Large language models are increasingly capable of producing creative texts, yet most studies on AI-generated poetry focus on English -- a language that dominates training data. In this paper, we examine the perception of AI- and human-written Czech poetry. We ask if Czech native speakers are able to identify it and how they aesthetically judge it. Participants performed at chance level when guessing authorship (45.8\% correct on average), indicating that Czech AI-generated poems were largely indistinguishable from human-written ones. Aesthetic evaluations revealed a strong authorship bias: when participants believed a poem was AI-generated, they rated it as less favorably, even though AI poems were in fact rated equally or more favorably than human ones on average. The logistic regression model uncovered that the more the people liked a poem, the less probable was that they accurately assign the authorship. Familiarity with poetry or literary background had no effect on recognition accuracy. Our findings show that AI can convincingly produce poetry even in a morphologically complex, low-resource (with respect of the training data of AI models) Slavic language such as Czech. The results suggest that readers' beliefs about authorship and the aesthetic evaluation of the poem are interconnected.

Large Language Models (LLMs) have recently revolutionized machine learning on text-attributed graphs, but the application of LLMs to graph outlier detection, particularly in the context of fake news detection, remains significantly underexplored. One of the key challenges is the scarcity of large-scale, realistic, and well-annotated datasets that can serve as reliable benchmarks for outlier detection. To bridge this gap, we introduce TAGFN, a large-scale, real-world text-attributed graph dataset for outlier detection, specifically fake news detection. TAGFN enables rigorous evaluation of both traditional and LLM-based graph outlier detection methods. Furthermore, it facilitates the development of misinformation detection capabilities in LLMs through fine-tuning. We anticipate that TAGFN will be a valuable resource for the community, fostering progress in robust graph-based outlier detection and trustworthy AI. The dataset is publicly available at https://huggingface.co/datasets/kayzliu/TAGFN and our code is available at https://github.com/kayzliu/tagfn.

Incorporating metadata in Large Language Models (LLMs) pretraining has recently emerged as a promising approach to accelerate training. However prior work highlighted only one useful signal-URLs, leaving open the question of whether other forms of metadata could yield greater benefits. In this study, we investigate a wider range of metadata types and find other types of metadata, such as fine-grained indicators of document quality that can also accelerate pretraining when prepended. We identify a common feature among effective metadata: they encode information at a finer granularity. We further introduce metadata appending as a means of improving training efficiency, where predicting an appropriate metadata as auxiliary task can help speed up pretraining. In addition, learnable meta-tokens trained with masked loss can recover part of the speedup by inducing quality-aware latent structure. Using probing, we analyze latent representations to understand how metadata shapes learning. Together, these results yield practical guidelines for integrating metadata to improve both the efficiency and effectiveness of LLM pretraining.

Large language models (LLMs) exhibit remarkable capabilities across a wide range of tasks, yet their internal mechanisms remain largely opaque. In this paper, we introduce a simple, lightweight, and broadly applicable method with a focus on isolating neurons that encode specific skills. Building upon prior work that identified "skill neurons" via soft prompt training on classification tasks, our approach extends the analysis to complex scenarios involving multiple skills. We correlate neuron activations with auxiliary metrics -- such as external labels and the model's own confidence score -- thereby uncovering interpretable and task-specific behaviors without the need for manual token aggregation. We empirically validate our method on tasks spanning open-ended text generation and natural language inference, demonstrating its ability to detect neurons that not only drive known skills but also reveal previously unidentified shortcuts in arithmetic reasoning on BigBench.

Large language models (LLMs) achieve impressive results on many benchmarks, yet their capacity for planning and stateful reasoning remains unclear. We study these abilities directly, without code execution or other tools, using the 8-puzzle: a classic task that requires state tracking and goal-directed planning while allowing precise, step-by-step evaluation. Four models are tested under common prompting conditions (Zero-Shot, Chain-of-Thought, Algorithm-of-Thought) and with tiered corrective feedback. Feedback improves success rates for some model-prompt combinations, but many successful runs are long, computationally expensive, and indirect. We then examine the models with an external move validator that provides only valid moves. Despite this level of assistance, none of the models solve any puzzles in this setting. Qualitative analysis reveals two dominant deficits across all models: (1) brittle internal state representations, leading to frequent invalid moves, and (2) weak heuristic planning, with models entering loops or selecting actions that do not reduce the distance to the goal state. These findings indicate that, in the absence of external tools such as code interpreters, current LLMs have substantial limitations in planning and that further progress may require mechanisms for maintaining explicit state and performing structured search.

