2606.26530#bib.bib53)\)to assist rule editing and construct an alternative rule that is as close as possible to the semantic opposite of the original rule\. When a clean reversal is not feasible, GPT\-5\.4 instead generates a task\-specific alternative rule that stays close to the original transformation logic but leads to a different transformation behavior\. The resulting negative therefore stays close to the original solution in overall structure and execution pattern, while deviating in the task\-relevant rule itself\.
Figure 3:Preference\-learning pipeline\. Given an ARC task, we serialize the context, construct a chosen output and a rejected output, and optimize the model with DPO\.
### 3\.5Preference Optimization Pipeline
After obtaining the negative samples and the constructed preference data, we train the model with direct preference optimization \(DPO;Rafailovet al\.,[2023](https://arxiv.org/html/2606.26530#bib.bib19)\)\. Letπθ\\pi\_\{\\theta\}denote the policy model andπref\\pi\_\{\\mathrm\{ref\}\}the reference model\. For each preference triple\(x,y\+,y−\)\(x,y^\{\+\},y^\{\-\}\), DPO encourages the model to assign higher relative likelihood to the chosen output than to the rejected output:
ℒDPOθ=−𝔼\(x,y\+,y−\)∼𝒟preflogσ\(βlogπθ\(y\+∣x\)πref\(y\+∣x\)−βlogπθ\(y−∣x\)πref\(y−∣x\)\),\\begin\{aligned\} \\mathcal\{L\}\_\{\\mathrm\{DPO\}\}^\{\\theta\}=\-\\mathbb\{E\}\_\{\(x,y^\{\+\},y^\{\-\}\)\\sim\\mathcal\{D\}\_\{\\mathrm\{pref\}\}\}\\log\\sigma\\Bigg\(\\beta&\\log\\frac\{\\pi\_\{\\theta\}\(y^\{\+\}\\mid x\)\}\{\\pi\_\{\\mathrm\{ref\}\}\(y^\{\+\}\\mid x\)\}\\\\ \-\\beta&\\log\\frac\{\\pi\_\{\\theta\}\(y^\{\-\}\\mid x\)\}\{\\pi\_\{\\mathrm\{ref\}\}\(y^\{\-\}\\mid x\)\}\\Bigg\),\\end\{aligned\}
\(2\)whereβ\\betais a temperature parameter\. In practice, we initializeπθ\\pi\_\{\\theta\}fromπSFT\\pi\_\{\\mathrm\{SFT\}\}and use a frozen copy ofπSFT\\pi\_\{\\mathrm\{SFT\}\}as the reference modelπref\\pi\_\{\\mathrm\{ref\}\}\. Unlike reinforcement learning methods that require an explicit reward model, DPO directly optimizes pairwise preferences between correct and incorrect outputs\. This makes it particularly suitable for ARC, where supervision is sparse and many errors are better characterized as plausible but rule\-incorrect alternatives rather than compressed into a single reward score\. Figure[3](https://arxiv.org/html/2606.26530#S3.F3)shows the overall pipeline of preference data construction and DPO training\.
## 4Experiments
### 4\.1Experimental Setup
Dataset and Evaluation\.We evaluate on six benchmarks that cover the original ARC setting, its harder successor, and four related ARC\-style transfer or additional benchmarks\.ARC\-AGI\-1Chollet \([2019](https://arxiv.org/html/2606.26530#bib.bib9)\)is the original ARC benchmark for few\-shot abstract rule induction from grid\-based input\-output examples, and we use its 400\-task public evaluation set\.ARC\-AGI\-2Cholletet al\.\([2025](https://arxiv.org/html/2606.26530#bib.bib16)\)is a harder successor that preserves the same input\-output format while providing a more challenging test of abstract reasoning and problem solving, and we use its 120\-task public evaluation set\.MiniARCKimet al\.\([2022](https://arxiv.org/html/2606.26530#bib.bib6)\)is a compact 5×\\times5 ARC\-style benchmark designed to reduce variation in grid size and focus more directly on abductive reasoning\.ConceptARCMoskvichevet al\.\([2023](https://arxiv.org/html/2606.26530#bib.bib27)\)organizes ARC\-style tasks into concept groups in order to evaluate abstraction and generalization over recurring spatial and semantic concepts\.1D\-ARCXuet al\.\([2024](https://arxiv.org/html/2606.26530#bib.bib32)\)is a one\-dimensional ARC\-style benchmark that simplifies the original two\-dimensional grid setting into one\-dimensional transformation tasks while preserving the core requirement of abstract rule induction\.ARCcommunity111[https://github\.com/arc\-community/arc](https://github.com/arc-community/arc)is a collection of community\-created ARC\-style tasks that extends evaluation beyond the original benchmark distribution and provides an additional test of generalization to diverse human\-designed problems\.
For the standard evaluation on ARC\-AGI\-1 and ARC\-AGI\-2, each test input allows two candidate outputs and is counted as correct if either exactly matches the ground\-truth grid\. If a task contains multiple test inputs, the task score is the average over its test inputs\. For MiniARC, ConceptARC, 1D\-ARC, and ARCcommunity, we report single\-response exact\-match accuracy\.
Models\.We use LLMs across three families including Llama\-3\.2\-3BGrattafioriet al\.\([2024](https://arxiv.org/html/2606.26530#bib.bib36)\), Mistral\-NeMo\-Minitron\-8BSreenivaset al\.\([2024](https://arxiv.org/html/2606.26530#bib.bib35)\), and Qwen3\-4BYanget al\.\([2025](https://arxiv.org/html/2606.26530#bib.bib37)\)\. During training, we optimized their vocabulary and the effective fine\-tuned parameter size is slightly reduced\.
