In the realm of large language model (LLM), as the size of large models increases, it also brings higher training costs. There is a urgent need to minimize the data size in LLM training. Compared with data selection method, the data distillation method aims to synthesize a small number of data samples to achieve the training effect of the full data set and has better flexibility. Despite its successes in computer vision, the discreteness of text data has hitherto stymied its exploration in natural language processing (NLP). In this work, we proposed a method that involves learning pseudo prompt data based on trajectory matching and finding its nearest neighbor ID to achieve cross-architecture transfer. During the distillation process, we introduce a regularization loss to improve the robustness of our distilled data. To our best knowledge, this is the first data distillation work suitable for text generation tasks such as instruction tuning. Evaluations on two benchmarks, including ARC-Easy and MMLU instruction tuning datasets, established the superiority of our distillation approach over the SOTA data selection method LESS. Furthermore, our method demonstrates a good transferability over LLM structures (i.e., OPT to Llama).
Existing attacks against multimodal language models (MLLMs) primarily communicate instructions through text accompanied by adversarial images. In contrast, we exploit the capabilities of MLLMs to interpret non-textual instructions, specifically, adversarial images or audio generated by our novel method, Con Instruction. We optimize these adversarial examples to align closely with target instructions in the embedding space, revealing the detrimental implications of MLLMs' sophisticated understanding. Unlike prior work, our method does not require training data or preprocessing of textual instructions. While these non-textual adversarial examples can effectively bypass MLLM safety mechanisms, their combination with various text inputs substantially amplifies attack success. We further introduce a new Attack Response Categorization (ARC) framework, which evaluates both the quality of the model's response and its relevance to the malicious instructions. Experimental results demonstrate that Con Instruction effectively bypasses safety mechanisms in multiple vision- and audio-language models, including LLaVA-v1.5, InternVL, Qwen-VL, and Qwen-Audio, evaluated on two standard benchmarks: AdvBench and SafeBench. Specifically, our method achieves the highest attack success rates, reaching 81.3% and 86.6% on LLaVA-v1.5 (13B). On the defense side, we explore various countermeasures against our attacks and uncover a substantial performance gap among existing techniques. Our implementation is made publicly available.
Large language models (LLMs) have shown great success in text modeling tasks across domains. However, natural language exhibits inherent semantic hierarchies and nuanced geometric structure, which current LLMs do not capture completely owing to their reliance on Euclidean operations. Recent studies have also shown that not respecting the geometry of token embeddings leads to training instabilities and degradation of generative capabilities. These findings suggest that shifting to non-Euclidean geometries can better align language models with the underlying geometry of text. We thus propose to operate fully in Hyperbolic space, known for its expansive, scale-free, and low-distortion properties. We thus introduce HELM, a family of HypErbolic Large Language Models, offering a geometric rethinking of the Transformer-based LLM that addresses the representational inflexibility, missing set of necessary operations, and poor scalability of existing hyperbolic LMs. We additionally introduce a Mixture-of-Curvature Experts model, HELM-MICE, where each expert operates in a distinct curvature space to encode more fine-grained geometric structure from text, as well as a dense model, HELM-D. For HELM-MICE, we further develop hyperbolic Multi-Head Latent Attention (HMLA) for efficient, reduced-KV-cache training and inference. For both models, we develop essential hyperbolic equivalents of rotary positional encodings and RMS normalization. We are the first to train fully hyperbolic LLMs at billion-parameter scale, and evaluate them on well-known benchmarks such as MMLU and ARC, spanning STEM problem-solving, general knowledge, and commonsense reasoning. Our results show consistent gains from our HELM architectures -- up to 4% -- over popular Euclidean architectures used in LLaMA and DeepSeek, highlighting the efficacy and enhanced reasoning afforded by hyperbolic geometry in large-scale LM pretraining.
