Post-training quantization (PTQ) reduces the memory footprint of LLMs by quan-tizing weights to low-precision datatypes. Since LLM inference is usually memory-bound, PTQ methods can improve inference throughput.

Match-and-copy is a core retrieval primitive used at inference time by large language models to retrieve a matching token from the context then copy its successor. Yet, understanding how this behavior emerges on natural data is challenging because retrieval and memorization are entangled. To disentangle the two, we introduce Gaussian Match-and-Copy (GMC), a minimalist benchmark that isolates long-range retrieval through pure second-order correlation signals. Numerical investigations show that this task retains key qualitative aspects of how Transformers develop match-and-copy circuits in practice, and separates architectures by their retrieval capabilities. We also analyze the optimization dynamics in a simplified attention setting. Although many solutions are a priori possible under a regression objective, including ones that do not implement retrieval, we identify an implicit-bias regime in which gradient descent drives the parameters to diverge while their direction aligns with the max-margin separator, yielding hard match selection. We prove this max-margin alignment for GD trajectories that reach vanishing empirical loss under explicit technical conditions.

Abstract--Personalized review response generation presents a significant challenge in domains where user information is limited, such as food delivery platforms. While large language models (LLMs) offer powerful text generation capabilities, they often produce generic responses when lacking contextual user data, reducing engagement and effectiveness. In this work, we propose a two-stage prompting framework that infers both explicit (e.g., user-stated preferences) and implicit (e.g., demographic or stylistic cues) personas directly from short review texts. These inferred persona attributes are then incorporated into the response generation prompt to produce user-tailored replies. T o encourage diverse yet faithful generations, we adjust decoding temperature during inference. We evaluate our method using a real-world dataset collected from a Korean food delivery app, and assess its impact on precision, diversity, and semantic consistency. Our findings highlight the effectiveness of persona-augmented prompting in enhancing the relevance and personalization of automated responses without requiring model fine-tuning.

We investigate whether Large Language Models (LLMs) exhibit human-like cognitive patterns under four established frameworks from psychology: Thematic Apperception Test (TAT), Framing Bias, Moral Foundations Theory (MFT), and Cognitive Dissonance. We evaluated several proprietary and open-source models using structured prompts and automated scoring. Our findings reveal that these models often produce coherent narratives, show susceptibility to positive framing, exhibit moral judgments aligned with Liberty/Oppression concerns, and demonstrate self-contradictions tempered by extensive rationalization. Such behaviors mirror human cognitive tendencies yet are shaped by their training data and alignment methods. We discuss the implications for AI transparency, ethical deployment, and future work that bridges cognitive psychology and AI safety

Abstract--Generative Artificial Intelligence Models like Large Language Models (LLMs) and Large Vision Models (VLMs) exhibit state-of-the-art performance across a wide range of tasks but remain vulnerable to hardware-based threats, specifically bit-flip attacks (BF As), posing a serious risk to their security in safety-critical applications. Existing BF A discovery methods--gradient-based, static analysis, and search-based--lack generalizability and struggle to scale, often failing to analyze the vast parameter space and complex interdependencies of modern foundation models in a reasonable time. This paper proposes FlipLLM, a reinforcement learning (RL) architecture-agnostic framework that formulates BF A discovery as a sequential decision-making problem. FlipLLM combines sensitivity-guided layer pruning with Q-learning to efficiently identify minimal, high-impact bit sets capable of inducing catastrophic failure. We demonstrate the effectiveness and generalizability of FlipLLM by applying it to a diverse set of models, including prominent text-only LLMs (GPT -2 Large, LLaMA 3.1 8B, and DeepSeek-V2 7B), VLMs such as LLaV A 1.6, and datasets, such as MMLU, MMLU-Pro, VQA v2, and T extVQA. Our results show that FlipLLM can identify critical bits that are vulnerable to BF As up to 2.5 faster than SOT A methods. We demonstrate that flipping the FlipLLM-identified bits plummets the accuracy of LLaMA 3.1 8B from 69.9% to 0.2%, and for LLaV A's VQA score from 78% to almost 0%, by flipping as few as 5 and 7 bits, respectively. Further analysis shows that applying standard hardware protection mechanisms, such as ECC SECDED, to the FlipLLM-identified bit locations completely mitigates the BF A impact, demonstrating the practical value of our framework for guiding hardware-level defenses. FlipLLM offers the first scalable and adaptive methodology for exploring the BF A vulnerability of both language and multimodal foundation models, paving the way for comprehensive hardware-security evaluation. Generative Artificial Intelligence models like Large Language Models (LLMs) [1] and Large Vision Models (VLMs) represent a transformative advancement in artificial intelligence, finding integration into mission-critical systems spanning healthcare, finance, and autonomous navigation [2], [3]. Their effective deployment mandates reliable and secure operation across diverse hardware infrastructures, from expansive cloud accelerators to resource-constrained edge devices.