Models\#ParamsARC\-1ARC\-2\\rowcolorgray\!15Closed\-source Large Language ModelsOpus 4\.5\(Anthropic,[2025](https://arxiv.org/html/2606.26530#bib.bib51)\)†N/A40\.07\.8Kimi K2\.5\(Baiet al\.,[2026](https://arxiv.org/html/2606.26530#bib.bib52)\)†1T65\.311\.8GPT\-5\.4\(OpenAI,[2026](https://arxiv.org/html/2606.26530#bib.bib53)\)†N/A68\.229\.2Grok\-4\(xAI,[2025](https://arxiv.org/html/2606.26530#bib.bib54)\)‡1\.7T79\.629\.4\\rowcolorgray\!15ARC\-specialized Recursive ModelsHRM\(Wanget al\.,[2025](https://arxiv.org/html/2606.26530#bib.bib30)\)27M40\.35\.0GRAM\(Baeket al\.,[2026](https://arxiv.org/html/2606.26530#bib.bib33)\)10\.9M52\.011\.1\\rowcolorgray\!15ARC\-specialized Vision ModelsVARC\(Huet al\.,[2025](https://arxiv.org/html/2606.26530#bib.bib12)\)18M54\.58\.3VARC \(ensemble\)\(Huet al\.,[2025](https://arxiv.org/html/2606.26530#bib.bib12)\)73M60\.411\.1\\rowcolorgray\!15Open\-source Large Language ModelsLlama\-3\.2Grattafioriet al\.\([2024](https://arxiv.org/html/2606.26530#bib.bib36)\)3B0\.500\.00Llama\-3\.2 \+ SFT only\(Franzenet al\.,[2025](https://arxiv.org/html/2606.26530#bib.bib17)\)2\.6B59\.384\.44Llama\-3\.2 \+DiARC\(Ours\)2\.6B61\.506\.53MinitronSreenivaset al\.\([2024](https://arxiv.org/html/2606.26530#bib.bib35)\)8B1\.120\.00Minitron \+ SFT only\(Franzenet al\.,[2025](https://arxiv.org/html/2606.26530#bib.bib17)\)6\.9B71\.259\.86Minitron \+DiARC\(Ours\)6\.9B72\.3810\.28Qwen3Yanget al\.\([2025](https://arxiv.org/html/2606.26530#bib.bib37)\)4B0\.630\.00Qwen3 \+ SFT only\(Sorokin and Puget,[2025](https://arxiv.org/html/2606.26530#bib.bib14)\)3\.6B96\.2530\.83Qwen3 \+DiARC\(Ours\)3\.6B97\.0033\.06\\rowcolorgray\!15Human PerformanceHuman panel \(average\)\(ARC Prize,[2025](https://arxiv.org/html/2606.26530#bib.bib21)\)–64\.2060\.00
Table 1:Results on ARC\-AGI\-1 and ARC\-AGI\-2 benchmarks\.†\{\\dagger\}: Reported in leaderboard ofARC Prize \([2026](https://arxiv.org/html/2606.26530#bib.bib13)\)\.‡\{\\ddagger\}: Reported inBerman \([2025](https://arxiv.org/html/2606.26530#bib.bib15)\)\.Models\#ParamsMiniConcept1DComm\.\\rowcolorgray\!15Closed\-source Large Language Modelso1\-mini\(OpenAI,[2024](https://arxiv.org/html/2606.26530#bib.bib43)\)†N/A30\.20–––DeepSeek\-R1\(Guoet al\.,[2025](https://arxiv.org/html/2606.26530#bib.bib44)\)†671B28\.86–––GPT\-4\(Achiamet al\.,[2023](https://arxiv.org/html/2606.26530#bib.bib47)\)‡N/A24\.1024\.052\.00–Qwen\-2\.5\(Yanget al\.,[2024a](https://arxiv.org/html/2606.26530#bib.bib46)\)‡7B22\.0026\.055\.00–Claude Sonnet 4\(Anthropic,[2024](https://arxiv.org/html/2606.26530#bib.bib48)\)♮N/A–60\.2––Gemini 2\.5 Pro\(Comaniciet al\.,[2025](https://arxiv.org/html/2606.26530#bib.bib49)\)♮N/A–66\.0––o4\-mini\(OpenAI,[2025](https://arxiv.org/html/2606.26530#bib.bib50)\)♮N/A–70\.8––\\rowcolorgray\!15Closed\-source Large Vision Modelso4\-mini \(visual\)\(OpenAI,[2025](https://arxiv.org/html/2606.26530#bib.bib50)\)♮N/A–8\.1––\\rowcolorgray\!15Open\-source Large Language ModelsLlama\-3\.2Grattafioriet al\.\([2024](https://arxiv.org/html/2606.26530#bib.bib36)\)3B1\.341\.040\.112\.17Llama\-3\.2 \+ SFT only\(Franzenet al\.,[2025](https://arxiv.org/html/2606.26530#bib.bib17)\)2\.6B70\.4758\.1351\.7223\.91Llama\-3\.2 \+DiARC\(Ours\)2\.6B73\.1559\.1752\.3930\.43MinitronSreenivaset al\.\([2024](https://arxiv.org/html/2606.26530#bib.bib35)\)8B6\.045\.833\.556\.52Minitron \+ SFT only\(Franzenet al\.,[2025](https://arxiv.org/html/2606.26530#bib.bib17)\)6\.9B74\.5060\.4257\.0534\.78Minitron \+DiARC\(Ours\)6\.9B76\.5162\.7158\.8243\.48Qwen3Yanget al\.\([2025](https://arxiv.org/html/2606.26530#bib.bib37)\)4B0\.671\.250\.110\.00Qwen3 \+ SFT only\(Sorokin and Puget,[2025](https://arxiv.org/html/2606.26530#bib.bib14)\)3\.6B95\.3096\.4642\.8458\.70Qwen3 \+DiARC\(Ours\)3\.6B97\.3297\.2945\.9563\.04\\rowcolorgray\!15Human PerformanceHuman \(pass@1\)\(Begeret al\.,[2026](https://arxiv.org/html/2606.26530#bib.bib22)\)––73\.00––
Table 2:Results on MiniARC, ConceptARC, 1D\-ARC, and ARCcommunity benchmarks\.†\{\\dagger\}: Reported inZhenget al\.\([2025](https://arxiv.org/html/2606.26530#bib.bib28)\)\.‡\{\\ddagger\}: Reported inWang and Huang \([2025](https://arxiv.org/html/2606.26530#bib.bib29)\)\.♮\\natural: Reported inBegeret al\.\([2026](https://arxiv.org/html/2606.26530#bib.bib22)\)\.BenchmarksModelsSFT Init\.Grid\-BlockMorphologyRandom PerturbRigid ShiftARC\-AGI\-1Llama\-3\.2\-3B59\.3859\.88\(\+0\.50\)60\.75\(\+1\.37\)61\.50\(\+2\.12\)59\.75\(\+0\.37\)Minitron\-8B71\.2571\.88\(\+0\.63\)71\.88\(\+0\.63\)71\.38\(\+0\.13\)71\.38\(\+0\.13\)Qwen3\-4B96\.2596\.75\(\+0\.50\)97\.00\(\+0\.75\)96\.50\(\+0\.25\)96\.75\(\+0\.50\)ARC\-AGI\-2Llama\-3\.2\-3B4\.446\.11\(\+1\.67\)5\.69\(\+1\.25\)6\.53\(\+2\.09\)5\.69\(\+1\.25\)Minitron\-8B9\.869\.31\(\-0\.55\)10\.28\(\+0\.42\)9\.17\(\-0\.69\)9\.03\(\-0\.83\)Qwen3\-4B30\.8331\.39\(\+0\.56\)31\.81\(\+0\.98\)32\.64\(\+1\.81\)33\.06\(\+2\.23\)MiniARCLlama\-3\.2\-3B70\.4771\.14\(\+0\.67\)73\.15\(\+2\.68\)72\.48\(\+2\.01\)72\.48\(\+2\.01\)Minitron\-8B74\.5075\.17\(\+0\.67\)76\.51\(\+2\.01\)75\.17\(\+0\.67\)76\.51\(\+2\.01\)Qwen3\-4B95\.3097\.32\(\+2\.02\)97\.32\(\+2\.02\)97\.32\(\+2\.02\)96\.64\(\+1\.34\)ConceptARCLlama\-3\.2\-3B58\.1358\.75\(\+0\.62\)59\.17\(\+1\.04\)59\.17\(\+1\.04\)58\.13\(\+0\.00\)Minitron\-8B60\.4262\.71\(\+2\.29\)62\.29\(\+1\.87\)62\.71\(\+2\.29\)62\.08\(\+1\.66\)Qwen3\-4B96\.4696\.88\(\+0\.42\)96\.88\(\+0\.42\)96\.67\(\+0\.21\)97\.29\(\+0\.83\)1D\-ARCLlama\-3\.2\-3B51\.72––52\.39\(\+0\.67\)52\.28\(\+0\.56\)Minitron\-8B57\.05––58\.82\(\+1\.77\)58\.16\(\+1\.11\)Qwen3\-4B42\.84––45\.95\(\+3\.11\)45\.17\(\+2\.33\)ARCcommunityLlama\-3\.2\-3B23\.9130\.43\(\+6\.52\)26\.09\(\+2\.18\)26\.09\(\+2\.18\)26\.09\(\+2\.18\)Minitron\-8B34\.7843\.48\(\+8\.70\)43\.48\(\+8\.70\)43\.48\(\+8\.70\)43\.48\(\+8\.70\)Qwen3\-4B58\.7060\.87\(\+2\.17\)63\.04\(\+4\.34\)58\.70\(\+0\.00\)63\.04\(\+4\.34\)Avg\.–56\.3158\.14\(\+1\.83\)58\.36\(\+2\.04\)57\.97\(\+1\.66\)58\.09\(\+1\.78\)Table 3:Results of output\-level negative construction families across six benchmarks\. Values are formatted as result \(gain\>\>1\.00,gain≤\\leq1\.00, orloss\) relative to the corresponding SFT initialization model\. For 1D\-ARC, only random perturb and rigid shift operation are applicable, so results for grid\-block and morphology are not reported\.