Recent breakthroughs in large language models (LLMs) have effectively improved their reasoning abilities, particularly on mathematical and logical problems that have verifiable answers, through techniques such as supervised finetuning (SFT) and reinforcement learning (RL). Prior research indicates that RL effectively internalizes search strategies, enabling long chain-of-thought (CoT) reasoning, with backtracking emerging naturally as a learned capability. However, the precise benefits of backtracking, specifically, how significantly it contributes to reasoning improvements and the optimal extent of its use, remain poorly understood. In this work, we systematically investigate the dynamics between SFT and RL on eight reasoning tasks: Countdown, Sudoku, Arc 1D, Geometry, Color Cube Rotation, List Functions, Zebra Puzzles, and Self Reference. Our findings highlight that short CoT sequences used in SFT as a warm-up do have moderate contribution to RL training, compared with cold-start RL; however such contribution diminishes when tasks become increasingly difficult. Motivated by this observation, we construct synthetic datasets varying systematically in the number of backtracking steps and conduct controlled experiments to isolate the influence of either the correctness (content) or the structure (i.e., backtrack frequency). We find that (1) longer CoT with backtracks generally induce better and more stable RL training, (2) more challenging problems with larger search space tend to need higher numbers of backtracks during the SFT stage. Additionally, we demonstrate through experiments on distilled data that RL training is largely unaffected by the correctness of long CoT sequences, suggesting that RL prioritizes structural patterns over content correctness. Collectively, our results offer practical insights into designing optimal training strategies to effectively scale reasoning in LLMs.
Integrating structured information has long improved the quality of abstractive summarization, particularly in retaining salient content. In this work, we focus on a specific form of structure: argument roles, which are crucial for summarizing documents in high-stakes domains such as law. We investigate whether instruction-tuned large language models (LLMs) adequately preserve this information. To this end, we introduce Argument Representation Coverage (ARC), a framework for measuring how well LLM-generated summaries capture salient arguments. Using ARC, we analyze summaries produced by three open-weight LLMs in two domains where argument roles are central: long legal opinions and scientific articles. Our results show that while LLMs cover salient argument roles to some extent, critical information is often omitted in generated summaries, particularly when arguments are sparsely distributed throughout the input. Further, we use ARC to uncover behavioral patterns -- specifically, how the positional bias of LLM context windows and role-specific preferences impact the coverage of key arguments in generated summaries, emphasizing the need for more argument-aware summarization strategies.
Large Language Models (LLMs), such as OpenAI's o1 and DeepSeek's R1, excel at advanced reasoning tasks like math and coding via Reinforcement Learning with Verifiable Rewards (RLVR), but still struggle with puzzles solvable by humans without domain knowledge. We introduce Enigmata, the first comprehensive suite tailored for improving LLMs with puzzle reasoning skills. It includes 36 tasks across seven categories, each with 1) a generator that produces unlimited examples with controllable difficulty and 2) a rule-based verifier for automatic evaluation. This generator-verifier design supports scalable, multi-task RL training, fine-grained analysis, and seamless RLVR integration. We further propose Enigmata-Eval, a rigorous benchmark, and develop optimized multi-task RLVR strategies. Our trained model, Qwen2.5-32B-Enigmata, consistently surpasses o3-mini-high and o1 on the puzzle reasoning benchmarks like Enigmata-Eval, ARC-AGI (32.8%), and ARC-AGI 2 (0.6%). It also generalizes well to out-of-domain puzzle benchmarks and mathematical reasoning, with little multi-tasking trade-off. When trained on larger models like Seed1.5-Thinking (20B activated parameters and 200B total parameters), puzzle data from Enigmata further boosts SoTA performance on advanced math and STEM reasoning tasks such as AIME (2024-2025), BeyondAIME and GPQA (Diamond), showing nice generalization benefits of Enigmata. This work offers a unified, controllable framework for advancing logical reasoning in LLMs. Resources of this work can be found at https://seed-enigmata.github.io.
The nociceptive withdrawal reflex (NWR) is a mechanism to mediate interactions and protect the body from damage in a potentially dangerous environment. To better convey warning signals to users of prosthetic arms or autonomous robots and protect them by triggering a proper NWR, it is useful to use a biological representation of temperature information for fast and effective processing. In this work, we present a neuromorphic spiking network for heat-evoked NWR by mimicking the structure and encoding scheme of the reflex arc. The network is trained with the bio-plausible reward modulated spike timing-dependent plasticity learning algorithm. We evaluated the proposed model and three other methods in recent studies that trigger NWR in an experiment with radiant heat. We found that only the neuromorphic model exhibits the spatial summation (SS) effect and temporal summation (TS) effect similar to humans and can encode the reflex strength matching the intensity of the stimulus in the relative spike latency online. The improved bio-plausibility of this neuromorphic model could improve sensory feedback in neural prostheses.
