We present Butter-Bench, a benchmark evaluating large language model (LLM) controlled robots for practical intelligence, defined as the ability to navigate the messiness of the physical world. Current state-of-the-art robotic systems use a hierarchical architecture with LLMs in charge of high-level reasoning, and a Vision Language Action (VLA) model for low-level control. Butter-Bench evaluates the LLM part in isolation from the VLA. Although LLMs have repeatedly surpassed humans in evaluations requiring analytical intelligence, we find humans still outperform LLMs on Butter-Bench. The best LLMs score 40% on Butter-Bench, while the mean human score is 95%. LLMs struggled the most with multi-step spatial planning and social understanding. We also evaluate LLMs that are fine-tuned for embodied reasoning and conclude that this training does not improve their score on Butter-Bench. Language models (LMs) were initially intended for narrow text understanding tasks. The first Transformer-based LM (V aswani et al., 2017) was explicitly trained for translation. However, large-scale training runs of LMs eventually resulted in emergent behaviour - model capabilities that were not explicitly trained for (Brown et al., 2020). For example, LLMs are not trained to be robots, yet companies such as Figure (Helix, 2025) and Google DeepMind (Gemini Robotics 1.5, 2025) use LLMs in their robotic stack.

Policy optimization (PO) algorithms are used to refine Large Language Models (LLMs) for complex, multi-step reasoning. Current state-of-the-art pipelines enforce a strict think-then-answer format to elicit chain-of-thought (CoT); however, the behavior of PO when these rigid constraints are relaxed into an open-ended CoT structure remains an under-studied question. We investigate this gap with an extensive suite of controlled experiments and identify a consistent principle: policy optimization consistently follows the path of least resistance. When afforded the flexibility to interleave reasoning and response, policy optimization consistently learns to discard explicit reasoning, causing the policy to degenerate to a direct -only format. This outcome holds true across a rigorous evaluation suite spanning 5 model families (4B-12B), 3 reasoning domains (math, code, logic), and 3 distinct PO algorithms (GRPO, DAPO, REINFORCE++). We find that this collapse in format is persistent even when the complex format is assigned up to 4x larger reward weights. We formalize this principle through a series of controlled reward decomposition experiments, demonstrating a clear hierarchy: PO systematically optimizes for the simplest reward component first, a preference that holds even when faced with mutually exclusive choices or strong incentives for more complex behaviors. Finally, we show that successful convergence on the high-reward shortcut is not a low-effort drift but is driven by the optimization process that requires the KL-regularized policy to have sufficient freedom to make a significant shift from its initial prior. Our findings reveal that granting policies the freedom to diverge is a double-edged sword: while necessary for discovering high-reward shortcuts, it also creates a powerful incentive to game the simplest aspects of the reward function, posing a critical challenge for reward hacking under alignment. The central finding investigated in this paper is the emergent policy learned under this freedom: the model consistently bypasses explicit reasoning to converge on the simplest valid format, a phenomenon we term the Cognitive Shortcut. The community's dominant approach is one of careful enforcement: to elicit a chain-of-thought, we engineer a rigid reward function that strongly compels the model to follow a strict "think-then-answer" format Liu et al. (2025a); DeepSeek-AI et al. (2025); Xie et al. (2025); Chen et al. (2025). While this enforced structure is effective in producing a desired output, it obscures a deeper, more fundamental question that has been largely overlooked: What is the optimizer's innate preference when the external guidance is removed?

We present a differentiable framework that leverages the Discrete Empirical Interpolation Method (DEIM) for interpretable deep learning and dynamical system analysis. Although DEIM efficiently approximates nonlinear terms in projection-based reduced-order models (POD-ROM), its fixed interpolation points limit the adaptability to complex and time-varying dynamics. To address this limitation, we first develop a differentiable adaptive DEIM formulation for the one-dimensional viscous Burgers equation, which allows neural networks to dynamically select interpolation points in a computationally efficient and physically consistent manner. We then apply DEIM as an interpretable analysis tool for examining the learned dynamics of a pre-trained Neural Ordinary Differential Equation (NODE) on a two-dimensional vortex-merging problem. The DEIM trajectories reveal physically meaningful features in the learned dynamics of NODE and expose its limitations when extrapolating to unseen flow configurations. These findings demonstrate that DEIM can serve not only as a model reduction tool but also as a diagnostic framework for understanding and improving the generalization behavior of neural differential equation models.