Preference Data Construction\.For ARC\-AGI\-1, all three negative\-construction strategies are built on top of RE\-ARC\(Hodel,[2024](https://arxiv.org/html/2606.26530#bib.bib10)\), which provides task\-specific generators for the 400 original ARC training tasks\. Foroutput\-level visual transformation, we first use these generators to sample ARC tasks and then apply output\-space transformations to construct rejected outputs\. ForDSL\-level rule inversionandtask\-specific rule editing, we operate on the corresponding task\-specific transformation programs to generate chosen/rejected pairs\. For the remaining five benchmarks, RE\-ARC\-style per\-task generators are unavailable\. We therefore use only theoutput\-level visual transformationstrategy for preference construction\. For ARC\-AGI\-2, preference data is constructed from the released NVARC\-fullSorokin and Puget \([2025](https://arxiv.org/html/2606.26530#bib.bib14)\)resources\.
For MiniARC, ConceptARC, 1D\-ARC, and ARCcommunity, we use their official training sets to construct preference data with the same ARC\-oriented negative\-construction pipeline\. Across all six benchmarks, we keep the overall preference\-learning setup as consistent as possible: the original support demonstrations and query input are always preserved, and only the rejected output is varied according to the available negative\-construction strategy\. This design helps isolate the contribution ofDiARCfrom other sources of variation and makes cross\-benchmark and cross\-backbone comparisons easier to interpret\.
Detailed statistics of the constructed preference datasets used in these experiments are provided in Appendix[A](https://arxiv.org/html/2606.26530#A1)\.
Training Details\.We train all models with a shared setup including learning rate1×10−61\\times 10^\{\-6\}, effective batch size 8, cosine learning schedule, warmup ratio 0\.1, maximum sequence length 4096, and maximum prompt length 3584\. LoRAHuet al\.\([2022](https://arxiv.org/html/2606.26530#bib.bib62)\)fine\-tuning is used and the model\-specific settings are summarized in Appendix[B](https://arxiv.org/html/2606.26530#A2)\. All experiments are conducted on a single NVIDIA L40 GPU\.
### 4\.2Main Results
Tables[4\.1](https://arxiv.org/html/2606.26530#S4.SS1)and[4\.1](https://arxiv.org/html/2606.26530#S4.SS1)show the results on all six ARC\-style benchmarks, Table[3](https://arxiv.org/html/2606.26530#S4.T3)breaks down the results with different output\-level negative families, and Table[4](https://arxiv.org/html/2606.26530#S4.T4)compares the DSL\-level inversion and task\-specific editing on ARC\-AGI\-1\.
#### DiARCbrings broad gains across backbones and benchmarks\.
As shown in Tables[4\.1](https://arxiv.org/html/2606.26530#S4.SS1)and[4\.1](https://arxiv.org/html/2606.26530#S4.SS1),DiARCimproves the performance of SFT baselines on all six benchmarks for Llama\-3\.2\-3B, Qwen3\-4B, and Minitron\-8B, with an average absolute gain of 2\.48 points across the 18 model–benchmark pairs\. The gains are not limited to weaker SFT baselines but also strong ones like Qwen3\-4B\. The gains also do not scale monotonically with model size: all three backbones benefit fromDiARC, with average improvements of 2\.52, 2\.21, and 2\.72 points for Llama\-3\.2\-3B, Qwen3\-4B, and Minitron\-8B, respectively\. When using the Qwen3 backbone model,DiARCachieves an accuracy of over 96 on ARC\-AGI\-1, MiniARC, and ConceptARC, surpassing both closed\-source models and existing open\-source model methods, demonstrating its effectiveness\.
#### Useful preference signals can be constructed at multiple levels\.
Table[3](https://arxiv.org/html/2606.26530#S4.T3)shows that the four output\-level families improve over the initialization models on average, with morphology giving the strongest mean performance, although the best family varies across datasets and backbones\. One possible reason is that ARC is built around a set of human\-like priors, especially objectness as well as basic geometry and topology\(Chollet,[2019](https://arxiv.org/html/2606.26530#bib.bib9)\)\. Compared with other perturbations, morphology\-based negatives more directly alter object boundaries, connectivity, and other structural properties while preserving the object layout\. This makes them more likely to stay within the kind of object\-centric transformation space that ARC is intended to test, and therefore more likely to provide informative near\-miss alternatives\. At the same time, the most effective negative family differs across backbones, which may be related to differences in the models’ underlying capabilities, as they may develop different inductive biases and characteristic error patterns, and thus benefit from different forms of hard negative supervision\.
Table[4](https://arxiv.org/html/2606.26530#S4.T4)shows that both DSL\-level inversion and task\-specific editing provide gains on ARC\-AGI\-1\. Under task\-specific editing, however, we observe a temporary performance drop when around 200 tasks are edited, followed by a recovery as the number of edited tasks increases\. This is because LLM\-assisted editing tends to map negative samples into a few recurring task types, such as set/extraction operations, compressed summaries, and bounding\-box or contour extraction\. When the edited set is relatively small, these recurring patterns can dominate the training signal and make the model biased toward a narrow set of negative constructions\. As the edited set becomes larger and more diverse, this bias is diluted, allowing the model to better capture the intended contrast between positive and negative rules\. Taken together, results suggest that the benefit ofDiARCdoes not depend on a single negative\-construction recipe\.
ModelsSFT Init\.DSL\-LevelTask\-Specific Edit100200400Llama\-3\.2\-3B59\.3860\.2560\.1259\.1359\.62Minitron\-8B71\.2572\.3871\.8871\.3872\.00Qwen3\-4B96\.2596\.7597\.0096\.5097\.00
Table 4:Comparison between DSL\-level inversion and task\-specific editing on ARC\-AGI\-1\. For task\-specific editing, we report results under three data settings of 100, 200, and 400 edited tasks, respectively\.## 5Analysis
### 5\.1Impacts of Generation and Discrimination
Our ARC system first generate candidate solutions and then use product\-of\-experts scoring to select the final answer\.DiARCcan therefore improve performance through two different channels: it may increase the number of correct candidate solutions produced by the model, or it may improve the model’s ability to assign higher preference to correct candidates that are already generated\.
BenchmarksMean Gen\. GainMean Disc\. GainMeanΔ\\DeltaSolvedARC\-AGI\-1\+1\.56\+1\.94\+3\.50ARC\-AGI\-2\+0\.75\+0\.50\+1\.25MiniARC\+1\.17\+1\.33\+2\.50ConceptARC\+1\.42\+3\.67\+5\.091D\-ARC\+7\.83\+6\.50\+14\.33ARCcommunity\+0\.42\+0\.71\+1\.13All Average\+2\.19\+2\.44\+4\.63
Table 5:Impacts of generation and discrimination afterDiARCtraining across six benchmarks\. MeanΔ\\Deltasolved denotes the average increase in the number of correctly solved items relative to the corresponding base model\.ModelsVariantsBase\+ TTT\+ Aug\+ PoE\+ DFSLlama\-3\.2\-3BInit\.12\.3734\.6247\.8852\.1259\.38\+DiARC12\.6235\.0053\.6253\.6261\.50Minitron\-8BInit\.17\.1241\.1260\.5063\.3871\.25\+DiARC17\.3741\.7561\.1264\.3772\.38Qwen3\-4BInit\.23\.5031\.2561\.2563\.7596\.25\+DiARC23\.5031\.8762\.0064\.6297\.00
Table 6:Compatibility ofDiARCwith the test\-time scaling techniques\. We replace the base model with itsDiARC\-enhanced counterpart while keeping the remaining inference pipeline unchanged for comparison\.To distinguish these effects, we define two gain quantities\. Letgeneration gainbe the number of correct candidates generated afterDiARCtraining but not generated by the base model, minus the number of correct candidates generated by the base model but no longer generated afterDiARCtraining\. Letdiscrimination gainbe the number of correct candidates that are generated in both settings, not selected by the base model, but selected afterDiARCtraining, minus the number of correct candidates that are generated in both settings, selected by the base model, but not selected afterDiARCtraining\. Positive generation gain therefore indicates a net gain in candidate generation, while positive discrimination gain indicates a net gain in candidate selection\.