Retrieval-Augmented Generation (RAG) leverages large language models (LLMs) combined with external contexts to enhance the accuracy and reliability of generated responses. However, reliably attributing generated content to specific context segments, context attribution, remains challenging due to the computationally intensive nature of current methods, which often require extensive fine-tuning or human annotation. In this work, we introduce a novel Jensen-Shannon Divergence driven method to Attribute Response to Context (ARC-JSD), enabling efficient and accurate identification of essential context sentences without additional fine-tuning or surrogate modelling. Evaluations on a wide range of RAG benchmarks, such as TyDi QA, Hotpot QA, and Musique, using instruction-tuned LLMs in different scales demonstrate superior accuracy and significant computational efficiency improvements compared to the previous surrogate-based method. Furthermore, our mechanistic analysis reveals specific attention heads and multilayer perceptron (MLP) layers responsible for context attribution, providing valuable insights into the internal workings of RAG models.
Inference-time scaling has attracted much attention which significantly enhance the performance of Large Language Models (LLMs) in complex reasoning tasks by increasing the length of Chain-of-Thought. These longer intermediate reasoning rationales embody various meta-reasoning skills in human cognition, such as reflection and decomposition, being difficult to create and acquire. In this work, we introduce \textit{Self-Reasoning Language Model} (SRLM), where the model itself can synthesize longer CoT data and iteratively improve performance through self-training. By incorporating a few demonstration examples (i.e., 1,000 samples) on how to unfold hidden reasoning chains from existing responses, which act as a reasoning catalyst, we demonstrate that SRLM not only enhances the model's initial performance but also ensures more stable and consistent improvements in subsequent iterations. Our proposed SRLM achieves an average absolute improvement of more than $+2.5$ points across five reasoning tasks: MMLU, GSM8K, ARC-C, HellaSwag, and BBH on two backbone models. Moreover, it brings more improvements with more times of sampling during inference, such as absolute $+7.89$ average improvement with $64$ sampling times, revealing the in-depth, diverse and creative reasoning paths in SRLM against the strong baseline.
To reproduce natural standing-up motion, recent studies have emphasized the importance of coordination between the assisting robot and the human. However, many non-wearable assistive devices have struggled to replicate natural motion trajectories. While wearable devices offer better coordination with the human body, they present challenges in completely isolating mechanical and electrical hazards. To address this, we developed a novel standing-assist robot that integrates features of both wearable and non-wearable systems, aiming to achieve high coordination while maintaining safety. The device employs a four-link mechanism aligned with the human joint structure, designed to reproduce the S-shaped trajectory of the hip and the arc trajectory of the knee during natural standing-up motion. Subject-specific trajectory data were obtained using a gyroscope, and the link lengths were determined to drive the seat along the optimal path. A feedforward speed control using a stepping motor was implemented, and the reproducibility of the trajectory was evaluated based on the geometric constraints of the mechanism. A load-bearing experiment with weights fixed to the seat was conducted to assess the trajectory accuracy under different conditions. Results showed that the reproduction errors for the hip and knee trajectories remained within approximately 4 percent of the seat's total displacement, demonstrating high fidelity to the target paths. In addition, durability testing, thermal safety evaluation, and risk assessment confirmed the reliability and safety of the system for indoor use. These findings suggest that the proposed design offers a promising approach for developing assistive technologies that adapt to individual physical characteristics, with potential applications in elderly care and rehabilitation.
Out-of-distribution (OOD) generalisation is considered a hallmark of human and animal intelligence. To achieve OOD through composition, a system must discover the environment-invariant properties of experienced input-output mappings and transfer them to novel inputs. This can be realised if an intelligent system can identify appropriate, task-invariant, and composable input features, as well as the composition methods, thus allowing it to act based not on the interpolation between learnt data points but on the task-invariant composition of those features. We propose that in order to confirm that an algorithm does indeed learn compositional structures from data, it is not enough to just test on an OOD setup, but one also needs to confirm that the features identified are indeed compositional. We showcase this by exploring two tasks with clearly defined OOD metrics that are not OOD solvable by three commonly used neural networks: a Multi-Layer Perceptron (MLP), a Convolutional Neural Network (CNN), and a Transformer. In addition, we develop two novel network architectures imbued with biases that allow them to be successful in OOD scenarios. We show that even with correct biases and almost perfect OOD performance, an algorithm can still fail to learn the correct features for compositional generalisation.
