We study transformer language models, analyzing attention heads whose attention patterns are spread out, and whose attention scores depend weakly on content. We argue that the softmax denominators of these heads are stable when the underlying token distribution is fixed. By sampling softmax denominators from a "calibration text", we can combine together the outputs of multiple such stable heads in the first layer of GPT2-Small, approximating their combined output by a linear summary of the surrounding text. This approximation enables a procedure where from the weights alone - and a single calibration text - we can uncover hundreds of first layer neurons that respond to high-level contextual properties of the surrounding text, including neurons that didn't activate on the calibration text.

The combination of linear transformations and non-linear activation functions forms the foundation of most modern deep neural networks, enabling them to approximate highly complex functions. This paper explores the introduction of quadratic transformations to further increase nonlinearity in neural networks, with the aim of enhancing the performance of existing architectures. To reduce parameter complexity and computational complexity, we propose a lightweight quadratic enhancer that uses low-rankness, weight sharing, and sparsification techniques. For a fixed architecture, the proposed approach introduces quadratic interactions between features at every layer, while only adding negligible amounts of additional model parameters and forward computations. We conduct a set of proof-of-concept experiments for the proposed method across three tasks: image classification, text classification, and fine-tuning large-language models. In all tasks, the proposed approach demonstrates clear and substantial performance gains.

Test-time adaptation (TTA) aims to transfer knowledge from a source model to unknown test data with potential distribution shifts in an online manner. Many existing TTA methods rely on entropy as a confidence metric to optimize the model. However, these approaches are sensitive to the predefined entropy threshold, influencing which samples are chosen for model adaptation. Consequently, potentially reliable target samples are often overlooked and underutilized. For instance, a sample's entropy might slightly exceed the threshold initially, but fall below it after the model is updated. Such samples can provide stable supervised information and offer a normal range of gradients to guide model adaptation. In this paper, we propose a general approach, \underline{POEM}, to promote TTA via ex\underline{\textbf{p}}loring the previously unexpl\underline{\textbf{o}}red reliabl\underline{\textbf{e}} sa\underline{\textbf{m}}ples. Additionally, we introduce an extra Adapt Branch network to strike a balance between extracting domain-agnostic representations and achieving high performance on target data. Comprehensive experiments across multiple architectures demonstrate that POEM consistently outperforms existing TTA methods in both challenging scenarios and real-world domain shifts, while remaining computationally efficient. The effectiveness of POEM is evaluated through extensive analyses and thorough ablation studies. Moreover, the core idea behind POEM can be employed as an augmentation strategy to boost the performance of existing TTA approaches. The source code is publicly available at \emph{https://github.com/ycarobot/POEM}

A recent breakthrough in nonconvex optimization is the online-to-nonconvex conversion framework of [Cutkosky et al., 2023], which reformulates the task of finding an $\varepsilon$-first-order stationary point as an online learning problem. When both the gradient and the Hessian are Lipschitz continuous, instantiating this framework with two different online learners achieves a complexity of $O(\varepsilon^{-1.75}\log(1/\varepsilon))$ in the deterministic case and a complexity of $O(\varepsilon^{-3.5})$ in the stochastic case. However, this approach suffers from several limitations: (i) the deterministic method relies on a complex double-loop scheme that solves a fixed-point equation to construct hint vectors for an optimistic online learner, introducing an extra logarithmic factor; (ii) the stochastic method assumes a bounded second-order moment of the stochastic gradient, which is stronger than standard variance bounds; and (iii) different online learning algorithms are used in the two settings. In this paper, we address these issues by introducing an online optimistic gradient method based on a novel doubly optimistic hint function. Specifically, we use the gradient at an extrapolated point as the hint, motivated by two optimistic assumptions: that the difference between the hint and the target gradient remains near constant, and that consecutive update directions change slowly due to smoothness. Our method eliminates the need for a double loop and removes the logarithmic factor. Furthermore, by simply replacing full gradients with stochastic gradients and under the standard assumption that their variance is bounded by $σ^2$, we obtain a unified algorithm with complexity $O(\varepsilon^{-1.75} + σ^2 \varepsilon^{-3.5})$, smoothly interpolating between the best-known deterministic rate and the optimal stochastic rate.

