Large language models (LLMs) have been used in many application domains, including cyber security. The application of LLMs in the cyber security domain presents significant opportunities, such as for enhancing threat analysis and malware detection, but it can also introduce critical risks and safety concerns, including potential personal data leakage and automated generation of new malware. Building on recent findings that fine-tuning LLMs with pseudo-malicious cyber security data significantly compromises their safety, this paper presents a comprehensive validation and extension of these safety risks using a different evaluation framework. We employ the garak red teaming framework with the OWASP Top 10 for LLM Applications to assess four open-source LLMs: Mistral 7B, Llama 3 8B, Gemma 2 9B, and DeepSeek R1 8B. Our evaluation confirms and extends previous findings, showing that fine-tuning reduces safety resilience across all tested LLMs (e.g., the failure rate of Mistral 7B against prompt injection increases from 9.1% to 68.7%). We further propose and evaluate a novel safety alignment approach that carefully rewords instruction-response pairs to include explicit safety precautions and ethical considerations. This work validates previous safety concerns through independent evaluation and introduces new methods for mitigating these risks, contributing towards the development of secure, trustworthy, and ethically aligned LLMs. This approach demonstrates that it is possible to maintain or even improve model safety while preserving technical utility, offering a practical path towards developing safer fine-tuning methodologies.

Following AI scaling trends, frontier models continue to grow in size and continue to be trained on larger datasets. Training these models requires huge investments in exascale computational resources, which has in turn driven developtment of distributed deep learning methods. Data parallelism is an essential approach to speed up training, but it requires frequent global communication between workers, which can bottleneck training at the largest scales. In this work, we propose a method called Pseudo-Asynchronous Local SGD (PALSGD) to improve the efficiency of data-parallel training. PALSGD is an extension of Local SGD (Stich, 2018) and DiLoCo (Douillard et al., 2023), designed to further reduce communication frequency by introducing a pseudo-synchronization mechanism. PALSGD allows the use of longer synchronization intervals compared to standard Local SGD. Despite the reduced communication frequency, the pseudo-synchronization approach ensures that model consistency is maintained, leading to performance results comparable to those achieved with more frequent synchronization. Furthermore, we provide a theoretical analysis of PALSGD, establishing its convergence and deriving its convergence rate. This analysis offers insights into the algorithm's behavior and performance guarantees. We evaluated PALSGD on image classification and language modeling tasks. Our results show that PALSGD achieves better performance in less time compared to existing methods like Distributed Data Parallel (DDP), and DiLoCo. Notably, PALSGD trains 18.4% faster than DDP on ImageNet-1K with ResNet-50, 24.4% faster than DDP on TinyStories with GPT-Neo-125M, and 21.1% faster than DDP on TinyStories with GPT-Neo-8M.

Federated Learning (FL) enables decentralized training of machine learning models on distributed data while preserving privacy. However, in real-world FL settings, client data is often non-identically distributed and imbalanced, resulting in statistical data heterogeneity which impacts the generalization capabilities of the server's model across clients, slows convergence and reduces performance. In this paper, we address this challenge by proposing first a characterization of statistical data heterogeneity by means of 6 metrics of global and client attribute imbalance, class imbalance, and spurious correlations. Next, we create and share 7 computer vision datasets for binary and multiclass image classification tasks in Federated Learning that cover a broad range of statistical data heterogeneity and hence simulate real-world situations. Finally, we propose FEDDIVERSE, a novel client selection algorithm in FL which is designed to manage and leverage data heterogeneity across clients by promoting collaboration between clients with complementary data distributions. Experiments on the seven proposed FL datasets demonstrate FEDDIVERSE's effectiveness in enhancing the performance and robustness of a variety of FL methods while having low communication and computational overhead.