Table[5](https://arxiv.org/html/2606.26530#S5.T5)summarizes these effects across six benchmarks\. Overall, both quantities are positive on average: the mean generation gain is\+2\.19\+2\.19and the mean discrimination gain is\+2\.44\+2\.44, while the mean improvement in solved items is\+4\.63\+4\.63under the benchmark\-specific evaluation metric\. These results suggest that the gains fromDiARCarise from both stages of the pipeline: the model produces more correct candidates and also becomes better at assigning higher preference to correct candidates once they are generated\.
### 5\.2Compatibility with Test\-Time Scaling
Different test\-time scaling \(TTS\) methods have been used to improve reasoning performance of LLMs\. We further analyze whetherDiARCis compatible with different techniques\. FollowingFranzenet al\.\([2025](https://arxiv.org/html/2606.26530#bib.bib17)\), we combine base andDiARC\-enhanced ones with test\-time training \(TTT\), task augmentation \(Aug\), reranking \(PoE, product\-of\-experts scoring\), and searching \(DFS, depth\-first search\) methods, comparing how improvements in the underlying models propagate to final accuracy\.
Table[6](https://arxiv.org/html/2606.26530#S5.T6)shows the results\. Across different TTS techniques,DiARC\-enhanced model remain compatible and preserve competitive performance relative to their corresponding base models\. This suggests thatDiARCis not limited to direct model evaluation, but can also be integrated into advanced test\-time scaling pipeline for model improvement\.
### 5\.3Case Study
Figure[4](https://arxiv.org/html/2606.26530#S5.F4)shows the prediction results of Llama\-3\.2\-3B model on ARC\-AGI\-1 task 31adaf00 and see Appendix[C](https://arxiv.org/html/2606.26530#A3)for more cases\. In this task, the main challenge lies in identifying all valid blank pockets rather than generating a new structure\. We suggests that the output\-level perturbation based negative sample is helpful when dealing with similar problems, because it preserves the global layout while removing small valid filled regions, closely matching the model’s typical under\-completion errors\. Thereby, our model gives the precise prediction while the original model does not\.
Figure 4:Case study on ARC\-AGI\-1 task 31adaf00\.## 6Related Work
#### Improving ARC Reasoning of Models\.
Existing approaches to improving ARC reasoning ability can be broadly grouped into three types: \(1\)recursive modelsWanget al\.\([2025](https://arxiv.org/html/2606.26530#bib.bib30)\); Jolicoeur\-Martineau \([2025](https://arxiv.org/html/2606.26530#bib.bib42)\); Baeket al\.\([2026](https://arxiv.org/html/2606.26530#bib.bib33)\)view ARC as an iterative reasoning process in which intermediate states must be repeatedly refined, and therefore improve performance by revisiting and updating candidate solutions through repeated computation; \(2\)vision modelstreat ARC as a visual reasoning problemLiet al\.\([2025b](https://arxiv.org/html/2606.26530#bib.bib40)\); Huet al\.\([2025](https://arxiv.org/html/2606.26530#bib.bib12)\); Shuet al\.\([2026](https://arxiv.org/html/2606.26530#bib.bib41)\), formulating ARC problems as an image\-to\-image or grid\-to\-grid prediction task and exploit visual inductive biases to improve performance on ARC\-style benchmarks; and \(3\)large language modelssolve ARC\-style problems by redesigning task representations, prompting strategies, and ARC\-oriented training data\.Xuet al\.\([2024](https://arxiv.org/html/2606.26530#bib.bib32)\)shows that object\-based representations can substantially improve GPT\-style reasoning on ARC\. Subsequent works explore explicit hypothesis generation and program verificationWanget al\.\([2024a](https://arxiv.org/html/2606.26530#bib.bib60)\); Qiuet al\.\([2024](https://arxiv.org/html/2606.26530#bib.bib63)\), and self\-improving program synthesis through hindsight replayButtet al\.\([2024](https://arxiv.org/html/2606.26530#bib.bib23)\)\. More recent approaches improve ARC reasoning through data augmentation and synthetic program\-based supervision\(Bikovet al\.,[2024](https://arxiv.org/html/2606.26530#bib.bib61); Wang and Huang,[2025](https://arxiv.org/html/2606.26530#bib.bib29)\)\. These efforts have made some progress, but have yet to tap into the potential improvement of negative samples for ARC reasoning ability\.
#### Preference Alignment of LLMs\.
Preference alignment has become a standard paradigm for adapting LLMs to desired behaviors\. Early work studies learning from human preferences and reinforcement learning from human feedback, with the goal of making model outputs more helpful, honest, and harmless\(Ouyanget al\.,[2022](https://arxiv.org/html/2606.26530#bib.bib11)\)\. More recently, direct preference optimization methods simplify this pipeline by optimizing pairwise preferences without separately training an explicit reward model\(Rafailovet al\.,[2023](https://arxiv.org/html/2606.26530#bib.bib19)\)\. A growing body of work has further explored simpler and more efficient preference\-based objectives, such as reference\-free or monolithic optimization variants\(Honget al\.,[2024](https://arxiv.org/html/2606.26530#bib.bib57); Menget al\.,[2024](https://arxiv.org/html/2606.26530#bib.bib58)\)\. Our work extends this line of research to ARC\-style reasoning by constructing task\-specific chosen\-rejected output pairs, so that the model learns not only to fit correct outputs, but also to distinguish them from plausible yet rule\-incorrect alternatives\.
## 7Conclusion
We introducedDiARC, a preference alignment framework for ARC\-like tasks by training LLMs to distinguish correct outputs from task\-specific rejected alternatives\. We proposed three negative sample construction strategies operating at different levels: output\-level visual transformations, DSL\-level rule inversion, and task\-specific rule editing\. Across six ARC\-style benchmarks, our method consistently improves the ARC reasoning ability of open\-source LLMs, outperforms ARC\-specialized and strong closed\-source models\. These results suggest that, with suitable constructed negative samples, preference alignment can also be a effective post\-training techniques for improving ARC\-style reasoning ability of LLMs\.
## Limitations
First, the full negative\-construction pipeline of our method depends on auxiliary ARC resources, especially RE\-ARC task generators and task\-specific transformation programs\. At present, such standardized per\-task generation resources are available for ARC\-AGI\-1, but comparable RE\-ARC\-style generators are not yet available for the remaining benchmarks\. As a result, all three negative\-construction strategies can be applied on ARC\-AGI\-1, but only the output\-level visual transformation strategy is used on other benchmarks\. Second, our rejected outputs are constructed automatically rather than annotated by humans, so their informativeness and semantic quality may vary across tasks and construction methods, and the performance can be further improved when negatives are reasonably manually evaluated and assigned\.
## Acknowledgments
This work was supported in parts by Guangdong Basic and Applied Basic Research Foundation \(2026A1515011358\), Shenzhen Natural Science Foundation \(JCYJ20250604181610014\), and Intelligent Computing Center of Shenzhen University\.
## References
- J\. Achiam, S\. Adler, S\. Agarwal, L\. Ahmad, I\. Akkaya, F\. L\. Aleman, D\. Almeida, J\. Altenschmidt, S\. Altman, S\. Anadkat,et al\.\(2023\)GPT\-4 technical report\.arXiv preprint arXiv:2303\.08774\.External Links:[Link](https://arxiv.org/abs/2303.08774)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.11.11.3.3.3.1)\.