Consistency-based methods have emerged as an effective approach to uncertainty quantification (UQ) in large language models. These methods typically rely on several generations obtained via multinomial sampling, measuring their agreement level. However, in short-form QA, multinomial sampling is prone to producing duplicates due to peaked distributions, and its stochasticity introduces considerable variance in uncertainty estimates across runs. We introduce a new family of methods that employ beam search to generate candidates for consistency-based UQ, yielding improved performance and reduced variance compared to multinomial sampling. We also provide a theoretical lower bound on the beam set probability mass under which beam search achieves a smaller error than multinomial sampling. We empirically evaluate our approach on six QA datasets and find that its consistent improvements over multinomial sampling lead to state-of-the-art UQ performance.

Verifiers can improve language model capabilities by scoring and ranking responses from generated candidates. Currently, high-quality verifiers are either unscalable (e.g., humans) or limited in utility (e.g., tools like Lean). While LM judges and reward models have become broadly useful as general-purpose verifiers, a significant performance gap remains between them and oracle verifiers (verifiers with perfect accuracy). To help close this gap, we introduce Weaver, a framework for designing a strong verifier by combining multiple weak, imperfect verifiers. We find weighted ensembles of verifiers, which typically require learning from labeled data, significantly outperform unweighted combinations due to differences in verifier accuracies. To reduce dependency on labeled data, Weaver leverages weak supervision to estimate each verifier's accuracy and combines outputs into a unified score that better reflects true response quality. However, directly applying weak supervision algorithms poses challenges, including inconsistent verifier output formats and handling low-quality verifiers. Weaver addresses these using dataset statistics to normalize outputs and filter specific verifiers. We study Weaver's effectiveness in test-time repeated sampling, where a model generates multiple candidate responses and selects one. Our evaluations show Weaver significantly improves over Pass@1-performance when selecting the first candidate-across reasoning and math tasks, achieving o3-mini-level accuracy with Llama 3.3 70B Instruct as generator, and an ensemble of 70B or smaller judge and reward models as verifiers (87.7% average). This gain mirrors the jump between GPT-4o and o3-mini (69.0% vs. 86.7%), which required extensive finetuning and post-training. To reduce computational costs of verifier ensembles, we train a 400M cross-encoder using Weaver's combined output scores.

Large Language Models (LLMs) demonstrate exceptional capabilities across general domains, yet their application to specialized sectors such as mortgage finance requires domain-specific knowledge augmentation while preserving instruction-following fidelity. We present MortgageLLM, a novel domain-specific large language model that addresses this dual challenge. It is developed using a dual-track specialization framework from a single base model (LLaMA-3.1-8B). We opted for this dual-expert approach as a single multi-task model suffers from performance trade-offs, where optimizing for structured tasks (via SFT) degrades conversational fidelity (via DPO). Our dual-track method solves this by creating two specialists, allowing each to be optimally trained for its distinct capability. Our approach applies the instruction residual technique to restore instruction-following capabilities post-domain adaptation without supervised fine-tuning. We contribute: (1) application of this residual technique to the highly specialized mortgage finance domain; (2) a dual-expert architecture combining a conversational Q&A model and a structured task model for classification and summarization; and (3) an intelligent task routing mechanism using few-shot classification performed by one of the expert models itself. We validate our approach on domain-specific benchmarks, where our final model (MLM v2) significantly outperforms the base LLaMA-3.1-8B-Instruct, achieving an LLM-as-a-Judge summarization score of 4.58 (vs. 3.99), a Q&A score of 4.09 (vs. 4.0), and a classification score of 2.6 (vs. 1.2). On semantic similarity, our model achieved a BERTScore of 0.77 for summarization (vs. 0.74), 0.68 for Q&A (vs. 0.58), and 0.75 for classification (vs. 0.73), substantially outperforming baseline approaches.

Modern Large Language Models (LLMs) are increasingly trained to support very large context windows. Unfortunately the ability to use long contexts in generation is complicated by the large memory requirement of the KV cache, which scales linearly with the context length. This memory footprint is often the dominant resource bottleneck in real-world deployments, limiting throughput and increasing serving costs. One way to address this is by compressing the KV cache, which can be done either with knowledge of the question being asked (query-aware) or without knowledge of the query (query-agnostic). We present Compactor, a training-free, query-agnostic KV compression strategy that uses approximate leverage scores to determine token importance. We show that Compactor can achieve the same performance as competing methods while retaining 20% fewer tokens in both synthetic and real-world context tasks, while being far more task-robust. We further introduce a procedure for context-calibrated compression: inferring the maximum compression a given context supports before significant performance loss. Using context-calibrated compression, we show that Compactor achieves full KV performance on Longbench while reducing the KV memory burden by 68%, on average. To demonstrate the efficacy and generalizability of our approach, we apply Compactor to 27 synthetic and real-world tasks from RULER and Longbench, with models from both the Qwen 2.5 and Llama 3.1 families.