The Abstraction and Reasoning Corpus (ARC) poses a significant challenge to artificial intelligence, demanding broad generalization and few-shot learning capabilities that remain elusive for current deep learning methods, including large language models (LLMs). While LLMs excel in program synthesis, their direct application to ARC yields limited success. To address this, we introduce ConceptSearch, a novel function-search algorithm that leverages LLMs for program generation and employs a concept-based scoring method to guide the search efficiently. Unlike simplistic pixel-based metrics like Hamming distance, ConceptSearch evaluates programs on their ability to capture the underlying transformation concept reflected in the input-output examples. We explore three scoring functions: Hamming distance, a CNN-based scoring function, and an LLM-based natural language scoring function. Experimental results demonstrate the effectiveness of ConceptSearch, achieving a significant performance improvement over direct prompting with GPT-4. Moreover, our novel concept-based scoring exhibits up to 30% greater efficiency compared to Hamming distance, measured in terms of the number of iterations required to reach the correct solution. These findings highlight the potential of LLM-driven program search when integrated with concept-based guidance for tackling challenging generalization problems like ARC.
As of December 2024, the ARC-AGI benchmark is five years old and remains unbeaten. We believe it is currently the most important unsolved AI benchmark in the world because it seeks to measure generalization on novel tasks -- the essence of intelligence -- as opposed to skill at tasks that can be prepared for in advance. This year, we launched ARC Prize, a global competition to inspire new ideas and drive open progress towards AGI by reaching a target benchmark score of 85\%. As a result, the state-of-the-art score on the ARC-AGI private evaluation set increased from 33\% to 55.5\%, propelled by several frontier AGI reasoning techniques including deep learning-guided program synthesis and test-time training. In this paper, we survey top approaches, review new open-source implementations, discuss the limitations of the ARC-AGI-1 dataset, and share key insights gained from the competition.
Excellent progress has been made recently in solving ARC Challenge problems. However, it seems that new techniques may be required to push beyond 60% accuracy. Even commercial Large Language Models (LLMs) struggle to 'understand' many of the problems (when given the input and output grids), which makes discovering solutions by LLM-lead program search somewhat futile. In this work, LLM 'understanding' is attempted from a stronger starting position : An LLM is given complete solutions to tasks in code, and then asked to explain how the task is being solved at various levels of abstraction. Specifically, the LLM was given code solutions implemented in arc-dsl-llm (an LLM-legible version of Hodel's arc-dsl to obtain: (a) commented code; (b) code refactored into reusable functional chunks; (c) problem solution steps; and (d) high-level problem-solving tactics. We demonstrate that 'Sparks of Abstraction' can be extracted from the LLM output - in a form that could be used in downstream tasks with Local LLMs eligible to enter the ARC Prize. Both the arc-dsl-llm DSL framework (with the re-engineered solutions) and the Gemini LLM-generated data (along with the generation code) are made Open Source.
Program synthesis methods aim to automatically generate programs restricted to a language that can explain a given specification of input-output pairs. While purely symbolic approaches suffer from a combinatorial search space, recent methods leverage neural networks to learn distributions over program structures to narrow this search space significantly, enabling more efficient search. However, for challenging problems, it remains difficult to train models to perform program synthesis in one shot, making test-time search essential. Most neural methods lack structured search mechanisms during inference, relying instead on stochastic sampling or gradient updates, which can be inefficient. In this work, we propose the Latent Program Network (LPN), a general algorithm for program induction that learns a distribution over latent programs in a continuous space, enabling efficient search and test-time adaptation. We explore how to train these networks to optimize for test-time computation and demonstrate the use of gradient-based search both during training and at test time. We evaluate LPN on ARC-AGI, a program synthesis benchmark that evaluates performance by generalizing programs to new inputs rather than explaining the underlying specification. We show that LPN can generalize beyond its training distribution and adapt to unseen tasks by utilizing test-time computation, outperforming algorithms without test-time adaptation mechanisms.
