Transformers are the de-facto choice for sequence modelling, yet their quadratic self-attention and weak temporal bias can make long-range forecasting both expensive and brittle. We introduce FreezeTST, a lightweight hybrid that interleaves frozen random-feature (reservoir) blocks with standard trainable Transformer layers. The frozen blocks endow the network with rich nonlinear memory at no optimisation cost; the trainable layers learn to query this memory through self-attention. The design cuts trainable parameters and also lowers wall-clock training time, while leaving inference complexity unchanged. On seven standard long-term forecasting benchmarks, FreezeTST consistently matches or surpasses specialised variants such as Informer, Autoformer, and PatchTST; with substantially lower compute. Our results show that embedding reservoir principles within Transformers offers a simple, principled route to efficient long-term time-series prediction.

Biomedical question answering (QA) requires accurate interpretation of complex medical knowledge. Large language models (LLMs) have shown promising capabilities in this domain, with retrieval-augmented generation (RAG) systems enhancing performance by incorporating external medical literature. However, RAG-based approaches in biomedical QA suffer from hallucinations due to post-retrieval noise and insufficient verification of retrieved evidence, undermining response reliability. We propose MedTrust-Guided Iterative RAG, a framework designed to enhance factual consistency and mitigate hallucinations in medical QA. Our method introduces three key innovations. First, it enforces citation-aware reasoning by requiring all generated content to be explicitly grounded in retrieved medical documents, with structured Negative Knowledge Assertions used when evidence is insufficient. Second, it employs an iterative retrieval-verification process, where a verification agent assesses evidence adequacy and refines queries through Medical Gap Analysis until reliable information is obtained. Third, it integrates the MedTrust-Align Module (MTAM) that combines verified positive examples with hallucination-aware negative samples, leveraging Direct Preference Optimization to reinforce citation-grounded reasoning while penalizing hallucination-prone response patterns.

Ensuring the long-term reliability of AI models in clinical practice requires continuous performance monitoring and corrective actions when degradation occurs. Addressing this need, this manuscript presents ReclAIm, a multi-agent framework capable of autonomously monitoring, evaluating, and fine-tuning medical image classification models. The system, built on a large language model core, operates entirely through natural language interaction, eliminating the need for programming expertise. ReclAIm successfully trains, evaluates, and maintains consistent performance of models across MRI, CT, and X-ray datasets. Once ReclAIm detects significant performance degradation, it autonomously executes state-of-the-art fine-tuning procedures that substantially reduce the performance gap. In cases with performance drops of up to -41.1% (MRI InceptionV3), ReclAIm managed to readjust performance metrics within 1.5% of the initial model results. ReclAIm enables automated, continuous maintenance of medical imaging AI models in a user-friendly and adaptable manner that facilitates broader adoption in both research and clinical environments.

Large language models (LLMs) were shown to encode word form variations, such as "walk"->"walked", as linear directions in embedding space. However, standard tokenization algorithms treat these variations as distinct tokens -- filling the size-capped vocabulary with surface form variants (e.g., "walk", "walking", "Walk"), at the expense of less frequent words and multilingual coverage. We show that many of these variations can be captured by transformation vectors -- additive offsets that yield the appropriate word's representation when applied to the base form word embedding -- in both the input and output spaces. Building on this, we propose a compact reshaping of the vocabulary: rather than assigning unique tokens to each surface form, we compose them from shared base form and transformation vectors (e.g., "walked" = "walk" + past tense). We apply our approach to multiple LLMs and across five languages, removing up to 10% of vocabulary entries -- thereby freeing space to allocate new, more diverse tokens. Importantly, we do so while also expanding vocabulary coverage to out-of-vocabulary words, with minimal impact on downstream performance, and without modifying model weights. Our findings motivate a foundational rethinking of vocabulary design, moving from string enumeration to a compositional vocabulary that leverages the underlying structure of language.

The reliable execution of high-level missions in multi-robot systems with heterogeneous agents, requires robust methods for detecting spurious behaviors. In this paper, we address the challenge of identifying spurious executions of plans specified as a Linear Temporal Logic (LTL) formula, as incorrect task sequences, violations of spatial constraints, timing inconsis- tencies, or deviations from intended mission semantics. To tackle this, we introduce a structured data generation framework based on the Nets-within-Nets (NWN) paradigm, which coordinates robot actions with LTL-derived global mission specifications. We further propose a Transformer-based anomaly detection pipeline that classifies robot trajectories as normal or anomalous. Experi- mental evaluations show that our method achieves high accuracy (91.3%) in identifying execution inefficiencies, and demonstrates robust detection capabilities for core mission violations (88.3%) and constraint-based adaptive anomalies (66.8%). An ablation experiment of the embedding and architecture was carried out, obtaining successful results where our novel proposition performs better than simpler representations.

Memory-efficient optimization is critical for training increasingly large language models (LLMs). A popular strategy involves gradient low-rank projection, storing only the projected optimizer states, with GaLore being a representative example. However, a significant drawback of many such methods is their lack of convergence guarantees, as various low-rank projection approaches introduce inherent biases relative to the original optimization algorithms, which contribute to performance gaps compared to full-parameter training. Aiming to tackle this problem, this paper investigates the layerwise sampling technique for debiasing low-rank projection mechanisms. In particular, an instantiation of the paradigm gives rise to a novel and unbiased low-rank optimization method built upon GaLore's mechanism and the Muon algorithm, named GaLore Unbiased with Muon (GUM). We theoretically prove our method matches the convergence guarantees of the base Muon algorithm while preserving the memory efficiency of low-rank techniques. Empirical experiments on LLM fine-tuning and pretraining also demonstrate non-trivial improvements over GaLore and even better performance than full-parameter training. Further investigation shows that the improvement of this technique comes from a more uniform distribution of knowledge inside layers, leading to more efficient utilization of the model parameter space and better memorization.

