Large language models (LLMs) are widely adopted to generate synthetic datasets for various natural language processing (NLP) tasks, such as text classification and summarization. However, accurately measuring the diversity of these synthetic datasets-an aspect crucial for robust model performance-remains a significant challenge. In this paper, we introduce DCScore, a novel method for measuring synthetic dataset diversity from a classification perspective. Specifically, DCScore formulates diversity evaluation as a sample classification task, leveraging mutual relationships among samples. We further provide theoretical verification of the diversity-related axioms satisfied by DCScore, highlighting its role as a principled diversity evaluation method. Experimental results on synthetic datasets reveal that DCScore enjoys a stronger correlation with multiple diversity pseudo-truths of evaluated datasets, underscoring its effectiveness. Moreover, both empirical and theoretical evidence demonstrate that DCScore substantially reduces computational costs compared to existing approaches. Code is available at: https://github.com/BlueWhaleLab/DCScore.

This study examines how temperature settings and model architectures affect the generation of structured fictional data (names, birthdates) across three large language models (LLMs): llama3.1:8b, deepseek-r1:8b, and mistral:latest. By systematically testing temperature values from 0.0 to 1.0 in increments of 0.1, we conducted 330 trials yielding 889 structured entities, validated for syntactic consistency. Key findings reveal that model architecture significantly influences computational efficiency, with mistral:latest and llama3.1:8b processing data 8x faster than deepseek-r1:8b. Contrary to expectations, temperature showed no correlation with processing time, challenging assumptions about stochastic sampling costs. Output diversity remained limited, as models consistently defaulted to common name archetypes (e.g., 'John Doe' and 'Jane Smith') across all temperatures, though rare names clustered at intermediate values (0.3-0.7). These results demonstrate that architectural optimizations, rather than temperature adjustments, dominate performance in structured generation tasks. The findings emphasize prioritizing model selection over hyperparameter tuning for efficiency and suggest explicit diversity constraints are necessary to mitigate default output biases in synthetic data pipelines.

Context-aware compression techniques have gained increasing attention as model sizes continue to grow, introducing computational bottlenecks that hinder efficient deployment. A structured encoding approach was proposed to selectively eliminate redundant parameter groups while ensuring that representational fidelity was preserved across multiple layers. Contextual Compression Encoding (CCE) introduced a multi-stage encoding mechanism that dynamically restructured parameter distributions, allowing for significant reductions in memory footprint and computational complexity. Experimental evaluations demonstrated that models compressed through CCE retained linguistic expressivity and coherence, maintaining accuracy across a range of text generation and classification tasks. Layer-wise analysis revealed that middle-network layers exhibited higher compression ratios, aligning with the observation that self-attention and feed-forward transformations contained redundancies that could be reorganized without impairing functional capacity. Comparisons against conventional quantization and pruning methods confirmed that CCE provided a more balanced trade-off between efficiency and model retention, achieving reductions in energy consumption and inference latency without requiring extensive retraining. Computational efficiency improvements were particularly evident in deployment scenarios involving resource-constrained environments, where reductions in memory usage enabled more scalable implementations. Further analyses of internal network behavior showed that compressed models exhibited stable activation distributions and adapted dynamically to input variations, reinforcing the viability of structured compression strategies for optimizing large-scale architectures.

We present Label Space Reduction (LSR), a novel method for improving zero-shot classification performance of Large Language Models (LLMs). LSR iteratively refines the classification label space by systematically ranking and reducing candidate classes, enabling the model to concentrate on the most relevant options. By leveraging unlabeled data with the statistical learning capabilities of data-driven models, LSR dynamically optimizes the label space representation at test time. Our experiments across seven benchmarks demonstrate that LSR improves macro-F1 scores by an average of 7.0% (up to 14.2%) with Llama-3.1-70B and 3.3% (up to 11.1%) with Claude-3.5-Sonnet compared to standard zero-shot classification baselines. To reduce the computational overhead of LSR, which requires an additional LLM call at each iteration, we propose distilling the model into a probabilistic classifier, allowing for efficient inference.

There is growing demand for performing inference with hundreds of thousands of input tokens on trained transformer models. Inference at this extreme scale demands significant computational resources, hindering the application of transformers at long contexts on commodity (i.e not data center scale) hardware. To address the inference time costs associated with running self-attention based transformer language models on long contexts and enable their adoption on widely available hardware, we propose a tunable mechanism that reduces the cost of the forward pass by attending to only the most relevant tokens at every generation step using a top-k selection mechanism. We showcase the efficiency gains afforded by our method by performing inference on context windows up to 1M tokens using approximately 16GB of GPU RAM. Our experiments reveal that models are capable of handling the sparsity induced by the reduced number of keys and values. By attending to less than 2% of input tokens, we achieve over 95% of model performance on common benchmarks (RULER, AlpacaEval, and Open LLM Leaderboard).