Language models (LMs) have shown impressive performance on tasks within their training distribution, but often struggle with structurally novel tasks even when given a small number of in-context task examples. We investigate the effectiveness of test-time training (TTT) -- temporarily updating model parameters during inference using a loss derived from input data -- as a mechanism for improving LMs' reasoning and few-shot learning capabilities. On the Abstraction and Reasoning Corpus (ARC), performing TTT with in-context examples yields up to $6\times$ higher accuracy compared to fine-tuned baselines -- reaching $53.0\%$ on the public validation set with an 8B-parameter LM and $61.9\%$ when ensembled with program-synthesis methods, matching average human performance. On BIG-Bench Hard (BBH), TTT on in-context examples surpasses standard few-shot prompting in the $10$-shot setting by $7.3$ percentage points ($50.5\%$ to $57.8\%$). Our findings highlight the limitations of in-context learning for novel tasks and demonstrate the potential of test-time training to enhance language model adaptability.
Large language models (LLMs) have shown impressive capabilities, but still struggle with complex reasoning tasks requiring multiple steps. While prompt-based methods like Chain-of-Thought (CoT) can improve LLM reasoning at inference time, optimizing reasoning capabilities during training remains challenging. We introduce LaTent Reasoning Optimization (LaTRO), a principled framework that formulates reasoning as sampling from a latent distribution and optimizes it via variational approaches. LaTRO enables LLMs to concurrently improve both their reasoning process and ability to evaluate reasoning quality, without requiring external feedback or reward models. We validate LaTRO through experiments on GSM8K and ARC-Challenge datasets using multiple model architectures. On GSM8K, LaTRO improves zero-shot accuracy by an average of 12.5% over base models and 9.6% over supervised fine-tuning across Phi-3.5-mini, Mistral-7B, and Llama-3.1-8B. Our findings suggest that pre-trained LLMs possess latent reasoning capabilities that can be unlocked and enhanced through our proposed optimization approach in a self-improvement manner. The code of LaTRO is available at \url{https://github.com/SalesforceAIResearch/LaTRO}.
Touch is a crucial sensing modality that provides rich information about object properties and interactions with the physical environment. Humans and robots both benefit from using touch to perceive and interact with the surrounding environment (Johansson and Flanagan, 2009; Li et al., 2020; Calandra et al., 2017). However, no existing systems provide rich, multi-modal digital touch-sensing capabilities through a hemispherical compliant embodiment. Here, we describe several conceptual and technological innovations to improve the digitization of touch. These advances are embodied in an artificial finger-shaped sensor with advanced sensing capabilities. Significantly, this fingertip contains high-resolution sensors (~8.3 million taxels) that respond to omnidirectional touch, capture multi-modal signals, and use on-device artificial intelligence to process the data in real time. Evaluations show that the artificial fingertip can resolve spatial features as small as 7 um, sense normal and shear forces with a resolution of 1.01 mN and 1.27 mN, respectively, perceive vibrations up to 10 kHz, sense heat, and even sense odor. Furthermore, it embeds an on-device AI neural network accelerator that acts as a peripheral nervous system on a robot and mimics the reflex arc found in humans. These results demonstrate the possibility of digitizing touch with superhuman performance. The implications are profound, and we anticipate potential applications in robotics (industrial, medical, agricultural, and consumer-level), virtual reality and telepresence, prosthetics, and e-commerce. Toward digitizing touch at scale, we open-source a modular platform to facilitate future research on the nature of touch.
Aligning large language models (LLMs) with human intent is critical for enhancing their performance across a variety of tasks. Standard alignment techniques, such as Direct Preference Optimization (DPO), often rely on the binary Bradley-Terry (BT) model, which can struggle to capture the complexities of human preferences -- particularly in the presence of noisy or inconsistent labels and frequent ties. To address these limitations, we introduce the Tie-rank Oriented Bradley-Terry model (TOBT), an extension of the BT model that explicitly incorporates ties, enabling more nuanced preference representation. Building on this, we propose Tie-rank Oriented Direct Preference Optimization (TODO), a novel alignment algorithm that leverages TOBT's ternary ranking system to improve preference alignment. In evaluations on Mistral-7B and Llama 3-8B models, TODO consistently outperforms DPO in modeling preferences across both in-distribution and out-of-distribution datasets. Additional assessments using MT Bench and benchmarks such as Piqa, ARC-c, and MMLU further demonstrate TODO's superior alignment performance. Notably, TODO also shows strong results in binary preference alignment, highlighting its versatility and potential for broader integration into LLM alignment. The implementation details can be found in https://github.com/XXares/TODO.