End-to-end (E2E) autonomous driving models have demonstrated strong performance in open-loop evaluations but often suffer from cascading errors and poor generalization in closed-loop settings. To address this gap, we propose Model-based Policy Adaptation (MPA), a general framework that enhances the robustness and safety of pretrained E2E driving agents during deployment. MPA first generates diverse counterfactual trajectories using a geometry-consistent simulation engine, exposing the agent to scenarios beyond the original dataset. Based on this generated data, MPA trains a diffusion-based policy adapter to refine the base policy's predictions and a multi-step Q value model to evaluate long-term outcomes. At inference time, the adapter proposes multiple trajectory candidates, and the Q value model selects the one with the highest expected utility. Experiments on the nuScenes benchmark using a photorealistic closed-loop simulator demonstrate that MPA significantly improves performance across in-domain, out-of-domain, and safety-critical scenarios. We further investigate how the scale of counterfactual data and inference-time guidance strategies affect overall effectiveness.

Latent reasoning represents a new development in Transformer language models that has shown potential in compressing reasoning lengths compared to chain-of-thought reasoning. By directly passing the information-rich previous final latent state into the next sequence, latent reasoning removes the restriction to human language tokens as the medium for reasoning. We develop adaptive-length latent reasoning models and introduce a post-SFT reinforcement-learning methodology to optimize latent reasoning length by minimizing reasoning length while maintaining accuracy. This, in turn, further reduces compute usage and raises the bar on the compressive capabilities of latent reasoning models. Experiments on the Llama 3.2 1B model and the GSM8K-Aug dataset show a $52\%$ drop in total reasoning length with no penalty to accuracy. In future work, we plan to extend to additional models and datasets, analyze relationships between training coefficients, experiment with architecture variations, and continue our knowledge distillation for latent reasoning SFT efforts. We make our code and pretrained weights available at https://github.com/apning/adaptive-latent-reasoning.

Large Language Models (LLMs) often exhibit inconsistent behavior when answering paraphrased questions, suggesting a reliance on surface-level patterns rather than true semantic understanding. To address this limitation, we introduce RoParQ, a benchmark specifically constructed to evaluate cross-paraphrase consistency in closed-book multiple-choice QA. This benchmark is derived from standard datasets by generating paraphrases via proprietary models and selectively retaining examples that elicit inconsistent confidence from a judge model. We further propose XParaCon, a novel evaluation metric that quantifies a model's robustness by measuring the standard deviation of accuracies across question variants. Additionally, we implement a reasoning-based, paraphrase-aware Supervised Fine-Tuning (SFT) strategy designed to align models toward semantic invariance. Our experiments demonstrate that this targeted alignment significantly enhances robustness. Notably, fine-tuned lightweight models achieved consistency levels comparable to much larger pre-trained models. These results highlight the efficacy of our approach in mitigating superficial memorization and fostering more robust, reliable LLMs.

The key limitation of the verification performance lies in the ability of error detection. With this intuition we designed several variants of pessimistic verification, which are simple workflows that could significantly improve the verification of open-ended math questions. In pessimistic verification we construct multiple parallel verifications for the same proof, and the proof is deemed incorrect if any one of them reports an error. This simple technique significantly improves the performance across many math verification benchmarks without incurring substantial computational resources. Its token efficiency even surpassed extended long-CoT in test-time scaling. Our case studies further indicate that the majority of false negatives in stronger models are actually caused by annotation errors in the original dataset, so our method's performance is in fact underestimated. Self-verification for mathematical problems can effectively improve the reliability and performance of language model outputs, and it also plays a critical role in enabling long-horizon mathematical tasks. We believe that research on pessimistic verification will help enhance the mathematical capabilities of language models across a wide range of tasks.

Deploying Transformer models on edge devices is limited by latency and energy budgets. While INT8 quantization effectively accelerates the primary matrix multiplications, it exposes the softmax as the dominant bottleneck. This stage incurs a costly dequantize-softmax-requantize detour, which can account for up to 65% of total attention latency and disrupts the end-to-end integer dataflow critical for edge hardware efficiency. To address this limitation, we present IntAttention, the first fully integer, plug-and-play attention pipeline without retraining. At the core of our approach lies IndexSoftmax, a hardware-friendly operator that replaces floating-point exponentials entirely within the integer domain. IntAttention integrates sparsity-aware clipping, a 32-entry lookup-table approximation, and direct integer normalization, thereby eliminating all datatype conversion overhead. We evaluate IntAttention and demonstrate consistent and substantial gains. Our method achieves up to 3.7x speedup and 61% energy reduction over FP16 baselines and 2.0x faster than conventional INT8 attention pipelines on Armv8 CPUs. These gains are achieved with high-fidelity accuracy comparable to baselines across diverse language and vision models, enabling practical and efficient Transformer inference on commodity edge devices. Code will be released in later version of this work.