Leaderboards for large reasoning models (LRMs) have turned evaluation into a competition, incentivizing developers to optimize directly on benchmark suites. A shortcut to achieving higher rankings is to incorporate evaluation benchmarks into the training data, thereby yielding inflated performance, known as benchmark contamination. Despite that numerous contamination detection approaches have been proposed, surprisingly, our studies find that evading contamination detections for LRMs is alarmingly easy. We focus on the two scenarios where contamination may occur in practice: (I) when the base model evolves into LRM via supervised fine-tuning (SFT) and reinforcement learning (RL), we find that contamination during SFT can be originally identified by contamination detection methods. Y et, even a brief Group Relative Policy Optimization (GRPO) training can markedly conceal contamination signals that most detection methods rely on. Further empirical experiments and theoretical analysis indicate that Proximal Policy Optimization (PPO) style importance sampling and clipping objectives are the root cause of this detection concealment, indicating that a broad class of RL methods may inherently exhibit similar concealment capability; (II) when SFT contamination with CoT is applied to advanced LRMs as the final stage, most contamination detection methods perform near random guesses. Without exposure to non-members, contaminated LRMs would still have more confidence when responding to those unseen samples that share similar distributions to the training set, and thus, evade existing memorization-based detection methods. Together, our findings reveal the unique vulnerability of LRMs evaluations: Model developers could easily contaminate LRMs to achieve inflated leaderboards performance while leaving minimal traces of contamination, thereby strongly undermining the fairness of evaluation and threatening the integrity of public leaderboards. This underscores the urgent need for advanced contamination detection methods and trustworthy evaluation protocols tailored to LRMs. Our code is available at https://github.com/ASTRAL-Group/ Competition among model developers has intensified as Large Language Models (LLMs) have demonstrated remarkable capabilities in various real-world tasks (Achiam et al., 2023; Wang et al., 2024). The leaderboards for performance are becoming a competitive arena for all state-of-the-art (SOT A) LLMs. However, inadvertently, benchmark samples may appear during LLMs' pre-training due to vast amounts of web-scraped training data. In addition, in the pursuit of publicity, some model developers may even deliberately incorporate benchmark data into their training sets (Sun et al., 2025), resulting in inflated benchmark performance and leaderboard rankings. We refer to this as the benchmark contamination problem in LLMs (Xu et al., 2024; Balloccu et al., 2024).

Large Language Models (LLMs) have demonstrated exceptional capabilities, yet selecting the most reliable response from multiple LLMs remains a challenge, particularly in resource-constrained settings. Existing approaches often depend on costly external verifiers, human evaluators, or self-consistency techniques that require multiple samples from a single model. While multi-LLM systems produce more diverse responses than single models and thus have greater potential, they often underperform compared to single LLM self-consistency. We propose a principled, novel and computationally efficient method to select the best response from multiple different LLMs using a calibrated log-likelihood score, implicitly leveraging the inherent knowledge and confidence of these models. Our method demonstrates improvements of approx. 4%, 3%, and 5% across both debate (multi-round LLM discussions) and non-debate (Best-of-N with multiple LLMs) settings on GSM8K, MMLU (6 subsets), and ARC datasets respectively.

In this case study, we select a set of controversial Romanian historical questions and ask multiple Large Language Models to answer them across languages and contexts, in order to assess their biases. Besides being a study mainly performed for educational purposes, the motivation also lies in the recognition that history is often presented through altered perspectives, primarily influenced by the culture and ideals of a state, even through large language models. Since they are often trained on certain data sets that may present certain ambiguities, the lack of neutrality is subsequently instilled in users. The research process was carried out in three stages, to confirm the idea that the type of response expected can influence, to a certain extent, the response itself; after providing an affirmative answer to some given question, an LLM could shift its way of thinking after being asked the same question again, but being told to respond with a numerical value of a scale. Results show that binary response stability is relatively high but far from perfect and varies by language. Models often flip stance across languages or between formats; numeric ratings frequently diverge from the initial binary choice, and the most consistent models are not always those judged most accurate or neutral. Our research brings to light the predisposition of models to such inconsistencies, within a specific contextualization of the language for the question asked.

Transformer-based large models excel in natural language processing and computer vision, but face severe computational inefficiencies due to the self-attention's quadratic complexity with input tokens. Recently, researchers have proposed a series of methods based on block selection and compression to alleviate this problem, but they either have issues with semantic incompleteness or poor training-inference efficiency. To comprehensively address these challenges, we propose ChunkLLM, a lightweight and pluggable training framework. Specifically, we introduce two components: QK Adapter (Q-Adapter and K-Adapter) and Chunk Adapter. The former is attached to each Transformer layer, serving dual purposes of feature compression and chunk attention acquisition. The latter operates at the bottommost layer of the model, functioning to detect chunk boundaries by leveraging contextual semantic information. During the training phase, the parameters of the backbone remain frozen, with only the QK Adapter and Chunk Adapter undergoing training. Notably, we design an attention distillation method for training the QK Adapter, which enhances the recall rate of key chunks. During the inference phase, chunk selection is triggered exclusively when the current token is detected as a chunk boundary, thereby accelerating model inference. Experimental evaluations are conducted on a diverse set of long-text and short-text benchmark datasets spanning multiple tasks. ChunkLLM not only attains comparable performance on short-text benchmarks but also maintains 98.64% of the performance on long-context benchmarks while preserving a 48.58% key-value cache retention rate. Particularly, ChunkLLM attains a maximum speedup of 4.48x in comparison to the vanilla Transformer in the processing of 120K long texts.

The adoption of generative AI and large language models (LLMs) in education is still emerging. In this study, we explore the development and evaluation of AI teaching assistants that provide curriculum-based guidance using a retrieval-augmented generation (RAG) pipeline applied to selected open-source small language models (SLMs). We benchmarked eight SLMs, including LLaMA 3.1, IBM Granite 3.3, and Gemma 3 (7-17B parameters), against GPT-4o. Our findings show that with proper prompting and targeted retrieval, SLMs can match LLMs in delivering accurate, pedagogically aligned responses. Importantly, SLMs offer significant sustainability benefits due to their lower computational and energy requirements, enabling real-time use on consumer-grade hardware without depending on cloud infrastructure. This makes them not only cost-effective and privacy-preserving but also environmentally responsible, positioning them as viable AI teaching assistants for educational institutions aiming to scale personalized learning in a sustainable and energy-efficient manner.

Our results resolve an open question by Gillen et al. (2018) by showing that online learning under an unknown individual fairness constraint is possible even without assuming a strong parametric form of the

Our results resolve an open question by Gillen et al. (2018) by showing that online learning under an unknown individual fairness constraint is possible even without assuming a strong parametric form of the