The impressive performance gains of modern language models currently rely on scaling parameters: larger models store more world knowledge and reason better. Yet compressing all world knowledge into parameters is unnecessary, as only a fraction is used per prompt, and impractical for edge devices with limited inference-time memory and compute. We address this shortcoming by a memory-augmented architecture and a pretraining strategy aligned with existing hardware paradigms. We introduce small language models that access large hierarchical parametric memory banks encoding world knowledge. During pretraining and inference, we fetch a small, context-dependent memory block and add it to the model. Our pretraining learns to store long-tail world knowledge in the memory parameters, while the small language model acts as an anchor capturing common knowledge and general reasoning abilities. Through trillion-token-scale experiments, we show significant gains: a 160M-parameters model augmented with an 18M-parameters memory fetched from a 4.6B memory bank obtains comparable performance to a regular model with more than 2x the parameters. Through extensive experiments, we study the optimal type and size of parametric memories in transformers, scaling them to over 21B parameters. We find that our proposed hierarchical feed-forward memories work robustly across transformer architectures, whether added during pretraining or post-hoc.

In this short paper, we present work evaluating an AI agent's understanding of spoken conversations about data visualizations in an online meeting scenario. There is growing interest in the development of AI-assistants that support meetings, such as by providing assistance with tasks or summarizing a discussion. The quality of this support depends on a model that understands the conversational dialogue. To evaluate this understanding, we introduce a dual-axis testing framework for diagnosing the AI agent's comprehension of spoken conversations about data. Using this framework, we designed a series of tests to evaluate understanding of a novel corpus of 72 spoken conversational dialogues about data visualizations. We examine diverse pipelines and model architectures, LLM vs VLM, and diverse input formats for visualizations (the chart image, its underlying source code, or a hybrid of both) to see how this affects model performance on our tests. Using our evaluation methods, we found that text-only input modalities achieved the best performance (96%) in understanding discussions of visualizations in online meetings.

Recent advances in large language models (LLMs) highlight the importance of training data structure and quality in shaping reasoning behavior. However, most existing approaches focus on transforming data formats while neglecting the internal reasoning complexity of training samples, leaving the reasoning potential of data under-explored and underutilized. In this work, we posit that LLM logical reasoning performance is jointly constrained by the potential of the training data and the cognitive capacity of the model. To make this relationship measurable, we introduce Data Reasoning Intensity (DRI), a novel metric that quantifies the latent logical reasoning complexity of samples by decomposing and aggregating their logical structures. This allows us to analyze how well current LLMs utilize logical reasoning signals and identify performance gaps relative to data potential. Based on this insight, we introduce a re-cognizing optimization strategy that systematically enhances the logical reasoning intensity of training data. Rather than increasing data volume, our method re-optimizes existing samples to better align with the LLM's logical reasoning boundary. Extensive experiments show that our approach significantly improves performance and generalization over data-centric strategies. We further validate our method under a reinforcement learning framework. Our results indicate that prioritizing reasoning complexity in data rather than sheer scale or superficial form is essential to realizing LLMs' full cognitive potential.

