Concurrent to the rapid progress in the development of neural-network based models in areas like natural language processing and computer vision, the need for creating explanations for the predictions of these black-box models has risen steadily. We propose a new method to generate extractive explanations for predictions made by neural networks, that is based on masking parts of the input which the model does not consider to be indicative of the respective class. The masking is done using gradient-based optimization combined with a new regularization scheme that enforces sufficiency, comprehensiveness and compactness of the generated explanation, three properties that are known to be desirable from the related field of rationale extraction in natural language processing. In this way, we bridge the gap between model interpretability and rationale extraction, thereby proving that the latter of which can be performed without training a specialized model, only on the basis of a trained classifier. We further apply the same method to image inputs and obtain high quality explanations for image classifications, which indicates that the conditions proposed for rationale extraction in natural language processing are more broadly applicable to different input types.

This technical report details a novel approach to combining reasoning and retrieval augmented generation (RAG) within a single, lean language model architecture. While existing RAG systems typically rely on large-scale models and external APIs, our work addresses the increasing demand for performant and privacy-preserving solutions deployable in resource-constrained or secure environments. Building on recent developments in test-time scaling and small-scale reasoning models, we develop a retrieval augmented conversational agent capable of interpreting complex, domain-specific queries using a lightweight backbone model. Our system integrates a dense retriever with fine-tuned Qwen2.5-Instruct models, using synthetic query generation and reasoning traces derived from frontier models (e.g., DeepSeek-R1) over a curated corpus, in this case, the NHS A-to-Z condition pages. We explore the impact of summarisation-based document compression, synthetic data design, and reasoning-aware fine-tuning on model performance. Evaluation against both non-reasoning and general-purpose lean models demonstrates that our domain-specific fine-tuning approach yields substantial gains in answer accuracy and consistency, approaching frontier-level performance while remaining feasible for local deployment. All implementation details and code are publicly released to support reproducibility and adaptation across domains.

Large Language Models (LLMs) are highly sensitive to subtle, non-semantic variations in prompt phrasing and formatting. In this work, we present the first systematic evaluation of 5 methods for improving prompt robustness within a unified experimental framework. We benchmark these techniques on 8 models from Llama, Qwen and Gemma families across 52 tasks from Natural Instructions dataset. Our evaluation covers robustness methods from both fine-tuned and in-context learning paradigms, and tests their generalization against multiple types of distribution shifts. Finally, we extend our analysis to GPT-4.1 and DeepSeek V3 to assess frontier models' current robustness to format perturbations. Our findings offer actionable insights into the relative effectiveness of these robustness methods, enabling practitioners to make informed decisions when aiming for stable and reliable LLM performance in real-world applications. Code: https://github.com/AIRI-Institute/when-punctuation-matters.

Knowledge Distillation is crucial for optimizing face recognition models for deployment in computationally limited settings, such as edge devices. Traditional KD methods, such as Raw L2 Feature Distillation or Feature Consistency loss, often fail to capture both fine-grained instance-level details and complex relational structures, leading to suboptimal performance. We propose a unified approach that integrates two novel loss functions, Instance-Level Embedding Distillation and Relation-Based Pairwise Similarity Distillation. Instance-Level Embedding Distillation focuses on aligning individual feature embeddings by leveraging a dynamic hard mining strategy, thereby enhancing learning from challenging examples. Relation-Based Pairwise Similarity Distillation captures relational information through pairwise similarity relationships, employing a memory bank mechanism and a sample mining strategy. This unified framework ensures both effective instance-level alignment and preservation of geometric relationships between samples, leading to a more comprehensive distillation process. Our unified framework outperforms state-of-the-art distillation methods across multiple benchmark face recognition datasets, as demonstrated by extensive experimental evaluations. Interestingly, when using strong teacher networks compared to the student, our unified KD enables the student to even surpass the teacher's accuracy.

Automated feedback generation has the potential to enhance students' learning progress by providing timely and targeted feedback. Moreover, it can assist teachers in optimizing their time, allowing them to focus on more strategic and personalized aspects of teaching. To generate high-quality, information-rich formative feedback, it is essential first to extract relevant indicators, as these serve as the foundation upon which the feedback is constructed. Teachers often employ feedback criteria grids composed of various indicators that they evaluate systematically. This study examines the initial phase of extracting such indicators from students' submissions of a language learning course using the large language model Llama 3.1. Accordingly, the alignment between indicators generated by the LLM and human ratings across various feedback criteria is investigated. The findings demonstrate statistically significant strong correlations, even in cases involving unanticipated combinations of indicators and criteria. The methodology employed in this paper offers a promising foundation for extracting indicators from students' submissions using LLMs. Such indicators can potentially be utilized to auto-generate explainable and transparent formative feedback in future research.

