In this paper, we investigate how individuals evaluate human and large langue models generated responses to popular questions when the source of the content is either concealed or disclosed. Through a controlled field experiment, participants were presented with a set of questions, each accompanied by a response generated by either a human or an AI. In a randomized design, half of the participants were informed of the response's origin while the other half remained unaware. Our findings indicate that, overall, participants tend to prefer AI-generated responses. However, when the AI origin is revealed, this preference diminishes significantly, suggesting that evaluative judgments are influenced by the disclosure of the response's provenance rather than solely by its quality. These results underscore a bias against AI-generated content, highlighting the societal challenge of improving the perception of AI work in contexts where quality assessments should be paramount.

This paper introduces LogitLens4LLMs, a toolkit that extends the Logit Lens technique to modern large language models. While Logit Lens has been a crucial method for understanding internal representations of language models, it was previously limited to earlier model architectures. Our work overcomes the limitations of existing implementations, enabling the technique to be applied to state-of-the-art architectures (such as Qwen-2.5 and Llama-3.1) while automating key analytical workflows. By developing component-specific hooks to capture both attention mechanisms and MLP outputs, our implementation achieves full compatibility with the HuggingFace transformer library while maintaining low inference overhead. The toolkit provides both interactive exploration and batch processing capabilities, supporting large-scale layer-wise analyses. Through open-sourcing our implementation, we aim to facilitate deeper investigations into the internal mechanisms of large-scale language models. The toolkit is openly available at https://github.com/zhenyu-02/LogitLens4LLMs.

Advances in Large Language Models revolutionized medical education by enabling scalable and efficient learning solutions. This paper presents a pipeline employing Retrieval-Augmented Generation (RAG) system to prepare comments generation for Poland's State Specialization Examination (PES) based on verified resources. The system integrates these generated comments and source documents with a spaced repetition learning algorithm to enhance knowledge retention while minimizing cognitive overload. By employing a refined retrieval system, query rephraser, and an advanced reranker, our modified RAG solution promotes accuracy more than efficiency. Rigorous evaluation by medical annotators demonstrates improvements in key metrics such as document relevance, credibility, and logical coherence of generated content, proven by a series of experiments presented in the paper. This study highlights the potential of RAG systems to provide scalable, high-quality, and individualized educational resources, addressing non-English speaking users.

Key-value (KV) caching has emerged as a crucial optimization technique for accelerating inference in large language models (LLMs). By allowing the attention operation to scale linearly rather than quadratically with the total sequence length, KV caching significantly enhances generation throughput. However, due to large context lengths in the modern LLMs, the memory footprint of the KV is a huge bottleneck for model deployment directly impacting the model's batch size, hindering its ability to deliver high-throughput. Existing research addresses this challenge using several techniques, such as discarding low-attention tokens, quantization, and matrix approximation which typically lead to a negative impact on the model accuracy. In this paper, We propose KVCrush technology which can be combined with many KV compression technologies to improve the model accuracy at a much smaller memory. KVCrush provides an alternate representation scheme for key-value states, along with a low-overhead token pruning algorithm that accounts for the token distribution in the KV cache, which in turn allows for a a smaller footprint while maintaining the accuracy of the model. Based on our results, KVCrush reduces LongBench KV Cache size by 4x with less than 1% accuracy drop and achieves state-of-the-art average accuracy with minimal overhead, incurring less than 0.5% total inference latency. KVCrush not only outperforms the accuracy of state-of-the-art importance-based token retention schemes but is also compatible with typical practical LLM deployments using KV cache paging schemes such as vLLM and mixed precision quantization.

Modern large language models (LLMs) exhibit critical vulnerabilities to poison pill attacks: localized data poisoning that alters specific factual knowledge while preserving overall model utility. We systematically demonstrate these attacks exploit inherent architectural properties of LLMs, achieving 54.6% increased retrieval inaccuracy on long-tail knowledge versus dominant topics and up to 25.5% increase retrieval inaccuracy on compressed models versus original architectures. Through controlled mutations (e.g., temporal/spatial/entity alterations) and, our method induces localized memorization deterioration with negligible impact on models' performance on regular standard benchmarks (e.g., <2% performance drop on MMLU/GPQA), leading to potential detection evasion. Our findings suggest: (1) Disproportionate vulnerability in long-tail knowledge may result from reduced parameter redundancy; (2) Model compression may increase attack surfaces, with pruned/distilled models requiring 30% fewer poison samples for equivalent damage; (3) Associative memory enables both spread of collateral damage to related concepts and amplification of damage from simultaneous attack, particularly for dominant topics. These findings raise concerns over current scaling paradigms since attack costs are lowering while defense complexity is rising. Our work establishes poison pills as both a security threat and diagnostic tool, revealing critical security-efficiency trade-offs in language model compression that challenges prevailing safety assumptions.