We present NanoFlux, a novel adversarial framework for generating targeted training data to improve LLM reasoning, where adversarially-generated datasets of 200 examples outperform conventional fine-tuning approaches. The framework employs a competitive dynamic between models alternating as Attacker and Defender, supervised by a tool-augmented Judge, synthesizing multi-step questions with explanatory annotations that target specific reasoning capabilities. Fine-tuning a 4B-parameter model on NanoFlux-generated data yields performance gains across diverse domains compared to full-benchmark fine-tuning: +5.9% on mathematical reasoning (GSMHard), +3.6% on scientific reasoning (GenomeBench), and +16.6% on medical reasoning (MultiMedQA), while reducing computational requirements by 3-14 . Ablation studies reveal a nonmonotonic relationship between dataset characteristics and model performance, uncovering domain-specific optimal points for question complexity and reasoning quality. NanoFlux automates training data generation through embedding-based novelty filtering, tool-augmented evaluation, and multi-hop reasoning, suggesting that future model improvements may lie in the intelligent synthesis of small, precisely targeted training datasets. As large language models (LLMs) rapidly approach and surpass human-level performance on established benchmarks, we confront a fundamental limitation: the finite nature of high-quality training data. Today's frontier models have effectively consumed the entirety of available text on the internet, yet continue to exhibit critical reasoning failures and knowledge gaps. This "benchmark exhaustion" phenomenon raises crucial questions about how to advance AI capabilities beyond the constraints of existing data. While generating synthetic training examples represents one potential path forward, creating effective synthetic data remains challenging - naive generation approaches often produce low-information samples that fail to improve model performance, while synthesizing effective datasets typically requires precisely the kind of human expertise and curation that we seek to automate. Recent work, notably LIMO (Y e et al., 2025), has demonstrated that small, carefully cu-rated datasets of high-quality chain-of-thought solutions can unlock strong reasoning performance, but still depends on human effort in curation. We introduce NanoFlux, a fully generative adversarial framework that reimagines data-efficient reasoning improvement. NanoFlux orchestrates a competitive dynamic between two models alternating as Attacker and Defender, supervised by a tool-augmented Judge that evaluates responses for accuracy, coherence, and safety (as shown in Figure 1).

Reinforcement learning (RL) and sim-to-real transfer have advanced robotic manipulation of rigid objects. Yet, policies remain brittle when applied to articulated mechanisms due to contact-rich dynamics and under-modeled joint phenomena such as friction, stiction, backlash, and clearances. We address this challenge through dexterous in-hand manipulation of articulated tools using a robotic hand with reduced articulation and kinematic redundancy relative to the human hand. Our controller augments a simulation-trained base policy with a sensor-driven refinement learned from hardware demonstrations, conditioning on proprioception and target articulation states while fusing whole-hand tactile and force feedback with the policy's internal action intent via cross-attention-based integration. This design enables online adaptation to instance-specific articulation properties, stabilizes contact interactions, regulates internal forces, and coordinates coupled-link motion under perturbations. We validate our approach across a diversity of real-world examples, including scissors, pliers, minimally invasive surgical tools, and staplers. We achieve robust transfer from simulation to hardware, improved disturbance resilience, and generalization to previously unseen articulated tools, thereby reducing reliance on precise physical modeling in contact-rich settings.

Abstract--This study introduces a haptic shared control framework designed to teach human drivers advanced driving skills. In this context, shared control refers to a driving mode where the human driver collaborates with an autonomous driving system to control the steering of a vehicle simultaneously. Advanced driving skills are those necessary to safely push the vehicle to its handling limits in high-performance driving such as racing and emergency obstacle avoidance. Previous research has demonstrated the performance and safety benefits of shared control schemes using both subjective and objective evaluations. However, these schemes have not been assessed for their impact on skill acquisition on complex and demanding tasks. Prior research on long-term skill acquisition either applies haptic shared control to simple tasks or employs other feedback methods like visual and auditory aids. T o bridge this gap, this study creates a cyber racing coach framework based on the haptic shared control paradigm and evaluates its performance in helping human drivers acquire high-performance driving skills. The framework introduces (1) an autonomous driving system that is capable of cooperating with humans in a highly performant driving scenario; and (2) a haptic shared control mechanism along with a fading scheme to gradually reduce the steering assistance from autonomy based on the human driver's performance during training. Two benchmarks are considered: self-learning (no assistance) and full assistance during training. Results from a human subject study indicate that the proposed framework helps human drivers develop superior racing skills compared to the benchmarks, resulting in better performance and consistency. Advanced driving skills refer to a set of competencies that go beyond basic driving abilities in terms of situational awareness, hazard perception, risk management, and vehicle handling [1]. They are crucial in high-performance driving tasks such as racing, and can also improve safety in everyday driving [1], [2]. This work has been submitted to the IEEE for possible publication.