- Large language models for mathematical reasoning: progresses and challenges\.InProceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop,N\. Falk, S\. Papi, and M\. Zhang \(Eds\.\),St\. Julian’s, Malta,pp\. 225–237\.External Links:[Link](https://aclanthology.org/2024.eacl-srw.17/),[Document](https://dx.doi.org/10.18653/v1/2024.eacl-srw.17)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p1.1)\.
- Anthropic \(2024\)System card: claude opus 4 & claude sonnet 4\.External Links:[Link](https://www-cdn.anthropic.com/6d8a8055020700718b0c49369f60816ba2a7c285.pdf)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.13.13.5.5.5.1)\.
- Anthropic \(2025\)System card: claude opus 4\.5\.External Links:[Link](https://www-cdn.anthropic.com/bf10f64990cfda0ba858290be7b8cc6317685f47.pdf)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.1.1.1.1.1)\.
- ARC Prize \(2025\)ARC\-agi\-2 \+ arc prize 2025 is live\!\.Note:Published March 24, 2025External Links:[Link](https://arcprize.org/blog/announcing-arc-agi-2-and-arc-prize-2025)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.4.4.4.24.1)\.
- ARC Prize \(2026\)ARC prize \- leaderboard\.Note:Accessed: 2026\-05\-07External Links:[Link](https://arcprize.org/leaderboard)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p2.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.23)\.
- J\. Baek, M\. Jo, M\. Kim, M\. Ren, Y\. Bengio, and S\. Ahn \(2026\)Generative recursive reasoning\.arXiv preprint arXiv:2605\.19376\.External Links:[Link](https://arxiv.org/abs/2605.19376)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.4.4.4.9.1),[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px1.p1.1)\.
- T\. Bai, Y\. Bai, Y\. Bao, S\. Cai, Y\. Cao, Y\. Charles, H\. Che, C\. Chen, G\. Chen,et al\.\(2026\)Kimi k2\.5: visual agentic intelligence\.arXiv preprint arXiv:2602\.02276\.Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.2.2.2.2.1)\.
- C\. Beger, R\. Yi, S\. Fu, A\. K\. Moskvichev, S\. Tsai, S\. Rajamanickam, and M\. Mitchell \(2026\)Do AI models perform human\-like abstract reasoning across modalities?\.External Links:[Link](https://openreview.net/forum?id=CcYCgF491G)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.16.16.8.8.23.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.23.23)\.
- J\. Berman \(2025\)How i got the highest score on arc\-agi again, swapping python for english\.Note:SubstackExternal Links:[Link](https://jeremyberman.substack.com/p/how-i-got-the-highest-score-on-arc-agi-again)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.23)\.
- I\. Bikov, B\. Prystawski, B\. Prystawski, and N\. D\. Goodman \(2024\)Reflection system for the abstraction and reasoning corpus\.External Links:[Link](https://openreview.net/forum?id=HBq4Jwl4Ja)Cited by:[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px1.p1.1)\.
- N\. Butt, B\. Manczak, A\. Wiggers, C\. Rainone, D\. W\. Zhang, M\. Defferrard, and T\. Cohen \(2024\)CodeIt: self\-improving language models with prioritized hindsight replay\.InProceedings of the 41st International Conference on Machine LearningICLR 2026 Workshop \- From Human Cognition to AI Reasoning: Models, Methods, and ApplicationsProceedings of the 42nd International Conference on Machine LearningProceedings of the 42nd International Conference on Machine LearningProceedings of the 24th China National Conference on Computational Linguistics \(CCL 2025\)Proceedings of the 2024 Conference on Empirical Methods in Natural Language ProcessingThe Thirty\-eighth Annual Conference on Neural Information Processing SystemsThe Twelfth International Conference on Learning RepresentationsNeurIPS 2024 Workshop on System 2 Reasoning at ScaleInternational Conference on Learning RepresentationsThe Twelfth International Conference on Learning RepresentationsICLR 2024 Workshop: How Far Are We From AGI,R\. Salakhutdinov, Z\. Kolter, K\. Heller, A\. Weller, N\. Oliver, J\. Scarlett, F\. Berkenkamp, A\. Singh, M\. Fazel, D\. Hsu, S\. Lacoste\-Julien, F\. Berkenkamp, T\. Maharaj, K\. Wagstaff, J\. Zhu, A\. Singh, M\. Fazel, D\. Hsu, S\. Lacoste\-Julien, F\. Berkenkamp, T\. Maharaj, K\. Wagstaff, J\. Zhu, M\. Sun, P\. Duan, Z\. Liu, R\. Xu, W\. Sun, Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Proceedings of Machine Learning ResearchProceedings of Machine Learning ResearchProceedings of Machine Learning Research, Vol\.235267267,pp\. 5013–5034\.External Links:[Link](https://proceedings.mlr.press/v235/butt24a.html)Cited by:[§3\.3](https://arxiv.org/html/2606.26530#S3.SS3.p1.1),[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px1.p1.1)\.
- F\. Chollet, M\. Knoop, G\. Kamradt, B\. Landers, and H\. Pinkard \(2025\)Arc\-agi\-2: a new challenge for frontier ai reasoning systems\.arXiv preprint arXiv:2505\.11831\.External Links:[Link](https://arxiv.org/abs/2505.11831)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p5.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.p1.1)\.
- F\. Chollet \(2019\)On the measure of intelligence\.arXiv preprint arXiv:1911\.01547\.External Links:[Link](https://arxiv.org/abs/1911.01547)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p1.1),[§1](https://arxiv.org/html/2606.26530#S1.p5.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.p1.1),[§4\.2](https://arxiv.org/html/2606.26530#S4.SS2.SSS0.Px2.p1.1)\.
- G\. Comanici, E\. Bieber, M\. Schaekermann, I\. Pasupat, N\. Sachdeva, I\. Dhillon, M\. Blistein, O\. Ram, D\. Zhang, E\. Rosen,et al\.\(2025\)Gemini 2\.5: pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities\.arXiv preprint arXiv:2507\.06261\.External Links:[Link](https://arxiv.org/abs/2507.06261)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.14.14.6.6.6.1)\.
- S\. Dieterich, S\. Rumann, and M\. Rodemer \(2025\)Conditions for effective learning from erroneous examples: a systematic review\.Educational Psychology Review37,pp\. 94\.External Links:[Document](https://dx.doi.org/10.1007/s10648-025-10071-x),[Link](https://doi.org/10.1007/s10648-025-10071-x)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p3.1)\.
- D\. Franzen, J\. Disselhoff, and D\. Hartmann \(2025\)Product of experts with LLMs: boosting performance on ARC is a matter of perspective\.InForty\-second International Conference on Machine Learning,External Links:[Link](https://openreview.net/forum?id=dsBjxI6l8W)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p2.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.16.16.8.8.14.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.16.16.8.8.17.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.4.4.4.15.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.4.4.4.18.1),[§5\.2](https://arxiv.org/html/2606.26530#S5.SS2.p1.1)\.
- G\. Gendron, Q\. Bao, M\. Witbrock, and G\. Dobbie \(2023\)Large language models are not strong abstract reasoners\.arXiv preprint arXiv:2305\.19555\.External Links:[Link](https://arxiv.org/abs/2305.19555)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p1.1)\.
- A\. Grattafiori, A\. Dubey, A\. Jauhri, A\. Pandey, A\. Kadian, A\. Al\-Dahle, A\. Letman, A\. Mathur, A\. Schelten, A\. Vaughan,et al\.\(2024\)The llama 3 herd of models\.arXiv preprint arXiv:2407\.21783\.External Links:[Link](https://arxiv.org/abs/2407.21783)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.16.16.8.8.13.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.4.4.4.14.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.p3.1)\.