When learning an input-output mapping from very few examples, is it better to first infer a latent function that explains the examples, or is it better to directly predict new test outputs, e.g. using a neural network? We study this question on ARC by training neural models for induction (inferring latent functions) and transduction (directly predicting the test output for a given test input). We train on synthetically generated variations of Python programs that solve ARC training tasks. We find inductive and transductive models solve different kinds of test problems, despite having the same training problems and sharing the same neural architecture: Inductive program synthesis excels at precise computations, and at composing multiple concepts, while transduction succeeds on fuzzier perceptual concepts. Ensembling them approaches human-level performance on ARC.
This paper introduces a novel method for reconstructing cone beam computed tomography (CBCT) images for arbitrary orbits using a differentiable shift-variant filtered backprojection (FBP) neural network. Traditional CBCT reconstruction methods for arbitrary orbits, like iterative reconstruction algorithms, are computationally expensive and memory-intensive. The proposed method addresses these challenges by employing a shift-variant FBP algorithm optimized for arbitrary trajectories through a deep learning approach that adapts to a specific orbit geometry. This approach overcomes the limitations of existing techniques by integrating known operators into the learning model, minimizing the number of parameters, and improving the interpretability of the model. The proposed method is a significant advancement in interventional medical imaging, particularly for robotic C-arm CT systems, enabling faster and more accurate CBCT reconstructions with customized orbits. Especially this method can also be used for the analytical reconstruction of non-continuous orbits like circular plus arc. The experimental results demonstrate that the proposed method significantly accelerates the reconstruction process compared to conventional iterative algorithms. It achieves comparable or superior image quality, as evidenced by metrics such as the mean squared error (MSE), the peak signal-to-noise ratio (PSNR), and the structural similarity index measure (SSIM). The validation experiments show that the method can handle data from different trajectories, demonstrating its flexibility and robustness across different scan geometries. Our method demonstrates a significant improvement, particularly for the sinusoidal trajectory, achieving a 38.6% reduction in MSE, a 7.7% increase in PSNR, and a 5.0% improvement in SSIM. Furthermore, the computation time for reconstruction was reduced by more than 97%.
Large Language Models (LLMs) are widely used for tasks such as natural language and code generation, but their outputs often suffer from issues like hallucination, toxicity, and incorrect results. Current libraries for structured LLM generation rely on left-to-right decoding without support for backtracking, limiting the ability to correct or refine outputs mid-generation. To address this, we introduce IterGen, a user-friendly library for iterative, grammar-guided LLM generation that enables users to move both forward and backward within the generated output based on grammar symbols. By leveraging a symbol-to-position mapping and maintaining the key-value (KV) cache state, IterGen ensures efficient and structured generation while allowing for corrections during the process. We demonstrate IterGen's effectiveness in two important applications: reducing privacy leakage in LLM outputs and improving the accuracy of LLM-generated SQL and Vega-Lite queries. Our code and additional resources are available at https://structuredllm.com.
English-centric large language models (LLMs) often show strong multilingual capabilities. However, the multilingual performance of these models remains unclear and is not thoroughly evaluated for many languages. Most benchmarks for multilinguality focus on classic NLP tasks, or cover a minimal number of languages. We introduce MEXA, a method for assessing the multilingual capabilities of pre-trained English-centric LLMs using parallel sentences, which are available for more languages than existing downstream tasks. MEXA leverages the fact that English-centric LLMs use English as a kind of pivot language in their intermediate layers. It computes the alignment between English and non-English languages using parallel sentences to evaluate the transfer of language understanding from English to other languages. This alignment can be used to estimate model performance in other languages. We conduct studies using various parallel datasets (FLORES-200 and Bible), models (Llama family, Gemma family, Mistral, and OLMo), and established downstream tasks (Belebele, m-MMLU, and m-ARC). We explore different methods to compute embeddings in decoder-only models. Our results show that MEXA, in its default settings, achieves a statistically significant average Pearson correlation of 0.90 with three established downstream tasks across nine models and two parallel datasets. This suggests that MEXA is a reliable method for estimating the multilingual capabilities of English-centric LLMs, providing a clearer understanding of their multilingual potential and the inner workings of LLMs. Leaderboard: https://huggingface.co/spaces/cis-lmu/Mexa, Code: https://github.com/cisnlp/Mexa.
