In recent years, neural network-based models have achieved state-of-the-art performance across a wide array of tasks. These models effectively capture relevant features or concepts from samples, tailored to the specific prediction tasks they address (Yang and Hu, 2021b; Bordelon and Pehlevan, 2022a; Ba et al., 2022b). A fundamental challenge lies in understanding how these models learn such features and determining whether these features can be interpreted or even retrieved directly (Radhakrishnan et al., 2024). Recent advancements in mechanistic interpretability have opened multiple avenues for elucidating how transformerbased models, including Large Language Models (LLMs), acquire and represent features (Bricken et al., 2023; Doshi-Velez and Kim, 2017). These advances include uncovering neural circuits that encode specific concepts (Marks et al., 2024b; Olah et al., 2020), understanding feature composition across attention layers (Yang and Hu, 2021b), and revealing how models develop structured representations (Elhage et al., 2022). One line of research posits that features are encoded linearly within the latent representation space through sparse activations, a concept known as the linear representation hypothesis (LRH) (Mikolov et al., 2013; Arora et al., 2016). However, this hypothesis faces challenges in explaining how neural networks function, as models often need to represent more distinct features than their layer dimensions would theoretically allow under purely linear encoding. This phenomenon has been studied extensively in the context of large language models through the lens of superposition (Elhage et al., 2022), where multiple features share the same dimensional space in structured ways.

To deliver coherent and personalized experiences in long-term conversations, existing approaches typically perform retrieval augmented response generation by constructing memory banks from conversation history at either the turn-level, session-level, or through summarization techniques. In this paper, we present two key findings: (1) The granularity of memory unit matters: Turn-level, session-level, and summarization-based methods each exhibit limitations in both memory retrieval accuracy and the semantic quality of the retrieved content. (2) Prompt compression methods, such as \textit{LLMLingua-2}, can effectively serve as a denoising mechanism, enhancing memory retrieval accuracy across different granularities. Building on these insights, we propose SeCom, a method that constructs a memory bank with topical segments by introducing a conversation Segmentation model, while performing memory retrieval based on Compressed memory units. Experimental results show that SeCom outperforms turn-level, session-level, and several summarization-based methods on long-term conversation benchmarks such as LOCOMO and Long-MT-Bench+. Additionally, the proposed conversation segmentation method demonstrates superior performance on dialogue segmentation datasets such as DialSeg711, TIAGE, and SuperDialSeg.

Decoder-only language models have the ability to dynamically switch between various computational tasks based on input prompts. Despite many successful applications of prompting, there is very limited understanding of the internal mechanism behind such flexibility. In this work, we investigate how different prompting methods affect the geometry of representations in these models. Employing a framework grounded in statistical physics, we reveal that various prompting techniques, while achieving similar performance, operate through distinct representational mechanisms for task adaptation. Our analysis highlights the critical role of input distribution samples and label semantics in few-shot in-context learning. We also demonstrate evidence of synergistic and interfering interactions between different tasks on the representational level. Our work contributes to the theoretical understanding of large language models and lays the groundwork for developing more effective, representation-aware prompting strategies.

We present FoQA, a Faroese extractive question-answering (QA) dataset with 2,000 samples, created using a semi-automated approach combining Large Language Models (LLMs) and human validation. The dataset was generated from Faroese Wikipedia articles using GPT-4-turbo for initial QA generation, followed by question rephrasing to increase complexity and native speaker validation to ensure quality. We provide baseline performance metrics for FoQA across multiple models, including LLMs and BERT, demonstrating its effectiveness in evaluating Faroese QA performance. The dataset is released in three versions: a validated set of 2,000 samples, a complete set of all 10,001 generated samples, and a set of 2,395 rejected samples for error analysis.

In this study, we introduce the process for creating BiaSWE, an expert-annotated dataset tailored for misogyny detection in the Swedish language. To address the cultural and linguistic specificity of misogyny in Swedish, we collaborated with experts from the social sciences and humanities. Our interdisciplinary team developed a rigorous annotation process, incorporating both domain knowledge and language expertise, to capture the nuances of misogyny in a Swedish context. This methodology ensures that the dataset is not only culturally relevant but also aligned with broader efforts in bias detection for low-resource languages. The dataset, along with the annotation guidelines, is publicly available for further research.