- D\. Guo, D\. Yang, H\. Zhang, J\. Song, P\. Wang, Q\. Zhu, R\. Xu, R\. Zhang, S\. Ma, X\. Bi, X\. Zhang, X\. Yu, Y\. Wu, Z\. F\. Wu, Z\. Gou, Z\. Shao, Z\. Li, Z\. Gao, A\. Liu, B\. Xue, B\. Wang, B\. Wu, B\. Feng, C\. Lu, C\. Zhao, C\. Deng, C\. Ruan, D\. Dai, D\. Chen, D\. Ji, E\. Li, F\. Lin, F\. Dai, F\. Luo, G\. Hao, G\. Chen, G\. Li, H\. Zhang, H\. Xu, H\. Ding, H\. Gao, H\. Qu, H\. Li, J\. Guo, J\. Li, J\. Chen, J\. Yuan, J\. Tu, J\. Qiu, J\. Li, J\. L\. Cai, J\. Ni, J\. Liang, J\. Chen, K\. Dong, K\. Hu, K\. You, K\. Gao, K\. Guan, K\. Huang, K\. Yu, L\. Wang, L\. Zhang, L\. Zhao, L\. Wang, L\. Zhang, L\. Xu, L\. Xia, M\. Zhang, M\. Zhang, M\. Tang, M\. Zhou, M\. Li, M\. Wang, M\. Li, N\. Tian, P\. Huang, P\. Zhang, Q\. Wang, Q\. Chen, Q\. Du, R\. Ge, R\. Zhang, R\. Pan, R\. Wang, R\. J\. Chen, R\. L\. Jin, R\. Chen, S\. Lu, S\. Zhou, S\. Chen, S\. Ye, S\. Wang, S\. Yu, S\. Zhou, S\. Pan, S\. S\. Li, S\. Zhou, S\. Wu, T\. Yun, T\. Pei, T\. Sun, T\. Wang, W\. Zeng, W\. Liu, W\. Liang, W\. Gao, W\. Yu, W\. Zhang, W\. L\. Xiao, W\. An, X\. Liu, X\. Wang, X\. Chen, X\. Nie, X\. Cheng, X\. Liu, X\. Xie, X\. Liu, X\. Yang, X\. Li, X\. Su, X\. Lin, X\. Q\. Li, X\. Jin, X\. Shen, X\. Chen, X\. Sun, X\. Wang, X\. Song, X\. Zhou, X\. Wang, X\. Shan, Y\. K\. Li, Y\. Q\. Wang, Y\. X\. Wei, Y\. Zhang, Y\. Xu, Y\. Li, Y\. Zhao, Y\. Sun, Y\. Wang, Y\. Yu, Y\. Zhang, Y\. Shi, Y\. Xiong, Y\. He, Y\. Piao, Y\. Wang, Y\. Tan, Y\. Ma, Y\. Liu, Y\. Guo, Y\. Ou, Y\. Wang, Y\. Gong, Y\. Zou, Y\. He, Y\. Xiong, Y\. Luo, Y\. You, Y\. Liu, Y\. Zhou, Y\. X\. Zhu, Y\. Huang, Y\. Li, Y\. Zheng, Y\. Zhu, Y\. Ma, Y\. Tang, Y\. Zha, Y\. Yan, Z\. Z\. Ren, Z\. Ren, Z\. Sha, Z\. Fu, Z\. Xu, Z\. Xie, Z\. Zhang, Z\. Hao, Z\. Ma, Z\. Yan, Z\. Wu, Z\. Gu, Z\. Zhu, Z\. Liu, Z\. Li, Z\. Xie, Z\. Song, Z\. Pan, Z\. Huang, Z\. Xu, Z\. Zhang, and Z\. Zhang \(2025\)DeepSeek\-r1 incentivizes reasoning in llms through reinforcement learning\.Nature645\(8081\),pp\. 633–638\.External Links:ISSN 1476\-4687,[Link](http://dx.doi.org/10.1038/s41586-025-09422-z),[Document](https://dx.doi.org/10.1038/s41586-025-09422-z)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.10.10.2.2.2.1)\.
- M\. Hodel \(2024\)Addressing the abstraction and reasoning corpus via procedural example generation\.arXiv preprint arXiv:2404\.07353\.External Links:[Link](https://arxiv.org/abs/2404.07353)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p2.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.23.23.16)\.
- J\. Hong, N\. Lee, and J\. Thorne \(2024\)ORPO: monolithic preference optimization without reference model\.Miami, Florida, USA,pp\. 11170–11189\.External Links:[Link](https://aclanthology.org/2024.emnlp-main.626/),[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.626)Cited by:[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px2.p1.1)\.
- E\. J\. Hu, yelong shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. Chen \(2022\)LoRA: low\-rank adaptation of large language models\.External Links:[Link](https://openreview.net/forum?id=nZeVKeeFYf9)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.23.23.15)\.
- K\. Hu, A\. Cy, L\. Qiu, X\. D\. Ding, R\. Wang, Y\. E\. Zhu, J\. Andreas, and K\. He \(2025\)ARC is a vision problem\!\.arXiv preprint arXiv:2511\.14761\.External Links:[Link](https://arxiv.org/abs/2511.14761)Cited by:[§3\.2](https://arxiv.org/html/2606.26530#S3.SS2.p1.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.4.4.4.11.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.4.4.4.12.1),[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px1.p1.1)\.
- A\. Jolicoeur\-Martineau \(2025\)Less is more: recursive reasoning with tiny networks\.arXiv preprint arXiv:2510\.04871\.External Links:[Link](https://arxiv.org/abs/2510.04871)Cited by:[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px1.p1.1)\.
- S\. Kim, P\. Phunyaphibarn, D\. Ahn, and S\. Kim \(2022\)Playgrounds for abstraction and reasoning\.InNeurIPS 2022 Workshop on Neuro Causal and Symbolic AI \(nCSI\),External Links:[Link](https://openreview.net/forum?id=F4RNpByoqP)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p5.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.p1.1)\.
- X\. Lai, Z\. Tian, Y\. Chen, S\. Yang, X\. Peng, and J\. Jia \(2024\)Step\-dpo: step\-wise preference optimization for long\-chain reasoning of llms\.arXiv preprint arXiv:2406\.18629\.External Links:[Link](https://arxiv.org/abs/2406.18629)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p3.1)\.
- S\. LeGris, W\. K\. Vong, B\. M\. Lake, and T\. M\. Gureckis \(2025\)A comprehensive behavioral dataset for the abstraction and reasoning corpus\.Scientific Data12,pp\. 1380\.External Links:[Link](https://www.nature.com/articles/s41597-025-05687-1)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p3.1)\.
- W\. Li, A\. Jan, B\. Ray, J\. Yang, C\. Mao, and K\. Pei \(2025a\)EditLord: learning code transformation rules for code editing\.pp\. 34876–34903\.External Links:[Link](https://proceedings.mlr.press/v267/li25al.html)Cited by:[§3\.4](https://arxiv.org/html/2606.26530#S3.SS4.p1.1)\.
- W\. Li, Y\. Xu, S\. Sanner, and E\. B\. Khalil \(2025b\)Tackling the abstraction and reasoning corpus with vision transformers: the importance of 2d representation, positions, and objects\.Transactions on Machine Learning Research\.Note:External Links:ISSN 2835\-8856,[Link](https://openreview.net/forum?id=Al72Fp0rCg)Cited by:[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px1.p1.1)\.
- Z\. Lu, A\. Zhou, K\. Wang, H\. Ren, W\. Shi, Y\. Yang, J\. Pan, M\. Zhan, and H\. Li \(2025\)Step\-controlled DPO: leveraging stepwise errors for enhancing mathematical reasoning of language models\.Transactions on Machine Learning Research\.Note:External Links:ISSN 2835\-8856,[Link](https://openreview.net/forum?id=jp1AdIcKTj)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p3.1)\.