Foundation models have become general-purpose assistants, exhibiting diverse capabilities across numerous domains through training on web-scale data. It remains challenging to precisely characterize even a fraction of the full spectrum of capabilities and potential risks in any new model. Existing evaluation approaches often require significant human effort, and it is taking increasing effort to design ever harder challenges for more capable models. We introduce Automated Capability Discovery (ACD), a framework that designates one foundation model as a scientist to systematically propose open-ended tasks probing the abilities of a subject model (potentially itself). By combining frontier models with ideas from the field of open-endedness, ACD automatically and systematically uncovers both surprising capabilities and failures in the subject model. We demonstrate ACD across a range of foundation models (including the GPT, Claude, and Llama series), showing that it automatically reveals thousands of capabilities that would be challenging for any single team to uncover. We further validate our method's automated scoring with extensive human surveys, observing high agreement between model-generated and human evaluations. By leveraging foundation models' ability to both create tasks and self-evaluate, ACD is a significant step toward scalable, automated evaluation of novel AI systems. All code and evaluation logs are open-sourced at https://github.com/conglu1997/ACD.

Recent methodologies utilizing synthetic datasets have aimed to address inconsistent hallucinations in large language models (LLMs); however,these approaches are primarily tailored to specific tasks, limiting their generalizability. Inspired by the strong performance of code-trained models in logic-intensive domains, we propose a novel framework that leverages event-based text to generate corresponding code and employs cyclic training to transfer the logical consistency of code to natural language effectively. Our method significantly reduces inconsistent hallucinations across three leading LLMs and two categories of natural language tasks while maintaining overall performance. This framework effectively alleviates hallucinations without necessitating adaptation to downstream tasks, demonstrating generality and providing new perspectives to tackle the challenge of inconsistent hallucinations.

Large language models' significant advances in capabilities are accompanied by significant increases in inference costs. Model routing is a simple technique for reducing inference cost, wherein one maintains a pool of candidate LLMs, and learns to route each prompt to the smallest feasible LLM. Existing works focus on learning a router for a fixed pool of LLMs. In this paper, we consider the problem of dynamic routing, where new, previously unobserved LLMs are available at test time. We propose a new approach to this problem that relies on representing each LLM as a feature vector, derived based on predictions on a set of representative prompts. Based on this, we detail two effective strategies, relying on cluster-based routing and a learned cluster map respectively. We prove that these strategies are estimates of a theoretically optimal routing rule, and provide an excess risk bound to quantify their errors. Experiments on a range of public benchmarks show the effectiveness of the proposed strategies in routing amongst more than 30 unseen LLMs.

Large language models (LLMs) are helping millions of users write texts about diverse issues, and in doing so expose users to different ideas and perspectives. This creates concerns about issue bias, where an LLM tends to present just one perspective on a given issue, which in turn may influence how users think about this issue. So far, it has not been possible to measure which issue biases LLMs actually manifest in real user interactions, making it difficult to address the risks from biased LLMs. Therefore, we create IssueBench: a set of 2.49m realistic prompts for measuring issue bias in LLM writing assistance, which we construct based on 3.9k templates (e.g. "write a blog about") and 212 political issues (e.g. "AI regulation") from real user interactions. Using IssueBench, we show that issue biases are common and persistent in state-of-the-art LLMs. We also show that biases are remarkably similar across models, and that all models align more with US Democrat than Republican voter opinion on a subset of issues. IssueBench can easily be adapted to include other issues, templates, or tasks. By enabling robust and realistic measurement, we hope that IssueBench can bring a new quality of evidence to ongoing discussions about LLM biases and how to address them.

Time is implicitly embedded in classification process: classifiers are usually built on existing data while to be applied on future data whose distributions (e.g., label and token) may change. However, existing state-of-the-art classification models merely consider the temporal variations and primarily focus on English corpora, which leaves temporal studies less explored, let alone under multilingual settings. In this study, we fill the gap by treating time as domains (e.g., 2024 vs. 2025), examining temporal effects, and developing a domain adaptation framework to generalize classifiers over time on multiple languages. Our framework proposes Mixture of Temporal Experts (MoTE) to leverage both semantic and data distributional shifts to learn and adapt temporal trends into classification models. Our analysis shows classification performance varies over time across different languages, and we experimentally demonstrate that MoTE can enhance classifier generalizability over temporal data shifts. Our study provides analytic insights and addresses the need for time-aware models that perform robustly in multilingual scenarios.