We present a self-supervised method to improve an agent's abilities in describing arbitrary objects while actively exploring a generic environment. This is a challenging problem, as current models struggle to obtain coherent image captions due to different camera viewpoints and clutter. We propose a three-phase framework to fine-tune existing captioning models that enhances caption accuracy and consistency across views via a consensus mechanism. First, an agent explores the environment, collecting noisy image-caption pairs. Then, a consistent pseudo-caption for each object instance is distilled via consensus using a large language model. Finally, these pseudo-captions are used to fine-tune an off-the-shelf captioning model, with the addition of contrastive learning. We analyse the performance of the combination of captioning models, exploration policies, pseudo-labeling methods, and fine-tuning strategies, on our manually labeled test set. Results show that a policy can be trained to mine samples with higher disagreement compared to classical baselines. Our pseudo-captioning method, in combination with all policies, has a higher semantic similarity compared to other existing methods, and fine-tuning improves caption accuracy and consistency by a significant margin. Code and test set annotations available at https://hsp-iit.github.io/embodied-captioning/

Large language models (LLMs) achieve impressive results on advanced mathematics benchmarks but sometimes fail on basic arithmetic tasks, raising the question of whether they have truly grasped fundamental arithmetic rules or are merely relying on pattern matching. To unravel this issue, we systematically probe LLMs' understanding of two-integer addition ($0$ to $2^{64}$) by testing three crucial properties: commutativity ($A+B=B+A$), representation invariance via symbolic remapping (e.g., $7 \mapsto Y$), and consistent accuracy scaling with operand length. Our evaluation of 12 leading LLMs reveals a stark disconnect: while models achieve high numeric accuracy (73.8-99.8%), they systematically fail these diagnostics. Specifically, accuracy plummets to $\le 7.5$% with symbolic inputs, commutativity is violated in up to 20% of cases, and accuracy scaling is non-monotonic. Interventions further expose this pattern-matching reliance: explicitly providing rules degrades performance by 29.49%, while prompting for explanations before answering merely maintains baseline accuracy. These findings demonstrate that current LLMs address elementary addition via pattern matching, not robust rule induction, motivating new diagnostic benchmarks and innovations in model architecture and training to cultivate genuine mathematical reasoning. Our dataset and generating code are available at https://github.com/kuri-leo/llm-arithmetic-diagnostic.

While we increasingly rely on large language models (LLMs) for various tasks, these models are known to produce inaccurate content or 'hallucinations' with potentially disastrous consequences. The recent integration of web search results into LLMs prompts the question of whether people utilize them to verify the generated content, thereby accurately detecting hallucinations. An online experiment (N=560) investigated how the provision of search results, either static (i.e., fixed search results provided by LLM) or dynamic (i.e., participant-led searches), affects participants' perceived accuracy of LLM-generated content (i.e., genuine, minor hallucination, major hallucination), self-confidence in accuracy ratings, as well as their overall evaluation of the LLM, as compared to the control condition (i.e., no search results). Results showed that participants in both static and dynamic conditions (vs. control) rated hallucinated content to be less accurate and perceived the LLM more negatively. However, those in the dynamic condition rated genuine content as more accurate and demonstrated greater overall self-confidence in their assessments than those in the static search or control conditions. We highlighted practical implications of incorporating web search functionality into LLMs in real-world contexts.

Large Language Models (LLMs) can generate creative and engaging narratives from user-specified input, but maintaining coherence and emotional depth throughout these AI-generated stories remains a challenge. In this work, we propose SCORE, a framework for Story Coherence and Retrieval Enhancement, designed to detect and resolve narrative inconsistencies. By tracking key item statuses and generating episode summaries, SCORE uses a Retrieval-Augmented Generation (RAG) approach to identify related episodes and enhance the overall story structure. Experimental results from testing multiple LLM-generated stories demonstrate that SCORE significantly improves the consistency and stability of narrative coherence compared to baseline GPT models, providing a more robust method for evaluating and refining AI-generated narratives.