Precise initialization plays a critical role in the performance of localization algorithms, especially in the context of robotics, autonomous driving, and computer vision. Poor localization accuracy is often a consequence of inaccurate initial poses, particularly noticeable in GNSS-denied environments where GPS signals are primarily relied upon for initialization. Recent advances in leveraging deep neural networks for pose regression have led to significant improvements in both accuracy and robustness, especially in estimating complex spatial relationships and orientations. In this paper, we introduce APR-Transformer, a model architecture inspired by state-of-the-art methods, which predicts absolute pose (3D position and 3D orientation) using either image or LiDAR data. We demonstrate that our proposed method achieves state-of-the-art performance on established benchmark datasets such as the Radar Oxford Robot-Car and DeepLoc datasets. Furthermore, we extend our experiments to include our custom complex APR-BeIntelli dataset. Additionally, we validate the reliability of our approach in GNSS-denied environments by deploying the model in real-time on an autonomous test vehicle. This showcases the practical feasibility and effectiveness of our approach. The source code is available at:https://github.com/GT-ARC/APR-Transformer.
A customizable method (OPTIKS) for designing fast trajectory-constrained gradient waveforms with optimized time domain properties was developed. Given a specified multidimensional k-space trajectory, the method optimizes traversal speed (and therefore timing) with position along the trajectory. OPTIKS facilitates optimization of objectives dependent on the time domain gradient waveform and the arc-length domain k-space speed. OPTIKS is applied to design waveforms which limit peripheral nerve stimulation (PNS), minimize mechanical resonance excitation, and reduce acoustic noise. A variety of trajectory examples are presented including spirals, circular echo-planar-imaging, and rosettes. Design performance is evaluated based on duration, standardized PNS models, field measurements, gradient coil back-EMF measurements, and calibrated acoustic measurements. We show reductions in back-EMF of up to 94% and field oscillations up to 91.1%, acoustic noise decreases of up to 9.22 dB, and with efficient use of PNS models speed increases of up to 11.4%. The design method implementation is made available as an open source Python package through GitHub.
The Abstraction and Reasoning Corpus (ARC-AGI) poses a significant challenge for large language models (LLMs), exposing limitations in their abstract reasoning abilities. In this work, we leverage task-specific data augmentations throughout the training, generation, and scoring phases, and employ a depth-first search algorithm to generate diverse, high-probability candidate solutions. Furthermore, we utilize the LLM not only as a generator but also as a scorer, using its output probabilities to select the most promising solutions. Our method achieves a score of 71.6% (286.5/400 solved tasks) on the public ARC-AGI evaluation set, demonstrating state-of-the-art performance among publicly available approaches. While concurrent closed-source work has reported higher scores, our method distinguishes itself through its transparency, reproducibility, and remarkably low inference cost, averaging only around 2ct per task on readily available hardware (we assume a price of 36ct/hour for a Nvidia 4090 GPU).
Text-to-audio systems, while increasingly performant, are slow at inference time, thus making their latency unpractical for many creative applications. We present Adversarial Relativistic-Contrastive (ARC) post-training, the first adversarial acceleration algorithm for diffusion/flow models not based on distillation. While past adversarial post-training methods have struggled to compare against their expensive distillation counterparts, ARC post-training is a simple procedure that (1) extends a recent relativistic adversarial formulation to diffusion/flow post-training and (2) combines it with a novel contrastive discriminator objective to encourage better prompt adherence. We pair ARC post-training with a number optimizations to Stable Audio Open and build a model capable of generating $\approx$12s of 44.1kHz stereo audio in $\approx$75ms on an H100, and $\approx$7s on a mobile edge-device, the fastest text-to-audio model to our knowledge.
The Abstraction and Reasoning Corpus (ARC), later renamed ARC-AGI, poses a fundamental challenge in artificial general intelligence (AGI), requiring solutions that exhibit robust abstraction and reasoning capabilities across diverse tasks, while only few (with median count of three) correct examples are presented. While ARC-AGI remains very challenging for artificial intelligence systems, it is rather easy for humans. This paper introduces ARC-NCA, a developmental approach leveraging standard Neural Cellular Automata (NCA) and NCA enhanced with hidden memories (EngramNCA) to tackle the ARC-AGI benchmark. NCAs are employed for their inherent ability to simulate complex dynamics and emergent patterns, mimicking developmental processes observed in biological systems. Developmental solutions may offer a promising avenue for enhancing AI's problem-solving capabilities beyond mere training data extrapolation. ARC-NCA demonstrates how integrating developmental principles into computational models can foster adaptive reasoning and abstraction. We show that our ARC-NCA proof-of-concept results may be comparable to, and sometimes surpass, that of ChatGPT 4.5, at a fraction of the cost.
