Underlying mechanisms of memorization in LLMs -- the verbatim reproduction of training data -- remain poorly understood. What exact part of the network decides to retrieve a token that we would consider as start of memorization sequence? How exactly is the models' behaviour different when producing memorized sentence vs non-memorized? In this work we approach these questions from mechanistic interpretability standpoint by utilizing transformer circuits -- the minimal computational subgraphs that perform specific functions within the model. Through carefully constructed contrastive datasets, we identify points where model generation diverges from memorized content and isolate the specific circuits responsible for two distinct aspects of memorization. We find that circuits that initiate memorization can also maintain it once started, while circuits that only maintain memorization cannot trigger its initiation. Intriguingly, memorization prevention mechanisms transfer robustly across different text domains, while memorization induction appears more context-dependent.

Evaluation benchmark characteristics may distort the true benefits of domain adaptation in retrieval models. This creates misleading assessments that influence deployment decisions in specialized domains. We show that two benchmarks with drastically different features such as topic diversity, boundary overlap, and semantic complexity can influence the perceived benefits of fine-tuning. Using environmental regulatory document retrieval as a case study, we fine-tune ColBERTv2 model on Environmental Impact Statements (EIS) from federal agencies. We evaluate these models across two benchmarks with different semantic structures. Our findings reveal that identical domain adaptation approaches show very different perceived benefits depending on evaluation methodology. On one benchmark, with clearly separated topic boundaries, domain adaptation shows small improvements (maximum 0.61% NDCG gain). However, on the other benchmark with overlapping semantic structures, the same models demonstrate large improvements (up to 2.22% NDCG gain), a 3.6-fold difference in the performance benefit. We compare these benchmarks through topic diversity metrics, finding that the higher-performing benchmark shows 11% higher average cosine distances between contexts and 23% lower silhouette scores, directly contributing to the observed performance difference. These results demonstrate that benchmark selection strongly determines assessments of retrieval system effectiveness in specialized domains. Evaluation frameworks with well-separated topics regularly underestimate domain adaptation benefits, while those with overlapping semantic boundaries reveal improvements that better reflect real-world regulatory document complexity. Our findings have important implications for developing and deploying AI systems for interdisciplinary domains that integrate multiple topics.

Recent advancements in Large Language Models (LLMs) have demonstrated remarkable capabilities in various domains. However, effective decision-making relies heavily on strong reasoning abilities. Reasoning is the foundation for decision-making, providing the analytical and logical framework to make sound choices. Reasoning involves analyzing information, drawing inferences, and reaching conclusions based on logic or evidence. Decision-making builds on this foundation by applying the insights from reasoning to select the best course of action among alternatives. Together, these processes create a continuous cycle of thought and action aimed at achieving goals effectively. As AI technology evolves, there is a growing trend to train LLMs to excel in general reasoning. This study explores how the general reasoning capabilities of LLMs connect to their performance in domain-specific reasoning tasks.

Sequential recommendation aims to predict users' future interactions by modeling collaborative filtering (CF) signals from historical behaviors of similar users or items. Traditional sequential recommenders predominantly rely on ID-based embeddings, which capture CF signals through high-order co-occurrence patterns. However, these embeddings depend solely on past interactions, lacking transferable knowledge to generalize to unseen domains. Recent advances in large language models (LLMs) have motivated text-based recommendation approaches that derive item representations from textual descriptions. While these methods enhance generalization, they fail to encode CF signals-i.e., latent item correlations and preference patterns-crucial for effective recommendation. We argue that an ideal embedding model should seamlessly integrate CF signals with rich semantic representations to improve both in-domain and out-of-domain recommendation performance. To this end, we propose LLM2Rec, a novel embedding model tailored for sequential recommendation, integrating the rich semantic understanding of LLMs with CF awareness. Our approach follows a two-stage training framework: (1) Collaborative Supervised Fine-tuning, which adapts LLMs to infer item relationships based on historical interactions, and (2) Item-level Embedding Modeling, which refines these specialized LLMs into structured item embedding models that encode both semantic and collaborative information. Extensive experiments on real-world datasets demonstrate that LLM2Rec effectively improves recommendation quality across both in-domain and out-of-domain settings. Our findings highlight the potential of leveraging LLMs to build more robust, generalizable embedding models for sequential recommendation. Our codes are available at https://github.com/HappyPointer/LLM2Rec.

The evaluation of Large Language Models (LLMs) in healthcare has been dominated by physician-centric, English-language benchmarks, creating a dangerous illusion of competence that ignores the interprofessional nature of patient care. To provide a more holistic and realistic assessment, we introduce HealthQA-BR, the first large-scale, system-wide benchmark for Portuguese-speaking healthcare. Comprising 5,632 questions from Brazil's national licensing and residency exams, it uniquely assesses knowledge not only in medicine and its specialties but also in nursing, dentistry, psychology, social work, and other allied health professions. We conducted a rigorous zero-shot evaluation of over 20 leading LLMs. Our results reveal that while state-of-the-art models like GPT 4.1 achieve high overall accuracy (86.6%), this top-line score masks alarming, previously unmeasured deficiencies. A granular analysis shows performance plummets from near-perfect in specialties like Ophthalmology (98.7%) to barely passing in Neurosurgery (60.0%) and, most notably, Social Work (68.4%). This "spiky" knowledge profile is a systemic issue observed across all models, demonstrating that high-level scores are insufficient for safety validation. By publicly releasing HealthQA-BR and our evaluation suite, we provide a crucial tool to move beyond single-score evaluations and toward a more honest, granular audit of AI readiness for the entire healthcare team.