Where do learning signals come from when there is no ground truth in post-training? We propose turning exploration into supervision through Compute as Teacher (CaT), which converts the model's own exploration at inference-time into reference-free supervision by synthesizing a single reference from a group of parallel rollouts and then optimizing toward it. Concretely, the current policy produces a group of rollouts; a frozen anchor (the initial policy) reconciles omissions and contradictions to estimate a reference, turning extra inference-time compute into a teacher signal. We turn this into rewards in two regimes: (i) verifiable tasks use programmatic equivalence on final answers; (ii) non-verifiable tasks use self-proposed rubrics-binary, auditable criteria scored by an independent LLM judge, with reward given by the fraction satisfied. Unlike selection methods (best-of-N, majority, perplexity, or judge scores), synthesis may disagree with the majority and be correct even when all rollouts are wrong; performance scales with the number of rollouts. As a test-time procedure, CaT improves Gemma 3 4B, Qwen 3 4B, and Llama 3.1 8B (up to +27% on MATH-500; +12% on HealthBench). With reinforcement learning (CaT-RL), we obtain further gains (up to +33% and +30%), with the trained policy surpassing the initial teacher signal.

To address these critical shortcomings, we introduce NIRV ANA, a novel pruning method explicitly designed to balance immediate zero-shot accuracy preservation with robust fine-tuning capability. Transformer-based (V aswani et al., 2017) large language models (LLMs) have revolutionized natural To alleviate this critical bottleneck, model compression techniques--particularly pruning (LeCun et al., 1989)--emerge as an essential strategy, aiming to create lighter, more accessible models These two can also be applied for semi-structured pruning. This oversight often results in suboptimal pruning choices, impairing model performance. To address these critical gaps, we introduce NIRV ANA (NTK-InfoRmed adaptiVe neuron & AttentioN heAd pruning), a novel structured pruning method that tightly integrates pruning decisions with model fine-tuning dynamics through the lens of the Neural Tangent Kernel (NTK) (Jacot et al., 2018). An adaptive sparsity allocation strategy that dynamically adjusts pruning ratios across layers and modules, explicitly addressing overlooked disparities in existing pruning methodologies. Recent unstructured pruning methods, such as SparseGPT (Frantar and Alistarh, 2023) and Wanda (Sun et al., 2023), prune individual weights Semi-structured methods address this by imposing fixed patterns (e.g., 2:4 sparsity (Fang et al., 2024; Zheng et al., 2024)), yet still struggle to support efficient training and require specialized hardware. ShortGPT (Men et al., 2024) introduce global or layer-wise pruning strategies, yet do not explicitly SliceGPT (Ashkboos et al., 2024) applies PCA-based transformations per block, but remains highly sensitive to calibration data, reflecting a broader Table 4. Since most of the current LLMs are based on SwiGLU Shazeer (2020) structure, we focus Neural Tangent Kernel (NTK) (Jacot et al., 2018) provides a kernel-based framework for analyzing See the details of the derivation in Section A.6 3.2 P Consequently, popular practices include fixing the weights (i.e., setting In Llama3's implementation, which employs Grouped Query Attention (GQA), multiple query heads share Without loss of generality, our analysis can be extended to the vector-output case.

Data-driven modeling of nonlinear dynamical systems is often hampered by measurement noise. We propose a denoising framework, called Runge-Kutta and Total Variation Based Implicit Neural Representation (RKTV-INR), that represents the state trajectory with an implicit neural representation (INR) fitted directly to noisy observations. Runge-Kutta integration and total variation are imposed as constraints to ensure that the reconstructed state is a trajectory of a dynamical system that remains close to the original data. The trained INR yields a clean, continuous trajectory and provides accurate first-order derivatives via automatic differentiation. These denoised states and derivatives are then supplied to Sparse Identification of Nonlinear Dynamics (SINDy) to recover the governing equations. Experiments demonstrate effective noise suppression, precise derivative estimation, and reliable system identification.