Scaled post-training now drives many of the largest capability gains in language models (LMs), yet its effect on pretrained knowledge remains poorly understood. Not all forgetting is equal: Forgetting one fact (e.g., a U.S. president or an API call) does not "average out" by recalling another. Hence, we propose a sample-wise paradigm to measure what is forgotten and when backward transfer occurs. Our metric counts 1->0 transitions (correct before post-training, incorrect after) to quantify forgetting and 0->1 transitions to quantify backward transfer. Traditional task averages conflate these effects and obscure large changes. For multiple-choice benchmarks, we add chance-adjusted variants that subtract the expected contribution of random guessing from pre- and post-training accuracies. We apply this framework across post-training stages, model sizes, and data scales. Our large-scale analysis shows that: (1) Domain-continual pretraining induces moderate forgetting with low-to-moderate backward transfer; (2) RL/SFT post-training applied to base models and Instruction tuning yields moderate-to-large backward transfer on math and logic with overall low-to-moderate forgetting; (3) Applying RL/SFT to instruction-tuned models is sensitive on data scale: at small scales, both forgetting and backward transfer are small; at larger scales, effects are mixed and warrant further study with better controls; (4) Model merging does not reliably mitigate forgetting. Overall, our framework offers a practical yardstick for mapping how post-training alters pretrained knowledge at scale -- enabling progress towards generally capable AI systems.

Skin cancer is a life-threatening disease where early detection significantly improves patient outcomes. Automated diagnosis from dermoscopic images is challenging due to high intra-class variability and subtle inter-class differences. Many deep learning models operate as "black boxes," limiting clinical trust. In this work, we propose a dual-encoder attention-based framework that leverages both segmented lesions and clinical metadata to enhance skin lesion classification in terms of both accuracy and interpretability. A novel Deep-UNet architecture with Dual Attention Gates (DAG) and Atrous Spatial Pyramid Pooling (ASPP) is first employed to segment lesions. The classification stage uses two DenseNet201 encoders-one on the original image and another on the segmented lesion whose features are fused via multi-head cross-attention. This dual-input design guides the model to focus on salient pathological regions. In addition, a transformer-based module incorporates patient metadata (age, sex, lesion site) into the prediction. We evaluate our approach on the HAM10000 dataset and the ISIC 2018 and 2019 challenges. The proposed method achieves state-of-the-art segmentation performance and significantly improves classification accuracy and average AUC compared to baseline models. To validate our model's reliability, we use Gradient-weighted Class Activation Mapping (Grad-CAM) to generate heatmaps. These visualizations confirm that our model's predictions are based on the lesion area, unlike models that rely on spurious background features. These results demonstrate that integrating precise lesion segmentation and clinical data with attention-based fusion leads to a more accurate and interpretable skin cancer classification model.

Language models often generate factually incorrect information unsupported by their training data, a phenomenon known as extrinsic hallucination. Existing mitigation approaches often degrade performance on open-ended generation and downstream tasks, limiting their practical utility. We propose an online reinforcement learning method using a novel binary retrieval-augmented reward (RAR) to address this tradeoff. Unlike continuous reward schemes, our approach assigns a reward of one only when the model's output is entirely factually correct, and zero otherwise. We evaluate our method on Qwen3 reasoning models across diverse tasks. For open-ended generation, binary RAR achieves a 39.3% reduction in hallucination rates, substantially outperforming both supervised training and continuous-reward RL baselines. In short-form question answering, the model learns calibrated abstention, strategically outputting "I don't know" when faced with insufficient parametric knowledge. This yields 44.4% and 21.7% fewer incorrect answers on PopQA and GPQA, respectively. Crucially, these factuality gains come without performance degradation on instruction following, math, or code, whereas continuous-reward RL, despite improving factuality, induces quality regressions.

Differentiating cardiogenic pulmonary oedema (CPE) from non-cardiogenic and structurally normal lungs in lung ultrasound (LUS) videos remains challenging due to the high visual variability of non-cardiogenic inflammatory patterns (NCIP/ARDS-like), interstitial lung disease, and healthy lungs. This heterogeneity complicates automated classification as overlapping B-lines and pleural artefacts are common. We introduce ZACH-ViT (Zero-token Adaptive Compact Hierarchical Vision Transformer), a 0.25 M-parameter Vision Transformer variant that removes both positional embeddings and the [CLS] token, making it fully permutation-invariant and suitable for unordered medical image data. To enhance generalization, we propose ShuffleStrides Data Augmentation (SSDA), which permutes probe-view sequences and frame orders while preserving anatomical validity. ZACH-ViT was evaluated on 380 LUS videos from 95 critically ill patients against nine state-of-the-art baselines. Despite the heterogeneity of the non-cardiogenic group, ZACH-ViT achieved the highest validation and test ROC-AUC (0.80 and 0.79) with balanced sensitivity (0.60) and specificity (0.91), while all competing models collapsed to trivial classification. It trains 1.35x faster than Minimal ViT (0.62M parameters) with 2.5x fewer parameters, supporting real-time clinical deployment. These results show that aligning architectural design with data structure can outperform scale in small-data medical imaging.