With the rapid progress of Large Language Models (LLMs), it becomes increasingly important to understand their abilities and limitations. In two experiments, we investigate the causal and compositional reasoning abilities of LLMs and humans in the domain o f object affordances, an area traditionally linked to embodied cognition. The tasks - designed from scratch to avoid data contamination - require decision - makers to select unconventional objects to replace a typical tool for a particular purpose, such as u sing a table tennis racket to dig a hole. In Experiment 1, we evaluated GPT - 3.5 and GPT - 4o, finding that GPT - 4o, when given chain - of - thought prompting, performed on par with human participants, while GPT - 3.5 lagged significantly. In Experiment 2, we introd uced two new conditions, Distractor (more object choices, increasing difficulty) and Image (object options presented visually), and evaluated Claude 3 Sonnet and Claude 3.5 Sonnet in addition to the GPT models. The Distractor condition significantly impair ed performance across humans and models, although GPT - 4o and Claude 3.5 still performed well above chance. Surprisingly, the Image condition had little impact on humans or GPT - 4o, but significantly lowered Claude 3.5's accuracy. Qualitative analysis showed that GPT - 4o and Claude 3.5 have a stronger ability than their predecessors to identify and flexibly apply causal ly relevant object properties . The improvement from GPT - 3.5 and Claude 3 to GPT - 4o and Claude 3.5 suggests that models are increasingly capable of causal and compositional reasoning in some domains, although further mechanistic research is necessary to understand how LLMs reason.

The explosion of high-performing conversational language models (LMs) has spurred a shift from classic natural language processing (NLP) benchmarks to expensive, time-consuming and noisy human evaluations - yet the relationship between these two evaluation strategies remains hazy. In this paper, we conduct a large-scale study of four Chat Llama 2 models, comparing their performance on 160 standard NLP benchmarks (e.g., MMLU, ARC, BIG-Bench Hard) against extensive human preferences on more than 11k single-turn and 2k multi-turn dialogues from over 2k human annotators. Our findings are striking: most NLP benchmarks strongly correlate with human evaluations, suggesting that cheaper, automated metrics can serve as surprisingly reliable predictors of human preferences. Three human evaluations, such as adversarial dishonesty and safety, are anticorrelated with NLP benchmarks, while two are uncorrelated. Moreover, through overparameterized linear regressions, we show that NLP scores can accurately predict human evaluations across different model scales, offering a path to reduce costly human annotation without sacrificing rigor. Overall, our results affirm the continued value of classic benchmarks and illuminate how to harness them to anticipate real-world user satisfaction - pointing to how NLP benchmarks can be leveraged to meet evaluation needs of our new era of conversational AI.

Data contamination has received increasing attention in the era of large language models (LLMs) due to their reliance on vast Internet-derived training corpora. To mitigate the risk of potential data contamination, LLM benchmarking has undergone a transformation from static to dynamic benchmarking. In this work, we conduct an in-depth analysis of existing static to dynamic benchmarking methods aimed at reducing data contamination risks. We first examine methods that enhance static benchmarks and identify their inherent limitations. We then highlight a critical gap-the lack of standardized criteria for evaluating dynamic benchmarks. Based on this observation, we propose a series of optimal design principles for dynamic benchmarking and analyze the limitations of existing dynamic benchmarks. This survey provides a concise yet comprehensive overview of recent advancements in data contamination research, offering valuable insights and a clear guide for future research efforts. We maintain a GitHub repository to continuously collect both static and dynamic benchmarking methods for LLMs. The repository can be found at this link.

State-of-the-art data stream mining in supervised classification has traditionally relied on ensembles of incremental decision trees. However, the emergence of large tabular models, i.e., transformers designed for structured numerical data, marks a significant paradigm shift. These models move beyond traditional weight updates, instead employing in-context learning through prompt tuning. By using on-the-fly sketches to summarize unbounded streaming data, one can feed this information into a pre-trained model for efficient processing. This work bridges advancements from both areas, highlighting how transformers' implicit meta-learning abilities, pre-training on drifting natural data, and reliance on context optimization directly address the core challenges of adaptive learning in dynamic environments. Exploring real-time model adaptation, this research demonstrates that TabPFN, coupled with a simple sliding memory strategy, consistently outperforms ensembles of Hoeffding trees across all non-stationary benchmarks. Several promising research directions are outlined in the paper. The authors urge the community to explore these ideas, offering valuable opportunities to advance in-context stream learning.

During crises, social media serves as a crucial coordination tool, but the vast influx of posts--from "actionable" requests and offers to generic content like emotional support, behavioural guidance, or outdated information--complicates effective classification. Although generative LLMs (Large Language Models) can address this issue with few-shot classification, their high computational demands limit real-time crisis response. While fine-tuning encoder-only models (e.g., BERT) is a popular choice, these models still exhibit higher inference times in resource-constrained environments. Moreover, although distilled variants (e.g., DistilBERT) exist, they are not tailored for the crisis domain. To address these challenges, we make two key contributions. First, we present CrisisHelpOffer, a novel dataset of 101k tweets collaboratively labelled by generative LLMs and validated by humans, specifically designed to distinguish actionable content from noise. Second, we introduce the first crisis-specific mini models optimized for deployment in resource-constrained settings. Across 13 crisis classification tasks, our mini models surpass BERT (also outperform or match the performance of RoBERTa, MPNet, and BERTweet), offering higher accuracy with significantly smaller sizes and faster speeds. The Medium model is 47% smaller with 3.8% higher accuracy at 3.5x speed, the Small model is 68% smaller with a 1.8% accuracy gain at 7.7x speed, and the Tiny model, 83% smaller, matches BERT's accuracy at 18.6x speed. All models outperform existing distilled variants, setting new benchmarks. Finally, as a case study, we analyze social media posts from a global crisis to explore help-seeking and assistance-offering behaviours in selected developing and developed countries.