The performance of large language models (LLMs) across diverse downstream applications is fundamentally governed by the quality and composition of their pretraining corpora.

The increasing reliance on generative AI models is rapidly increasing the volume of synthetic data, with some projections suggesting that most available new data for training could be machine-generated by 2030 Gartner, Inc. (2022). This shift to a mainly synthetic content presents a critical challenge: repeated training in synthetic data leads to a phenomenon known as model collapse, where model performance degrades over generations of training, eventually rendering the models ineffective. While the causes of model collapse are increasingly understood, effective mitigation strategies remain scarce. We address this challenge by leveraging a key insight: auto-regressive models tend to generate text sequences to which they assign high confidence (i.e., high log-likelihood). Based on this observation, we introduce the Truncated-Cross-Entropy (TCE) loss function. Our experiments demonstrate that models trained with TCE not only learn effectively but also exhibit significantly increased resilience, tolerating over 2.3 more synthetic data before the onset of collapse. In addition, we provide an open-source benchmark for collapse dynamics in mixed-data settings. Our results demonstrate that confidence-aware training objectives can substantially delay collapse onset, offering a practical and generalizable tool for model robustness under synthetic-data exposure. Generative models have become the foundation for modern AI applications in several modalities, including text, image, code, and audio. Large Language Models (LLMs) such as ChatGPT (OpenAI et al., 2024), LLaMA (Grattafiori et al., 2024) and Gemma (Team et al., 2025), as well as image generators DALL-E (Ramesh et al., 2021) and Imagen (Saharia et al., 2022), all rely on large datasets scraped from the Web. As these models are continuously updated to reflect recent knowledge and linguistic patterns, the need for ever larger and frequently refreshed training corpora has grown substantially. However, this demand is colliding with a shift in the data landscape: synthetic content is increasingly populating the Internet, contaminating the very datasets used for model training. This shift raises fundamental concerns.

When fine-tuning language models on new tasks, catastrophic forgetting -- performance degradation on previously-learned tasks -- is a ubiquitous problem. While Parameter-Efficient Fine-Tuning (PEFT) methods like LoRA address this through low-rank adapters, sparse adaptation offers an alternative that doesn't impose rank constraints. We introduce Random Initialization of Gated Sparse Adapters (RIGSA), which starts from randomly-initialized full-rank adapters, gates them with a ReZero analog, and sparsifies them with iterative magnitude pruning. We evaluate RIGSA on SmolLM2-1.7B-Instruct using a novel vision-in-text task (Textual MNIST) and measure forgetting on PIQA, HellaSwag, and GSM8k. SmolLM2-1.7B-Instruct initially performs around chance level on Textual MNIST, and is capable of learning the task through RIGSA, 4-bit QLoRA and random masking. In spite of having more trainable parameters than QLoRA, the RIGSA configurations that we studied displayed less forgetting than QLoRA, particularly on GSM8k, though it performs comparably to random masking.

Auto-regressive speech-text models are typically pre-trained on a large number of interleaved sequences of text tokens and raw speech encoded as speech tokens using vector quantization. These models have demonstrated state-of-the-art performance in speech-to-speech understanding and generation benchmarks, together with promising scaling laws, primarily enabled by the representational alignment between text and speech. Nevertheless, they suffer from shortcomings, partly owing to the disproportionately longer sequences of speech tokens in contrast to textual tokens. This results in a large compute imbalance between modalities during pre-training as well as during inference, and a potential hindrance to effectively aligning speech and text, ultimately translating to several orders of magnitude slower scaling laws. We introduce the Latent Speech-Text Transformer (LST), which makes pre-training speech-text models more data-efficient by dynamically and inexpensively aggregating speech tokens into latent speech patches. These patches serve as higher-level units that can either align with corresponding textual units to aid capability transfer or even encapsulate common speech sequences like silences to be more compute-efficient. We show that LST outperforms vanilla approaches on speech-to-speech as well as text-to-text benchmarks in both data- and compute-controlled settings, the former indicating more effective representational alignment and the latter indicating steeper scaling laws for speech-text models. On HellaSwag story completion, LST achieves 6.5% absolute gain in speech accuracy under compute-controlled training and 5.3% under data-controlled training, while also improving text performance. We will release our models, code, and the evaluation data to facilitate further research.

Empirical scaling laws have driven the evolution of large language models (LLMs), yet their coefficients shift whenever the model architecture or data pipeline changes. Mixture-of-Experts (MoE) models, now standard in state-of-the-art systems, introduce a new sparsity dimension that current dense-model frontiers overlook. We investigate how MoE sparsity influences two distinct capability regimes: memorization skills and reasoning skills. By training MoE families that vary total parameters, active parameters, and top-k routing under fixed compute budgets, we disentangle pre-training loss from downstream accuracy. Our results reveal two principles. First, Active FLOPs: models with identical training loss but greater active compute achieve higher reasoning accuracy. Second, T otal tokens per parameter (TPP): memorization tasks improve with more parameters, while reasoning tasks benefit from optimal TPP, indicating that reasoning is data-hungry. Neither reinforcement learning post-training (GRPO) nor increased test-time compute alters these trends. We therefore argue that optimal MoE sparsity must be determined jointly by active FLOPs and TPP, revising the classical picture of compute-optimal scaling. The recent evolution of large language models (LLMs) has been driven by empirical scaling laws (Hestness et al., 2017) that link training loss to model size, dataset size, and compute budget. Kaplan et al. (2020) showed that these laws hold across seven orders of magnitude, establishing them as a reliable extrapolation tool for dense Transformers. Subsequent work by Hoffmann et al. (2022) demonstrated that scaling curves can be inverted to choose the compute-optimal combination of parameters and tokens for a fixed budget. Together, these results have made scaling analysis a cornerstone of model planning at both academic and industrial labs.

