Zero-shot quantization aims to learn a quantized model from a pre-trained full-precision model with no access to original real training data. The common idea in zero-shot quantization approaches is to generate synthetic data for quantizing the full-precision model. While it is well-known that deep neural networks with low sharpness have better generalization ability, none of the previous zero-shot quantization works considers the sharpness of the quantized model as a criterion for generating training data. This paper introduces a novel methodology that takes into account quantized model sharpness in synthetic data generation to enhance generalization. Specifically, we first demonstrate that sharpness minimization can be attained by maximizing gradient matching between the reconstruction loss gradients computed on synthetic and real validation data, under certain assumptions. We then circumvent the problem of the gradient matching without real validation set by approximating it with the gradient matching between each generated sample and its neighbors. Experimental evaluations on CIFAR-100 and ImageNet datasets demonstrate the superiority of the proposed method over the state-of-the-art techniques in low-bit quantization settings.

The effectiveness of Large Language Models (LLMs) depends heavily on the availability of high-quality post-training data, particularly instruction-tuning and preference-based examples. Existing open-source datasets, however, often lack multilingual coverage, cultural grounding, and suffer from task diversity gaps that are especially pronounced for Indian languages. We introduce a human-in-the-loop pipeline that combines translations with synthetic expansion to produce reliable and diverse Indic post-training data. Using this pipeline, we curate two datasets: Pragyaan-IT (22.5K) and Pragyaan-Align (100K) across 10 Indian languages covering 13 broad and 56 sub-categories, leveraging 57 diverse datasets. Our dataset protocol incorporates several often-overlooked dimensions and emphasize task diversity, multi-turn dialogue, instruction fidelity, safety alignment, and preservation of cultural nuance, providing a foundation for more inclusive and effective multilingual LLMs.

Retrieval-Augmented Generation (RAG) is a promising approach to mitigate hallucinations in Large Language Models (LLMs) for legal applications, but its reliability is critically dependent on the accuracy of the retrieval step. This is particularly challenging in the legal domain, where large databases of structurally similar documents often cause retrieval systems to fail. In this paper, we address this challenge by first identifying and quantifying a critical failure mode we term Document-Level Retrieval Mismatch (DRM), where the retriever selects information from entirely incorrect source documents. To mitigate DRM, we investigate a simple and computationally efficient technique which we refer to as Summary-Augmented Chunking (SAC). This method enhances each text chunk with a document-level synthetic summary, thereby injecting crucial global context that would otherwise be lost during a standard chunking process. Our experiments on a diverse set of legal information retrieval tasks show that SAC greatly reduces DRM and, consequently, also improves text-level retrieval precision and recall. Interestingly, we find that a generic summarization strategy outperforms an approach that incorporates legal expert domain knowledge to target specific legal elements. Our work provides evidence that this practical, scalable, and easily integrable technique enhances the reliability of RAG systems when applied to large-scale legal document datasets.

Many users interact with AI tools like ChatGPT using a mental model that treats the system as human-like--what we call Model H. According to goal-setting theory, increased specificity in goals should reduce performance variance. If Model H holds, then prompting a chatbot with more detailed instructions should lead to more consistent evaluation behaviour. This paper tests that assumption through a controlled experiment in which ChatGPT evaluated 29 student submissions using four prompts with increasing specificity. We measured consistency using intra-rater reliability (Cohen's Kappa) across repeated runs. Contrary to expectations, performance did not improve consistently with increased prompt specificity, and performance variance remained largely unchanged. These findings challenge the assumption that LLMs behave like human evaluators and highlight the need for greater robustness and improved input integration in future model development.

Finetuning vision foundation models often improves in-domain accuracy but comes at the cost of robustness under distribution shift. We revisit Mixout, a stochastic regularizer that intermittently replaces finetuned weights with their pre-trained reference, through the lens of a single-run, weight-sharing implicit ensemble. This perspective reveals three key levers that govern robustness: the masking anchor, resampling frequency, and mask sparsity. Guided by this analysis, we introduce GMixout, which (i) replaces the fixed anchor with an exponential moving-average snapshot that adapts during training, and (ii) regulates masking period via an explicit resampling-frequency hyperparameter. Experiments on benchmarks covering covariate shift, corruption, and class imbalance, ImageNet / ImageNet-L T, DomainNet, iWildCam, and CIFAR100-C, GMixout consistently improves in-domain accuracy beyond zero-shot performance while surpassing both Model Soups and strong parameter-efficient finetuning baselines under distribution shift. A core assumption in standard machine/deep learning is that training and test examples are independently and identically distributed (i.i.d.) (V apnik, 1991). However, in practice, real-world datasets often violate this assumption due to shifts in input marginal distributions (Koh et al., 2021; Gul-rajani & Lopez-Paz, 2021), spurious correlations between inputs and labels (Geirhos et al., 2020), and imbalanced class distributions (Y ang et al., 2023; Liu et al., 2019). The advent of foundation models opens new opportunities for addressing distribution shifts (Bommasani, 2021; Radford et al., 2021; Oquab et al., 2023). Notably, they often perform well even in zero-shot settings (Radford et al., 2021), underscoring how large-scale, heterogeneous pretraining can alleviate many challenges.