This technical report describes an experiment on autoregressive pre-training of Gemma2 2 billion parameter large language model (LLM) with 10\% on the Lithuanian language component of CulturaX from the point of view of continual learning. We apply elastic weight consolidation (EWC) to the full set of the model's parameters and investigate language understanding benchmarks, consisting of Arc, Belebele, Gsm8K, Hellaswag, MMLU, TruthfulQA, and Winogrande sets (both in English and Lithuanian versions), and perplexity benchmarks. We empirically demonstrate that EWC regularisation allows us not only to mitigate catastrophic forgetting effects but also that it is potentially beneficial for learning of the new task with LLMs.
The Neomycin resistance cassette (Neo+) is commonly inserted in the genome of mice to generate knock-out (KO) models. The effect of gene deletion on social behaviors in mice is controversial between studies using different Neo+ and Neo-mouse lines, particularly Arc/Arg3.1 KO lines. In this study, we identified severe maternal behavior impairments in Neo+, but not Neo-Arc/Arg3.1 KO dams. These deficits resulted from reduced sociability and abnormal social information processing in Neo+ Arc/Arg3.1 KO dams, exacerbated by social communication impairments in pups. The expression of the Neo cassette product did not cause cytotoxicity, but led to altered ERK signaling, gene expression, and oxytocin system. However, oxytocin administration did not improve social impairments in Neo+ Arc/Arg3.1 KO animals. Interestingly, early social environment enrichment enhanced social interaction with familiar, but not unfamiliar conspecifics or maternal behavior. Overall, our findings reveal a major impact of the Neo cassette on behaviors, particularly social behaviors, in Arc/Arg3.1 KO mice, underscoring the need to re-examine phenotypes of animal models carrying the Neo cassette in neuroscience research.
Because large language models are expensive to pretrain on different datasets, using smaller-scale experiments to decide on data is crucial for reducing costs. Which benchmarks and methods of making decisions from observed performance at small scale most accurately predict the datasets that yield the best large models? To empower open exploration of this question, we release models, data, and evaluations in DataDecide -- the most extensive open suite of models over differences in data and scale. We conduct controlled pretraining experiments across 25 corpora with differing sources, deduplication, and filtering up to 100B tokens, model sizes up to 1B parameters, and 3 random seeds. We find that the ranking of models at a single, small size (e.g., 150M parameters) is a strong baseline for predicting best models at our larger target scale (1B) (~80% of com parisons correct). No scaling law methods among 8 baselines exceed the compute-decision frontier of single-scale predictions, but DataDecide can measure improvement in future scaling laws. We also identify that using continuous likelihood metrics as proxies in small experiments makes benchmarks including MMLU, ARC, HellaSwag, MBPP, and HumanEval >80% predictable at the target 1B scale with just 0.01% of the compute.
Language models rely on semantic priors to perform in-context learning, which leads to poor performance on tasks involving inductive reasoning. Instruction-tuning methods based on imitation learning can superficially enhance the in-context learning performance of language models, but they often fail to improve the model's understanding of the underlying rules that connect inputs and outputs in few-shot demonstrations. We propose ReDis, a reasoning distillation technique designed to improve the inductive reasoning capabilities of language models. Through a careful combination of data augmentation, filtering, supervised fine-tuning, and alignment, ReDis achieves significant performance improvements across a diverse range of tasks, including 1D-ARC, List Function, ACRE, and MiniSCAN. Experiments on three language model backbones show that ReDis outperforms equivalent few-shot prompting baselines across all tasks and even surpasses the teacher model, GPT-4o, in some cases. ReDis, based on the LLaMA-3 backbone, achieves relative improvements of 23.2%, 2.8%, and 66.6% over GPT-4o on 1D-ARC, ACRE, and MiniSCAN, respectively, within a similar hypothesis search space. The code, dataset, and model checkpoints will be made available at https://github.com/NafisSadeq/reasoning-distillation.git.