We present Dual-Feedback Actor (DFA), a reinforcement learning algorithm that fuses both individual rewards and pairwise preferences (if available) into a single update rule. DFA uses the policy's log-probabilities directly to model the preference probability, avoiding a separate reward-modeling step. Preferences can be provided by human-annotators (at state-level or trajectory-level) or be synthesized online from Q-values stored in an off-policy replay buffer. Under a Bradley-Terry model, we prove that minimizing DFA's preference loss recovers the entropy-regularized Soft Actor-Critic (SAC) policy. Our simulation results show that DFA trained on generated preferences matches or exceeds SAC on six control environments and demonstrates a more stable training process. With only a semi-synthetic preference dataset under Bradley-Terry model, our algorithm outperforms reward-modeling reinforcement learning from human feedback (RLHF) baselines in a stochastic GridWorld and approaches the performance of an oracle with true rewards.

As autonomous agents become adept at understanding and interacting with graphical user interface (GUI) environments, a new era of automated task execution is emerging. Recent studies have demonstrated that Reinforcement Learning (RL) can effectively enhance agents' performance in dynamic interactive GUI environments. However, these methods face two key limitations: (1) they overlook the significant variation in difficulty across different GUI tasks by treating the entire training data as a uniform set, which hampers the agent's ability to adapt its learning process; and (2) most approaches collapse task-specific nuances into a single, coarse reward, leaving the agent with a uniform signal that yields inefficient policy updates. To address these limitations, we propose CRAFT-GUI, a curriculum learning framework based on Group Relative Policy Optimization (GRPO) that explicitly accounts for the varying difficulty across trajectories. To enable more fine-grained policy optimization, we design a reward function that combines simple rule-based signals with model-judged evaluation, providing richer and more nuanced feedback during training. Experimental results demonstrate that our method achieves significant improvements over previous state-of-the-art approaches, outperforming them by 5.6% on public benchmarks Android Control and 10.3% on our internal online benchmarks, respectively. These findings empirically validate the effectiveness of integrating reinforcement learning with curriculum learning in GUI interaction tasks.

With the growing incorporation of deep neural network (DNN) models into modern software systems, the prohibitive construction costs have become a significant challenge. Model reuse has been widely applied to reduce training costs, but indiscriminately reusing entire models may incur significant inference overhead. Consequently, DNN modularization has gained attention, enabling module reuse by decomposing DNN models. The emerging modularizing-while-training (MwT) paradigm, which incorporates modularization into training, outperforms modularizing-after-training approaches. However, existing MwT methods focus on small-scale CNN models at the convolutional kernel level and struggle with diverse DNNs and large-scale models, particularly Transformer-based models. To address these limitations, we propose NeMo, a scalable and generalizable MwT approach. NeMo operates at the neuron level fundamental component common to all DNNs-ensuring applicability to Transformers and various architectures. We design a contrastive learning-based modular training method with an effective composite loss function, enabling scalability to large-scale models. Comprehensive experiments on two Transformer-based models and four CNN models across two classification datasets demonstrate NeMo's superiority over state-of-the-art MwT methods. Results show average gains of 1.72% in module classification accuracy and 58.10% reduction in module size, demonstrating efficacy across both CNN and large-scale Transformer-based models. A case study on open-source projects shows NeMo's potential benefits in practical scenarios, offering a promising approach for scalable and generalizable DNN modularization.

Conformal Prediction (CP) is a popular method for uncertainty quantification that converts a pretrained model's point prediction into a prediction set, with the set size reflecting the model's confidence. Although existing CP methods are guaranteed to achieve marginal coverage, they often exhibit imbalanced coverage across classes under long-tail label distributions, tending to over cover the head classes at the expense of under covering the remaining tail classes. This under coverage is particularly concerning, as it undermines the reliability of the prediction sets for minority classes, even with coverage ensured on average. In this paper, we propose the Tail-Aware Conformal Prediction (TACP) method to mitigate the under coverage of the tail classes by utilizing the long-tail structure and narrowing the head-tail coverage gap. Theoretical analysis shows that it consistently achieves a smaller head-tail coverage gap than standard methods. To further improve coverage balance across all classes, we introduce an extension of TACP: soft TACP (sTACP) via a reweighting mechanism. The proposed framework can be combined with various non-conformity scores, and experiments on multiple long-tail benchmark datasets demonstrate the effectiveness of our methods.

In targeted adversarial attacks on vision models, the selection of the target label is a critical yet often overlooked determinant of attack success. This target label corresponds to the class that the attacker aims to force the model to predict. Now, existing strategies typically rely on randomness, model predictions, or static semantic resources, limiting interpretability, reproducibility, or flexibility. This paper then proposes a semantics-guided framework for adversarial target selection using the cross-modal knowledge transfer from pretrained language and vision-language models. We evaluate several state-of-the-art models (BERT, TinyLLAMA, and CLIP) as similarity sources to select the most and least semantically related labels with respect to the ground truth, forming best- and worst-case adversarial scenarios. Our experiments on three vision models and five attack methods reveal that these models consistently render practical adversarial targets and surpass static lexical databases, such as WordNet, particularly for distant class relationships. We also observe that static testing of target labels offers a preliminary assessment of the effectiveness of similarity sources, \textit{a priori} testing. Our results corroborate the suitability of pretrained models for constructing interpretable, standardized, and scalable adversarial benchmarks across architectures and datasets.