The Abstraction and Reasoning Corpus (ARC) is a popular benchmark focused on visual reasoning in the evaluation of Artificial Intelligence systems. In its original framing, an ARC task requires solving a program synthesis problem over small 2D images using a few input-output training pairs. In this work, we adopt the recently popular data-driven approach to the ARC and ask whether a Vision Transformer (ViT) can learn the implicit mapping, from input image to output image, that underlies the task. We show that a ViT -- otherwise a state-of-the-art model for images -- fails dramatically on most ARC tasks even when trained on one million examples per task. This points to an inherent representational deficiency of the ViT architecture that makes it incapable of uncovering the simple structured mappings underlying the ARC tasks. Building on these insights, we propose ViTARC, a ViT-style architecture that unlocks some of the visual reasoning capabilities required by the ARC. Specifically, we use a pixel-level input representation, design a spatially-aware tokenization scheme, and introduce a novel object-based positional encoding that leverages automatic segmentation, among other enhancements. Our task-specific ViTARC models achieve a test solve rate close to 100% on more than half of the 400 public ARC tasks strictly through supervised learning from input-output grids. This calls attention to the importance of imbuing the powerful (Vision) Transformer with the correct inductive biases for abstract visual reasoning that are critical even when the training data is plentiful and the mapping is noise-free. Hence, ViTARC provides a strong foundation for future research in visual reasoning using transformer-based architectures.
Answer verification identifies correct solutions among candidates generated by large language models (LLMs). Current approaches typically train verifier models by labeling solutions as correct or incorrect based solely on whether the final answer matches the gold answer. However, this approach neglects any flawed rationale in the solution yielding the correct answer, undermining the verifier's ability to distinguish between sound and flawed rationales. We empirically show that in StrategyQA, only 19% of LLM-generated solutions with correct answers have valid rationales, thus leading to an unreliable verifier. Furthermore, we demonstrate that training a verifier on valid rationales significantly improves its ability to distinguish valid and flawed rationale. To make a better verifier without extra human supervision, we introduce REPS (Rationale Enhancement through Pairwise Selection), a method for selecting valid rationales from candidates by iteratively applying pairwise self-evaluation using the same LLM that generates the solutions. Verifiers trained on solutions selected by REPS outperform those trained using conventional training methods on three reasoning benchmarks (ARC-Challenge, DROP, and StrategyQA). Our results suggest that training reliable verifiers requires ensuring the validity of rationales in addition to the correctness of the final answers, which would be critical for models assisting humans in solving complex reasoning tasks.
Cellular automata have become a cornerstone for investigating emergence and self-organization across diverse scientific disciplines, spanning neuroscience, artificial life, and theoretical physics. However, the absence of a hardware-accelerated cellular automata library limits the exploration of new research directions, hinders collaboration, and impedes reproducibility. In this work, we introduce CAX (Cellular Automata Accelerated in JAX), a high-performance and flexible open-source library designed to accelerate cellular automata research. CAX offers cutting-edge performance and a modular design through a user-friendly interface, and can support both discrete and continuous cellular automata with any number of dimensions. We demonstrate CAX's performance and flexibility through a wide range of benchmarks and applications. From classic models like elementary cellular automata and Conway's Game of Life to advanced applications such as growing neural cellular automata and self-classifying MNIST digits, CAX speeds up simulations up to 2,000 times faster. Furthermore, we demonstrate CAX's potential to accelerate research by presenting a collection of three novel cellular automata experiments, each implemented in just a few lines of code thanks to the library's modular architecture. Notably, we show that a simple one-dimensional cellular automaton can outperform GPT-4 on the 1D-ARC challenge.
We propose to use the family of Mat\'ern kernels for tunable implicit surface reconstruction, building upon the recent success of kernel methods for 3D reconstruction of oriented point clouds. As we show, both, from a theoretical and practical perspective, Mat\'ern kernels have some appealing properties which make them particularly well suited for surface reconstruction -- outperforming state-of-the-art methods based on the arc-cosine kernel while being significantly easier to implement, faster to compute, and scaleable. Being stationary, we demonstrate that the Mat\'ern kernels' spectrum can be tuned in the same fashion as Fourier feature mappings help coordinate-based MLPs to overcome spectral bias. Moreover, we theoretically analyze Mat\'ern kernel's connection to SIREN networks as well as its relation to previously employed arc-cosine kernels. Finally, based on recently introduced Neural Kernel Fields, we present data-dependent Mat\'ern kernels and conclude that especially the Laplace kernel (being part of the Mat\'ern family) is extremely competitive, performing almost on par with state-of-the-art methods in the noise-free case while having a more than five times shorter training time.