Fine-tuning massive pre-trained language models across many tasks demands adapters that are both parameter-efficient and expressive. We introduce \textbf{Kron-LoRA}, a hybrid adapter that combines Kronecker-structured factorization with low-rank LoRA compression-an integration that, to our knowledge, has not been explored in parameter-efficient fine-tuning or in matrix approximation literature. Kron-LoRA achieves up to 4$\times$ fewer parameters than standard LoRA while retaining similar expressivity. Experiments on DistilBERT, Mistral-7B, LLaMA-2-7B, and LLaMA-3-8B across eight benchmarks show that Kron-LoRA matches or exceeds LoRA baselines with modest memory savings and only a 5-8\% speed overhead. In sequential fine-tuning, it also delivers competitive cross-task transfer despite using only one-quarter of the adapter parameters. Kron-LoRA thus offers a scalable, sustainable solution for multi-task adaptation of large language models.

Large language models (LLMs) have achieved remarkable success across various domains, driving significant technological advancements and innovations. Despite the rapid growth in model scale and capability, systematic, data-driven research on how structural configurations affect performance remains scarce. To address this gap, we present a large-scale dataset encompassing diverse open-source LLM structures and their performance across multiple benchmarks. Leveraging this dataset, we conduct a systematic, data mining-driven analysis to validate and quantify the relationship between structural configurations and performance. Our study begins with a review of the historical development of LLMs and an exploration of potential future trends. We then analyze how various structural choices impact performance across benchmarks and further corroborate our findings using mechanistic interpretability techniques. By providing data-driven insights into LLM optimization, our work aims to guide the targeted development and application of future models. We will release our dataset at https://huggingface.co/datasets/DX0369/LLM-Structure-Performance-Dataset

The rapid advancement of large language models (LLMs) and the development of increasingly large and diverse evaluation benchmarks have introduced substantial computational challenges for model assessment. In this paper, we present EffiEval, a training-free approach for efficient benchmarking that effectively addresses data redundancy while maintaining high evaluation reliability. Our method is specifically designed to meet three key criteria for high-quality evaluation: representativeness, by ensuring comprehensive coverage of model capabilities; fairness, by remaining independent of model performance during sample selection to avoid bias; and generalizability, by enabling flexible transfer across datasets and model families without reliance on large-scale evaluation data. Unlike traditional methods that rely on absolute performance or require extensive evaluation data, our approach adaptively selects high-quality representative subsets based on the Model Utility Index (MUI). Extensive experiments on multiple public benchmarks and diverse LLMs demonstrate that EffiEval achieves strong ranking consistency with full-dataset evaluation using only a small fraction of the original data. Furthermore, our method is flexible and scalable in size, allowing users to balance evaluation efficiency and representativeness according to specific needs. Overall, EffiEval provides a practical and generalizable solution for reliable, fair, and efficient evaluation in the era of LLMs.

Multilingual large language models (LLMs) exhibit strong cross-linguistic generalization, yet medium to low resource languages underperform on common benchmarks such as ARC-Challenge, MMLU, and HellaSwag. We analyze activation patterns in Gemma-2-2B across all 26 residual layers and 10 languages: Chinese (zh), Russian (ru), Spanish (es), Italian (it), medium to low resource languages including Indonesian (id), Catalan (ca), Marathi (mr), Malayalam (ml), and Hindi (hi), with English (en) as the reference. Using Sparse Autoencoders (SAEs), we reveal systematic disparities in activation patterns. Medium to low resource languages receive up to 26.27 percent lower activations in early layers, with a persistent gap of 19.89 percent in deeper layers. To address this, we apply activation-aware fine-tuning via Low-Rank Adaptation (LoRA), leading to substantial activation gains, such as 87.69 percent for Malayalam and 86.32 percent for Hindi, while maintaining English retention at approximately 91 percent. After fine-tuning, benchmark results show modest but consistent improvements, highlighting activation alignment as a key factor in enhancing multilingual LLM performance.

Pre-training state-of-the-art large language models (LLMs) requires vast amounts of clean and diverse text data. While the open development of large high-quality English pre-training datasets has seen substantial recent progress, training performant multilingual LLMs remains a challenge, in large part due to the inherent difficulty of tailoring filtering and deduplication pipelines to a large number of languages. In this work, we introduce a new pre-training dataset curation pipeline based on FineWeb that can be automatically adapted to support any language. We extensively ablate our pipeline design choices on a set of nine diverse languages, guided by a set of meaningful and informative evaluation tasks that were chosen through a novel selection process based on measurable criteria. Ultimately, we show that our pipeline can be used to create non-English corpora that produce more performant models than prior datasets. We additionally introduce a straightforward and principled approach to rebalance datasets that takes into consideration both duplication count and quality, providing an additional performance uplift. Finally, we scale our pipeline to over 1000 languages using almost 100 Common Crawl snapshots to produce FineWeb2, a new 20 terabyte (5 billion document) multilingual dataset which we release along with our pipeline, training, and evaluation codebases.