Precipitation nowcasting based on radar echoes plays a crucial role in monitoring extreme weather and supporting disaster prevention. Although deep learning approaches have achieved significant progress, they still face notable limitations. For example, deterministic models tend to produce over-smoothed predictions, which struggle to capture extreme events and fine-scale precipitation patterns. Probabilistic generative models, due to their inherent randomness, often show fluctuating performance across different metrics and rarely achieve consistently optimal results. Furthermore, precipitation nowcasting is typically evaluated using multiple metrics, some of which are inherently conflicting. For instance, there is often a trade-off between the Critical Success Index (CSI) and the False Alarm Ratio (FAR), making it challenging for existing models to deliver forecasts that perform well on both metrics simultaneously. To address these challenges, we introduce preference optimization into precipitation nowcasting for the first time, motivated by the success of reinforcement learning from human feedback in large language models. Specifically, we propose SynCast, a method that employs the two-stage post-training framework of Diffusion Sequential Preference Optimization (Diffusion-SPO), to progressively align conflicting metrics and consistently achieve superior performance. In the first stage, the framework focuses on reducing FAR, training the model to effectively suppress false alarms. Building on this foundation, the second stage further optimizes CSI with constraints that preserve FAR alignment, thereby achieving synergistic improvements across these conflicting metrics.

Membership inference attacks (MIAs) test whether a data point was part of a model's training set, posing serious privacy risks. Existing methods often depend on shadow models or heavy query access, which limits their practicality. We propose GP-MIA, an efficient and interpretable approach based on Gaussian process (GP) meta-modeling. Using post-hoc metrics such as accuracy, entropy, dataset statistics, and optional sensitivity features (e.g. gradients, NTK measures) from a single trained model, GP-MIA trains a GP classifier to distinguish members from non-members while providing calibrated uncertainty estimates. Experiments on synthetic data, real-world fraud detection data, CIFAR-10, and WikiText-2 show that GP-MIA achieves high accuracy and generalizability, offering a practical alternative to existing MIAs.

Brain-computer interfaces (BCIs) based on motor imagery (MI) translate covert movement intentions into actionable commands, yet reliable decoding from non-invasive EEG remains challenging due to nonstationarity, low SNR, and subject variability. The RDWT performs an undecimated, multi-resolution subband decomposition that preserves temporal length and shift-invariance, enhancing sensorimotor rhythms while mitigating jitter and mild artifacts; subbands are fused via lightweight grouped 1-D convolutions and passed to a multi-kernel CNN for local temporal-spatial feature extraction, a grouped-query attention encoder for long-range context, and a compact TCN head for causal temporal integration. Our goal is to test whether this principled wavelet front end improves robustness precisely where BCIs typically fail--on the hardest subjects--and whether such gains persist on average across seeds under both intra-and inter-subject protocols. On BCI-IV-2a and BCI-IV-2b, across five seeds, RatioWaveNet improves worst-subject accuracy over the Transformer backbone by +0.17 / +0.42 percentage points (Sub-Dependent / LOSO) on 2a and by +1.07 / +2.54 percentage points on 2b, with consistent average-case gains These results indicate that a simple, trainable wavelet front end is an effective plug-in to strengthen Transformer-based BCIs, improving worst-case reliability without sacrificing efficiency. Introduction Brain-computer interfaces (BCIs) establish a direct communication pathway between neural activity and external devices by decoding brain signals into actionable commands, with the potential to transform both clinical rehabilitation and human-computer interactions and thereby improve quality of life [1, 2]. Among brain-sensing modalities, electroencephalography (EEG) is particularly well suited to real-world deployment because it is noninvasive, inexpensive, portable, and offers millisecond-level temporal resolution; these properties have underpinned a broad spectrum of EEG-based applications spanning cognitive skill assessment [3], driver vigilance estimation [4], emotion recognition [5], and human-robot interaction [6]. In healthcare, EEG is a key enabler of smart health ecosystems [7], supporting tasks such as automated sleep staging [8], seizure detection and monitoring [9], and neurorehabilitation after stroke [10]. Within this context, Motor Imagery (MI) - the covert rehearsal of movement without execution - has become one of the most widely adopted BCI paradigms [1].