Learning performance data describe correct and incorrect answers or problem-solving attempts in adaptive learning, such as in intelligent tutoring systems (ITSs). Learning performance data tend to be highly sparse (80\%\(\sim\)90\% missing observations) in most real-world applications due to adaptive item selection. This data sparsity presents challenges to using learner models to effectively predict future performance explore new hypotheses about learning. This article proposes a systematic framework for augmenting learner data to address data sparsity in learning performance data. First, learning performance is represented as a three-dimensional tensor of learners' questions, answers, and attempts, capturing longitudinal knowledge states during learning. Second, a tensor factorization method is used to impute missing values in sparse tensors of collected learner data, thereby grounding the imputation on knowledge tracing tasks that predict missing performance values based on real observations. Third, a module for generating patterns of learning is used. This study contrasts two forms of generative Artificial Intelligence (AI), including Generative Adversarial Networks (GANs) and Generate Pre-Trained Transformers (GPT) to generate data associated with different clusters of learner data. We tested this approach on an adult literacy dataset from AutoTutor lessons developed for Adult Reading Comprehension (ARC). We found that: (1) tensor factorization improved the performance in tracing and predicting knowledge mastery compared with other knowledge tracing techniques without data augmentation, showing higher relative fidelity for this imputation method, and (2) the GAN-based simulation showed greater overall stability and less statistical bias based on a divergence evaluation with varying simulation sample sizes compared to GPT.
This paper introduces ARCLE, an environment designed to facilitate reinforcement learning research on the Abstraction and Reasoning Corpus (ARC). Addressing this inductive reasoning benchmark with reinforcement learning presents these challenges: a vast action space, a hard-to-reach goal, and a variety of tasks. We demonstrate that an agent with proximal policy optimization can learn individual tasks through ARCLE. The adoption of non-factorial policies and auxiliary losses led to performance enhancements, effectively mitigating issues associated with action spaces and goal attainment. Based on these insights, we propose several research directions and motivations for using ARCLE, including MAML, GFlowNets, and World Models.
Syntactic Transformer language models aim to achieve better generalization through simultaneously modeling syntax trees and sentences. While prior work has been focusing on adding constituency-based structures to Transformers, we introduce Dependency Transformer Grammars (DTGs), a new class of Transformer language model with explicit dependency-based inductive bias. DTGs simulate dependency transition systems with constrained attention patterns by modifying attention masks, incorporate the stack information through relative positional encoding, and augment dependency arc representation with a combination of token embeddings and operation embeddings. When trained on a dataset of sentences annotated with dependency trees, DTGs achieve better generalization while maintaining comparable perplexity with Transformer language model baselines. DTGs also outperform recent constituency-based models, showing that dependency can better guide Transformer language models. Our code is released at https://github.com/zhaoyd1/Dep_Transformer_Grammars.
Large Language Models (LLMs) vary in their abilities on a range of tasks. Initiatives such as the $\texttt{Open LLM Leaderboard}$ aim to quantify these differences with several large benchmarks (sets of test items to which an LLM can respond either correctly or incorrectly). However, high correlations within and between benchmark scores suggest that (1) there exists a small set of common underlying abilities that these benchmarks measure, and (2) items tap into redundant information and the benchmarks may thus be considerably compressed. We use data from $n > 5000$ LLMs to identify the most informative items of six benchmarks, ARC, GSM8K, HellaSwag, MMLU, TruthfulQA and WinoGrande (with $d=28,632$ items in total). From them we distill a sparse benchmark, $\texttt{metabench}$, that has less than $3\%$ of the original size of all six benchmarks combined. This new sparse benchmark goes beyond point scores by yielding estimators of the underlying benchmark-specific abilities. We show that these estimators (1) can be used to reconstruct each original $\textit{individual}$ benchmark score with, on average, $1.5\%$ root mean square error (RMSE), (2) reconstruct the original $\textit{total}$ score with $0.8\%$ RMSE, and (3) have a single underlying common factor whose Spearman correlation with the total score is $r = 0.93$.