- Y\. Meng, M\. Xia, and D\. Chen \(2024\)SimPO: simple preference optimization with a reference\-free reward\.External Links:[Link](https://openreview.net/forum?id=3Tzcot1LKb)Cited by:[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px2.p1.1)\.
- J\. Metcalfe \(2017\)Learning from errors\.Annual Review of Psychology68,pp\. 465–489\.External Links:[Link](https://www.annualreviews.org/doi/10.1146/annurev-psych-010416-044022)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p3.1)\.
- A\. K\. Moskvichev, V\. V\. Odouard, and M\. Mitchell \(2023\)The conceptARC benchmark: evaluating understanding and generalization in the ARC domain\.Transactions on Machine Learning Research\.Note:External Links:ISSN 2835\-8856,[Link](https://openreview.net/forum?id=8ykyGbtt2q)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p5.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.p1.1)\.
- OpenAI \(2024\)OpenAI o1 system card\.External Links:[Link](https://cdn.openai.com/o1-system-card-20241205.pdf)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.9.9.1.1.1.1)\.
- OpenAI \(2025\)OpenAI o3 and o4\-mini system card\.External Links:[Link](https://cdn.openai.com/pdf/2221c875-02dc-4789-800b-e7758f3722c1/o3-and-o4-mini-system-card.pdf)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.15.15.7.7.7.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.16.16.8.8.8.1)\.
- OpenAI \(2026\)GPT\-5\.4 thinking system card\.External Links:[Link](https://deploymentsafety.openai.com/gpt-5-4-thinking/gpt-5-4-thinking.pdf)Cited by:[§3\.4](https://arxiv.org/html/2606.26530#S3.SS4.p2.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.3.3.3.3.1)\.
- L\. Ouyang, J\. Wu, X\. Jiang, D\. Almeida, C\. Wainwright, P\. Mishkin, C\. Zhang, S\. Agarwal, K\. Slama, A\. Ray, J\. Schulman, J\. Hilton, F\. Kelton, L\. Miller, M\. Simens, A\. Askell, P\. Welinder, P\. F\. Christiano, J\. Leike, and R\. Lowe \(2022\)Training language models to follow instructions with human feedback\.35,pp\. 27730–27744\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2022/file/b1efde53be364a73914f58805a001731-Paper-Conference.pdf)Cited by:[§2](https://arxiv.org/html/2606.26530#S2.SS0.SSS0.Px3.p1.1),[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px2.p1.1)\.
- J\. Pourcel, C\. Colas, and P\. Oudeyer \(2025\)Self\-improving language models for evolutionary program synthesis: a case study on ARC\-AGI\.pp\. 49659–49688\.External Links:[Link](https://proceedings.mlr.press/v267/pourcel25a.html)Cited by:[§3\.4](https://arxiv.org/html/2606.26530#S3.SS4.p1.1)\.
- L\. Qiu, L\. Jiang, X\. Lu, M\. Sclar, V\. Pyatkin, C\. Bhagavatula, B\. Wang, Y\. Kim, Y\. Choi, N\. Dziri, and X\. Ren \(2024\)Phenomenal yet puzzling: testing inductive reasoning capabilities of language models with hypothesis refinement\.External Links:[Link](https://openreview.net/forum?id=bNt7oajl2a)Cited by:[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px1.p1.1)\.
- R\. Rafailov, A\. Sharma, E\. Mitchell, C\. D\. Manning, S\. Ermon, and C\. Finn \(2023\)Direct preference optimization: your language model is secretly a reward model\.InAdvances in Neural Information Processing Systems,A\. Oh, T\. Naumann, A\. Globerson, K\. Saenko, M\. Hardt, and S\. Levine \(Eds\.\),Vol\.36,pp\. 53728–53741\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2023/file/a85b405ed65c6477a4fe8302b5e06ce7-Paper-Conference.pdf)Cited by:[§3\.5](https://arxiv.org/html/2606.26530#S3.SS5.p1.3),[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px2.p1.1)\.
- W\. Shu, X\. Qiu, R\. Zhu, H\. H\. Chen, Y\. Liu, and H\. Yang \(2026\)LoopViT: scaling visual arc with looped transformers\.arXiv preprint arXiv:2602\.02156\.External Links:[Link](https://arxiv.org/abs/2602.02156)Cited by:[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px1.p1.1)\.
- M\. Simpson and S\. Banerjee \(2026\)Decomposing ARC programs to create simpler tasks\.External Links:[Link](https://openreview.net/forum?id=KDXHIVfkEz)Cited by:[§3\.3](https://arxiv.org/html/2606.26530#S3.SS3.p1.1)\.
- I\. Sorokin and J\. F\. Puget \(2025\)NVARC solution to arc\-agi\-2 2025\.External Links:[Link](https://github.com/1ytic/NVARC/blob/main/nvarc_2025.pdf)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p2.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.16.16.8.8.20.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.23.23.16),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.4.4.4.21.1)\.
- S\. T\. Sreenivas, S\. Muralidharan, R\. Joshi, M\. Chochowski, A\. S\. Mahabaleshwarkar, G\. Shen, J\. Zeng, Z\. Chen, Y\. Suhara, S\. Diao,et al\.\(2024\)Llm pruning and distillation in practice: the minitron approach\.arXiv preprint arXiv:2408\.11796\.External Links:[Link](https://arxiv.org/abs/2408.11796)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.16.16.8.8.16.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.4.4.4.17.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.p3.1)\.
- S\. Vahdati, A\. Aioanei, H\. Suresh, and J\. Lehmann \(2026\)The arc of progress towards agi: a living survey of abstraction and reasoning\.arXiv preprint arXiv:2603\.13372\.External Links:[Link](https://arxiv.org/abs/2603.13372)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p2.1)\.
- G\. Wang, J\. Li, Y\. Sun, X\. Chen, C\. Liu, Y\. Wu, M\. Lu, S\. Song, and Y\. A\. Yadkori \(2025\)Hierarchical reasoning model\.arXiv preprint arXiv:2506\.21734\.External Links:[Link](https://arxiv.org/abs/2506.21734)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.4.4.4.8.1),[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px1.p1.1)\.
- R\. Wang, E\. Zelikman, G\. Poesia, Y\. Pu, N\. Haber, and N\. D\. Goodman \(2024a\)Hypothesis search: inductive reasoning with language models\.External Links:[Link](https://openreview.net/forum?id=G7UtIGQmjm)Cited by:[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px1.p1.1)\.
- Y\. Wang, S\. Cheng, Z\. Sun, P\. Li, and Y\. Liu \(2024b\)Speak it out: solving symbol\-related problems with symbol\-to\-language conversion for language models\.External Links:[Link](https://openreview.net/forum?id=USvhSx3UJv)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p1.1)\.
- Y\. Wang and H\. Huang \(2025\)Improving abstract reasoning ability of large language models through mixture program\-based data synthesis\.Jinan, China,pp\. 904–921\.External Links:[Link](https://aclanthology.org/2025.ccl-1.69/)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p2.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.23.23),[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px1.p1.1)\.
- xAI \(2025\)Grok 4 model card\.External Links:[Link](https://data.x.ai/2025-08-20-grok-4-model-card.pdf)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.4.4.4.4.1)\.
- Y\. Xu, W\. Li, P\. Vaezipoor, S\. Sanner, and E\. B\. Khalil \(2024\)LLMs and the abstraction and reasoning corpus: successes, failures, and the importance of object\-based representations\.Transactions on Machine Learning Research\.Note:External Links:ISSN 2835\-8856,[Link](https://openreview.net/forum?id=E8m8oySvPJ)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p5.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.p1.1),[§6](https://arxiv.org/html/2606.26530#S6.SS0.SSS0.Px1.p1.1)\.