There are very few SotA deception systems based on Large Language Models. The existing ones are limited only to simulating one type of service, mainly SSH shells. These systems - but also the deception technologies not based on LLMs - lack an extensive evaluation that includes human attackers. Generative AI has recently become a valuable asset for cybersecurity researchers and practitioners, and the field of cyber-deception is no exception. Researchers have demonstrated how LLMs can be leveraged to create realistic-looking honeytokens, fake users, and even simulated systems that can be used as honeypots. This paper presents an AI-based deception framework called VelLMes, which can simulate multiple protocols and services such as SSH Linux shell, MySQL, POP3, and HTTP. All of these can be deployed and used as honeypots, thus VelLMes offers a variety of choices for deception design based on the users' needs. VelLMes is designed to be attacked by humans, so interactivity and realism are key for its performance. We evaluate the generative capabilities and the deception capabilities. Generative capabilities were evaluated using unit tests for LLMs. The results of the unit tests show that, with careful prompting, LLMs can produce realistic-looking responses, with some LLMs having a 100% passing rate. In the case of the SSH Linux shell, we evaluated deception capabilities with 89 human attackers. The results showed that about 30% of the attackers thought that they were interacting with a real system when they were assigned an LLM-based honeypot. Lastly, we deployed 10 instances of the SSH Linux shell honeypot on the Internet to capture real-life attacks. Analysis of these attacks showed us that LLM honeypots simulating Linux shells can perform well against unstructured and unexpected attacks on the Internet, responding correctly to most of the issued commands.

Large language models (LLMs) are increasingly deployed in user-facing applications, raising concerns about their potential to reflect and amplify social biases. We investigate social identity framing in Chinese LLMs using Mandarin-specific prompts across ten representative Chinese LLMs, evaluating responses to ingroup ("We") and outgroup ("They") framings, and extending the setting to 240 social groups salient in the Chinese context. To complement controlled experiments, we further analyze Chinese-language conversations from a corpus of real interactions between users and chatbots. Across models, we observe systematic ingroup-positive and outgroup-negative tendencies, which are not confined to synthetic prompts but also appear in naturalistic dialogue, indicating that bias dynamics might strengthen in real interactions. Our study provides a language-aware evaluation framework for Chinese LLMs, demonstrating that social identity biases documented in English generalize cross-linguistically and intensify in user-facing contexts.

Math word problems (MWPs) are critical K-12 educational tools, and customizing them to students' interests and ability levels can increase learning outcomes. However, teachers struggle to find time to customize MWPs for each student given large class sizes and increasing burnout. We propose that LLMs can support math education by generating MWPs customized to student interests and math education standards. To this end, we use a joint human expert-LLM judge approach to evaluate over 11,000 MWPs generated by open and closed LLMs and develop the first teacher-annotated dataset for standards-aligned educational MWP generation. We show the value of our data by using it to train a 12B open model that matches the performance of larger and more capable open models. We also use our teacher-annotated data to train a text classifier that enables a 30B open LLM to outperform existing closed baselines without any training. Next, we show our models' MWPs are more similar to human-written MWPs than those from existing models. We conclude by conducting the first study of customized LLM-generated MWPs with grade school students, finding they perform similarly on our models' MWPs relative to human-written MWPs but consistently prefer our customized MWPs.

The self-attention mechanism, while foundational to modern Transformer architectures, suffers from a critical inefficiency: it frequently allocates substantial attention to redundant or noisy context. Differential Attention addressed this by using subtractive attention maps for signal and noise, but its required balanced head allocation imposes rigid constraints on representational flexibility and scalability. To overcome this, we propose Grouped Differential Attention (GDA), a novel approach that introduces unbalanced head allocation between signal-preserving and noise-control groups. GDA significantly enhances signal focus by strategically assigning more heads to signal extraction and fewer to noise-control, stabilizing the latter through controlled repetition (akin to GQA). This design achieves stronger signal fidelity with minimal computational overhead. We further extend this principle to group-differentiated growth, a scalable strategy that selectively replicates only the signal-focused heads, thereby ensuring efficient capacity expansion. Through large-scale pretraining and continual training experiments, we demonstrate that moderate imbalance ratios in GDA yield substantial improvements in generalization and stability compared to symmetric baselines. Our results collectively establish that ratio-aware head allocation and selective expansion offer an effective and practical path toward designing scalable, computation-efficient Transformer architectures.

Modern astronomical surveys deliver immense volumes of transient detections, yet distinguishing real astrophysical signals (for example, explosive events) from bogus imaging artefacts remains a challenge. Convolutional neural networks are effectively used for real versus bogus classification; however, their reliance on opaque latent representations hinders interpretability. Here we show that large language models (LLMs) can approach the performance level of a convolutional neural network on three optical transient survey datasets (Pan-STARRS, MeerLICHT and ATLAS) while simultaneously producing direct, human-readable descriptions for every candidate. Using only 15 examples and concise instructions, Google's LLM, Gemini, achieves a 93% average accuracy across datasets that span a range of resolution and pixel scales. We also show that a second LLM can assess the coherence of the output of the first model, enabling iterative refinement by identifying problematic cases. This framework allows users to define the desired classification behaviour through natural language and examples, bypassing traditional training pipelines. Furthermore, by generating textual descriptions of observed features, LLMs enable users to query classifications as if navigating an annotated catalogue, rather than deciphering abstract latent spaces. As next-generation telescopes and surveys further increase the amount of data available, LLM-based classification could help bridge the gap between automated detection and transparent, human-level understanding.