Instruction-tuned for recommendation, we demonstrate that LLMs often exhibit uncertainty LLMs [4, 29, 64, 66] have shown remarkable performance for the in their recommendations. To ensure the trustworthy zero-shot ranking task [23, 25], and can be further fine-tuned with use of LLMs in generating recommendations, we emphasize the the user history logged on the system [2, 19, 81]. Recent methods importance of assessing the reliability of recommendations generated [10, 70, 79, 80] adopt the retrieval-augmented generation paradigm by LLMs. We start by introducing a novel framework for [3, 27], where LLMs are employed to generate ranking lists with candidates estimating the predictive uncertainty to quantitatively measure the retrieved by candidate generators. This approach exhibits reliability of LLM-based recommendations. We further propose to state-of-the-art recommendation performance over conventional decompose the predictive uncertainty into recommendation uncertainty sequential recommenders [31, 63], facilitating better online updates and prompt uncertainty, enabling in-depth analyses of and avoiding hallucination. the primary source of uncertainty. Through extensive experiments, While LLMs have been widely employed in real-world applications we (1) demonstrate predictive uncertainty effectively indicates the that can influence human behavior, there is a lack of exploration reliability of LLM-based recommendations, (2) investigate the origins in assessing the reliability of the LLM-based recommendation. of uncertainty with decomposed uncertainty measures, and Indeed, despite their superior performance, we demonstrate recommendations (3) propose uncertainty-aware prompting for a lower predictive generated by LLMs are highly volatile depending on uncertainty and enhanced recommendation. Our source code and the prompting details (e.g., word choice, number of user histories, model weights are available at https://github.com/WonbinKweon/

We introduce Glinthawk, an architecture for offline Large Language Model (LLM) inference. By leveraging a two-tiered structure, Glinthawk optimizes the utilization of the high-end accelerators ("Tier 1") by offloading the attention mechanism to lower-end compute tier ("Tier 2"). This separation allows the memory demand of the attention, known as the key-value cache, to scale independently from the model weights, enabling larger batch sizes and more efficient accelerator usage. Prototyped with NVIDIA T4 GPUs and standard CPU VMs, Glinthawk improves throughput by $5.9\times$ and reduces cost of generation by $2.8\times$, compared to paged attention baselines. For long sequence lengths, it achieves $16.3\times$ throughput improvement at $2.4\times$ less cost. Our evaluation shows that this architecture can tolerate moderate network latency with minimal performance degradation, making it highly effective for latency-tolerant, throughput-focused applications such as batch processing. The prototype is publicly available at https://github.com/microsoft/glinthawk.

Large language models (LLMs) deliver impressive performance but require large amounts of energy. In this work, we present a MatMul-free LLM architecture adapted for Intel's neuromorphic processor, Loihi 2. Our approach leverages Loihi 2's support for low-precision, event-driven computation and stateful processing. Our hardware-aware quantized model on GPU demonstrates that a 370M parameter MatMul-free model can be quantized with no accuracy loss. Based on preliminary results, we report up to 3x higher throughput with 2x less energy, compared to transformer-based LLMs on an edge GPU, with significantly better scaling. Further hardware optimizations will increase throughput and decrease energy consumption. These results show the potential of neuromorphic hardware for efficient inference and pave the way for efficient reasoning models capable of generating complex, long-form text rapidly and cost-effectively.

Humans have long relied on visual aids like sketches and diagrams to support reasoning and problem-solving. Visual tools, like auxiliary lines in geometry or graphs in calculus, are essential for understanding complex ideas. However, many tutoring systems remain text-based, providing feedback only through natural language. Leveraging recent advances in Large Multimodal Models (LMMs), this paper introduces Interactive Sketchpad, a tutoring system that combines language-based explanations with interactive visualizations to enhance learning. Built on a pre-trained LMM, Interactive Sketchpad is fine-tuned to provide step-by-step guidance in both text and visuals, enabling natural multimodal interaction with the student. Accurate and robust diagrams are generated by incorporating code execution into the reasoning process. User studies conducted on math problems such as geometry, calculus, and trigonometry demonstrate that Interactive Sketchpad leads to improved task comprehension, problem-solving accuracy, and engagement levels, highlighting its potential for transforming educational technologies.

In this work, we conduct an experiment using state-of-the-art LLMs to translate MiniF2F into Rocq. The translation task focuses on generating a Rocq theorem based on three sources: a natural language description, the Lean formalization, and the Isabelle formalization. We conducted our experiment in 3 stages of increasing complexity, from basic one-shot prompting to multi-turn conversations that incorporate feedback from unsuccessful attempts. At each stage, we perform multiple rounds of translation using increasingly advanced models: GPT-4o mini, Claude 3.5 Sonnet, o1 mini, and o1. We successfully translated 478 out of 488 theorems. The dataset is opensource: https://github.com/LLM4Rocq/miniF2F-rocq.