- A\. Yang, A\. Li, B\. Yang, B\. Zhang, B\. Hui, B\. Zheng, B\. Yu, C\. Gao, C\. Huang, C\. Lv,et al\.\(2025\)Qwen3 technical report\.arXiv preprint arXiv:2505\.09388\.External Links:[Link](https://arxiv.org/abs/2505.09388)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.16.16.8.8.19.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.4.4.4.20.1),[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.p3.1)\.
- A\. Yang, B\. Yang, B\. Zhang, B\. Hui, B\. Zheng, B\. Yu, C\. Li, D\. Liu, F\. Huang, H\. Wei,et al\.\(2024a\)Qwen2\.5 technical report\.arXiv preprint arXiv:2412\.15115\.External Links:[Link](https://arxiv.org/abs/2412.15115)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.12.12.4.4.4.1)\.
- J\. Yang, H\. Jin, R\. Tang, X\. Han, Q\. Feng, H\. Jiang, S\. Zhong, B\. Yin, and X\. Hu \(2024b\)Harnessing the power of llms in practice: a survey on chatgpt and beyond\.ACM Trans\. Knowl\. Discov\. Data18\(6\)\.External Links:ISSN 1556\-4681,[Link](https://doi.org/10.1145/3649506),[Document](https://dx.doi.org/10.1145/3649506)Cited by:[§1](https://arxiv.org/html/2606.26530#S1.p1.1)\.
- T\. Zheng, Y\. Chen, C\. Li, C\. Li, Q\. Zong, H\. Shi, B\. Xu, Y\. Song, G\. Wong, and S\. See \(2025\)The curse of cot: on the limitations of chain\-of\-thought in in\-context learning\.Transactions on Machine Learning Research\.Note:External Links:ISSN 2835\-8856,[Link](https://openreview.net/forum?id=7SIrvcYNYj)Cited by:[§4\.1](https://arxiv.org/html/2606.26530#S4.SS1.23.23)\.
## Appendix AData Statistics
This appendix provides summary statistics of the constructed preference data used in our experiments\. Table[7](https://arxiv.org/html/2606.26530#A1.T7)reports the number of samples for output\-level negative construction on the six ARC\-style benchmarks, together with the additional ARC\-AGI\-1 data used for the DSL\-level inversion and task\-specific editing settings\.
MetricOutput\-LevelDSL\-LevelTask\-SpecificARC\-1ARC\-2MiniARCConcept1DCommunityARC\-1\# Samples15,00012,00020,00012,00012,0002,0006,0004,000Table 7:Summary of data quantities used in our experiments\.
## Appendix BAdditional Training Details
Qwen3\-4BLlama\-3\.2\-3BMinitron\-8BDPOβ\\beta0\.10\.10\.1Learning rate1×10−61\\times 10^\{\-6\}1×10−61\\times 10^\{\-6\}1×10−61\\times 10^\{\-6\}Epochs111Per\-device batch size212Gradient accumulation steps484Effective batch size888LoRA rankrr2563232LoRAα\\alpha321616LoRA dropout0\.050\.050\.05Precision / loadingbf16; fp16 fallbackNF4 4\-bit \+ bf16NF4 4\-bit \+ bf16Optimizerpaged\_adamw\_8bitpaged\_adamw\_8bitpaged\_adamw\_8bitSchedulercosinecosinecosineWarmup ratio0\.10\.10\.1Max sequence length409640964096Max prompt length358435843584Hardware1x NVIDIA L401x NVIDIA L401x NVIDIA L40Table 8:Detailed training settings for the DPO\-based instantiation ofDiARC\. For all models, LoRA is applied toq\_proj,k\_proj,v\_proj,o\_proj,gate\_proj,up\_proj,down\_proj,embed\_tokens, andlm\_head\.
## Appendix CCase Study
We present more representative cases on ARC\-AGI\-1 in Figure[5](https://arxiv.org/html/2606.26530#A3.F5), and on ARC\-AGI\-2 ,MiniARC, ConceptARC, 1D\-ARC, and ARCcommunity in Figure[6](https://arxiv.org/html/2606.26530#A3.F6), respectively\.
Figure 5:Cases on ARC\-AGI\-1, where the base model fails while we predict correctly\.Figure 6:Cases on ARC\-AGI\-2,MiniARC, ConceptARC, 1D\-ARC, and ARCcommunity, where the base model fails while we predict correctly\.
## Appendix DAdditional Visualizations of Negative Construction Strategies
We provide additional visualizations of the negative\-construction operations that are not individually shown in the main pages\.
### D\.1Output\-Level Visual Transformations
Figures[7](https://arxiv.org/html/2606.26530#A4.F7)and[8](https://arxiv.org/html/2606.26530#A4.F8)show additional examples of output\-level visual transformations\.
Figure 7:Additional examples of output\-level visual transformations\.Figure 8:Additional examples of output\-level visual transformations\.
### D\.2DSL\-Level Rule Inversion
Figures[9](https://arxiv.org/html/2606.26530#A4.F9)and[10](https://arxiv.org/html/2606.26530#A4.F10)show additional examples of DSL\-level rule inversion\.
Figure 9:Additional examples of DSL\-level rule inversion\.Figure 10:Additional examples of DSL\-level rule inversion\.
### D\.3Task\-Specific Rule Editing
Figure[11](https://arxiv.org/html/2606.26530#A4.F11)shows additional examples of task\-specific rule editing\. We used the prompt template in Figure[12](https://arxiv.org/html/2606.26530#A5.F12)to guide GPT\-5\.4 in editing task\-specific transformation rules and constructing negative samples\.
Figure 11:Additional examples of task\-specific rule editing\.
## Appendix EAI Assistance Statement
We utilized GPT\-5\.4 to assist in polishing the text, refining grammar, and improving the clarity of the manuscript\. All content was reviewed and verified by the authors\. We also employed GPT\-5\.4 in the task\-specific rule editing process described in Appendix[D\.3](https://arxiv.org/html/2606.26530#A4.SS3)\. The generated content was reviewed and verified by the authors\.
Figure 12:Prompt template used for task\-specific rule editing with GPT\-5\.4\.Similar Articles
PotARCin: Multi-Dimensional Evaluation of Skill Acquisition in Abstract Reasoning Tasks
PotARCin is a multi-dimensional benchmark that extends ARC to evaluate abstract reasoning skills across five dimensions, showing that standard evaluations may not fully capture AI models' true capabilities.
GraphARC: A Comprehensive Benchmark for Graph-Based Abstract Reasoning
GraphARC is a new benchmark for abstract reasoning on graph-structured data, extending the ARC paradigm to graphs. Evaluations of state-of-the-art language models reveal a comprehension-execution gap and performance degradation on larger instances, highlighting scaling challenges.
RL-ARC: Calibrating Large Reasoning Models via Reasoning-guided Uncertainty
本文提出 RL-ARC,一个校准感知的训练框架,通过联合利用推理置信度与答案置信度来校准大型推理模型,缓解 RLVR 训练导致的过度自信问题,并在 ID 和 OOD 设置下提升校准表现而不显著牺牲推理性能。
From Reasoning Depth to Reasoning Breadth: Evaluating Multi-Point Associative Reasoning in Large Language Models
This paper introduces MPAR-Bench, a bilingual benchmark for evaluating multi-point associative reasoning in large language models, along with a perturbation suite and coarse-to-fine evaluation protocol. Results show that deeper reasoning does not automatically confer robust reasoning breadth.
ResRL: Boosting LLM Reasoning via Negative Sample Projection Residual Reinforcement Learning
This paper introduces ResRL, a method to boost LLM reasoning by decoupling semantic distributions between positive and negative responses through negative sample projection. It aims to maintain generation diversity while improving performance on various benchmarks.