Optimizing instructions for large language models (LLMs) is critical for harnessing their full potential in complex and diverse tasks. However, relying solely on white-box approaches demands extensive computational resources and offers limited representational capacity, while black-box models can incur prohibitive financial costs. To address these challenges, we introduce a novel framework that seamlessly merges the strengths of both paradigms. Black-box models provide high-quality, diverse instruction initializations, and white-box models supply fine-grained interpretability through hidden states and output features. By enforcing a semantic similarity constraint, these components fuse into a unified high-dimensional representation that captures deep semantic and structural nuances, enabling an iterative optimization process to refine instruction quality and adaptability. Extensive evaluations across a broad spectrum of tasks-ranging from complex reasoning to cross-lingual generalization-demonstrate that our approach consistently outperforms state-of-the-art baselines. This fusion of black-box initialization with advanced semantic refinement yields a scalable and efficient solution, paving the way for next-generation LLM-driven applications in diverse real-world scenarios. The source code will be released soon.

Recent works have demonstrated the effectiveness of adapting pre-trained language models (LMs) for forecasting time series in the low-data regime. We build upon these findings by analyzing the effective transfer from language models to time series forecasting under various design choices including upstream post-training, time series tokenizer and language backbone size. In the low-data regime, these design choices have a significant impact on the validation loss, with clear-cut choices that outperform others. Contrary to Hernandez et al. (2021), we observe that the validation loss of the LMs continues to smoothly decrease long after the validation loss of the randomly initialized models has converged, leading to a non-vanishing transfer gap that holds across design choices. These findings not only help shed light on the effective use of compute-efficient training for time series, but also open the way for the study of modality-agnostic properties of data distributions leveraged by these models.

--SystemV erilog Assertion s ( SV As) are critical for verifying the correctness of hardware designs, but manually writing them from natural language property descriptions, i.e., NL2SV A, remains a labor-intensive and error-prone task. T o enhance LLM performance in NL2SV A, we propose a customized retrieval-augmented generation (RAG) framework and a synthetic fine-tuning dataset that together improve LLM's performance. Our RAG framework (i) constructs a context-preserving database via dynamic splitting technique, (ii) combines global semantic retrieval with keyword-guided retrieval to extract SV A operator-related contexts via HybridRetrieval, and (iii) validate and correct the use of SV A operators in LLM-generated SV As via SV A operator-based rechecking . T o further improve lightweight models over NL2SV A, our fine-tuning dataset provides prompt-guided explanation s that teach LLMs the layer-by-layer construction process of concurrent SV As, enabling supervised fine-tuning that greatly improves syntax and functionality accuracy. T o evaluate the performance of LLMs over NL2SV A, we construct the largest evaluation dataset for NL2SV A, comprising 40 V erilog designs and 229 formally verified SV As with detailed annotations. Experimental results show that our customized RAG framework increases the number of functionality matched SV As by 58. SystemV erilog Assertion s ( SVAs) are essential tools in hardware verification, formally specifying expected design behaviors, namely the design properties, and serving as embedded checkers that continuously validate the implementation against its specification [1], [2]. However, writing SV As manually is difficult, which consists of two sub-tasks [3], [4]. The first task is to extract intended properties, described in natural language, from hardware designs and detailed specification documents. Once these are identified, the second task is to implement these natural language properties as SV As, i.e., the NL2SV A task.

Resource efficiency is a critical barrier to deploying large language models (LLMs) in edge and privacy-sensitive applications. This study evaluates the efficacy of two augmentation strategies--Retrieval-Augmented Generation (RAG) and Hypothetical Document Embeddings (HyDE)--on compact Gemma LLMs of 1 billion and 4 billion parameters, within the context of a privacy-first personal assistant. We implement short-term memory via MongoDB and long-term semantic storage via Qdrant, orchestrated through FastAPI and LangChain, and expose the system through a React.js frontend. Across both model scales, RAG consistently reduces latency by up to 17\% and eliminates factual hallucinations when responding to user-specific and domain-specific queries. HyDE, by contrast, enhances semantic relevance--particularly for complex physics prompts--but incurs a 25--40\% increase in response time and a non-negligible hallucination rate in personal-data retrieval. Comparing 1 B to 4 B models, we observe that scaling yields marginal throughput gains for baseline and RAG pipelines, but magnifies HyDE's computational overhead and variability. Our findings position RAG as the pragmatic choice for on-device personal assistants powered by small-scale LLMs.

While large language models (LLMs) have demonstrated impressive performance across a wide range of natural language processing (NLP) tasks in high-resource languages, their capabilities in low-resource and minority languages remain significantly underexplored. Formosan languages -- a subgroup of Austronesian languages spoken in Taiwan -- are both linguistically rich and endangered, largely due to the sociolinguistic dominance of Mandarin. In this work, we introduce FORMOSANBENCH, the first benchmark for evaluating LLMs on low-resource Austronesian languages. It covers three endangered Formosan languages: Atayal, Amis, and Paiwan, across three core NLP tasks: machine translation, automatic speech recognition (ASR), and text summarization. We assess model performance in zero-shot, 10-shot, and fine-tuned settings using FORMOSANBENCH. Our results reveal a substantial performance gap between high-resource and Formosan languages. Existing LLMs consistently underperform across all tasks, with 10-shot learning and fine-tuning offering only limited improvements. These findings underscore the urgent need for more inclusive NLP technologies that can effectively support endangered and underrepresented languages. We release our datasets and code to facilitate future research in this direction.