In this study, we evaluate the ability of OpenAI's gpt-4o model to classify chest X-ray images as either NORMAL or PNEUMONIA in a zero-shot setting, without any prior fine-tuning. A balanced test set of 400 images (200 from each class) was used to assess performance across four distinct prompt designs, ranging from minimal instructions to detailed, reasoning-based prompts. The results indicate that concise, feature-focused prompts achieved the highest classification accuracy of 74\%, whereas reasoning-oriented prompts resulted in lower performance. These findings highlight that while ChatGPT exhibits emerging potential for medical image interpretation, its diagnostic reliability remains limited. Continued advances in visual reasoning and domain-specific adaptation are required before such models can be safely applied in clinical practice.

Anomaly detection in cybersecurity is a challenging task, where normal events far outnumber anomalous ones with new anomalies occurring frequently. Classical autoencoders have been used for anomaly detection, but struggles in data-limited settings which quantum counterparts can potentially overcome. In this work, we apply Quantum Autoencoders (QAEs) for anomaly detection in cybersecurity, specifically on the BPF-extended tracking honeypot (BETH) dataset. QAEs are evaluated across multiple encoding techniques, ansatz types, repetitions, and feature selection strategies. Our results demonstrate that an 8-feature QAE using Dense-Angle encoding with a RealAmplitude ansatz can outperform Classical Autoencoders (CAEs), even when trained on substantially fewer samples. The effects of quantum encoding and feature selection for developing quantum models are demonstrated and discussed. In a data-limited setting, the best performing QAE model has a F1 score of 0.87, better than that of CAE (0.77). These findings suggest that QAEs may offer practical advantages for anomaly detection in data-limited scenarios.

Manually generating catchy descriptions and names is labor intensive and a slow process for retailers. Although generative AI provides an automation solution in form of Vision to Language Models (VLM), the current VLMs are prone to factual "hallucinations". Siloed, single task models are not only inefficient but also fail to capture interdependent relationships between features. To address these challenges, we propose an end to end, multi task system that generates factually grounded textual listings from a single image. The contributions of this study are two proposals for the model architecture. First, application of multi task learning approach for fine tuning a vision encoder where a single vision backbone is jointly trained on attribute prediction such as color, hemline and neck style and price regression. Second, introduction of a hierarchical generation process where the model's own predicted attributes are embedded in a prompt and fed to the text decoder to improve factual consistency. The experiments demonstrate the superiority of this architecture. The multi tasking approach outperforms both the independent price regression, with a 3.6% better R2 Value and attribute classification, with a 6.6% improvement F1 score. Critically, the hierarchical generation process proves highly effective, slashing the factual hallucination rate from 12.7% to 7.1%, a 44.5% relative reduction, compared to a non hierarchical ablation. The hierarchical approach also reduces the latency of the autoregressive text generation process by a factor of 3.5 when compared to direct vision to language model of similar size. One minor caveat is that the model does perform 3.5% worse than direct vision-to-language model on ROUGE-L score.

Pruning is a widely used technique to reduce the size and inference cost of large language models (LLMs), but it often causes performance degradation. To mitigate this, existing restoration methods typically employ parameter-efficient fine-tuning (PEFT), such as LoRA, to recover the pruned model's performance. However, most PEFT methods are designed for dense models and overlook the distinct properties of pruned models, often resulting in suboptimal recovery. In this work, we propose a targeted restoration strategy for pruned models that restores performance while preserving their low cost and high efficiency. We observe that pruning-induced information loss is reflected in attention activations, and selectively reintroducing components of this information can significantly recover model performance. Based on this insight, we introduce RestoreLCC (Restoring Pruned LLMs via Lost Component Compensation), a plug-and-play method that contrastively probes critical attention heads via activation editing, extracts lost components from activation differences, and finally injects them back into the corresponding pruned heads for compensation and recovery. RestoreLCC is compatible with structured, semi-structured, and unstructured pruning schemes. Extensive experiments demonstrate that RestoreLCC consistently outperforms state-of-the-art baselines in both general and task-specific performance recovery, without compromising the sparsity or inference efficiency of pruned models.