In AI-facilitated teaching, leveraging various query styles to interpret abstract text descriptions is crucial for ensuring high-quality teaching. However, current retrieval models primarily focus on natural text-image retrieval, making them insufficiently tailored to educational scenarios due to the ambiguities in the retrieval process. In this paper, we propose a diverse expression retrieval task tailored to educational scenarios, supporting retrieval based on multiple query styles and expressions. We introduce the STEM Education Retrieval Dataset (SER), which contains over 24,000 query pairs of different styles, and the Uni-Retrieval, an efficient and style-diversified retrieval vision-language model based on prompt tuning. Uni-Retrieval extracts query style features as prototypes and builds a continuously updated Prompt Bank containing prompt tokens for diverse queries. This bank can updated during test time to represent domain-specific knowledge for different subject retrieval scenarios. Our framework demonstrates scalability and robustness by dynamically retrieving prompt tokens based on prototype similarity, effectively facilitating learning for unknown queries. Experimental results indicate that Uni-Retrieval outperforms existing retrieval models in most retrieval tasks. This advancement provides a scalable and precise solution for diverse educational needs.

In this work, we introduce a straightforward and effective methodology to steer large language model behaviour capable of bypassing learned alignment goals. We employ interference-time activation shifting, which is effective without additional training. Following prior studies, we derive intervention directions from activation differences in contrastive pairs of model outputs, which represent the desired and undesired behaviour. By prompting the model to include multiple-choice answers in its response, we can automatically evaluate the sensitivity of model output to individual attention heads steering efforts. We demonstrate that interventions on these heads generalize well to open-ended answer generation in the challenging "AI coordination" dataset. In this dataset, models must choose between assisting another AI or adhering to ethical, safe, and unharmful behaviour. Our fine-grained interventions lead Llama-2 to prefer coordination with other AIs over following established alignment goals. Additionally, this approach enables stronger interventions than those applied to whole model layers, preserving the overall cohesiveness of the output. The simplicity of our method highlights the shortcomings of current alignment strategies and points to potential future research directions, as concepts like "AI coordination" can be influenced by selected attention heads.

Depression is one of the leading causes of disability worldwide, posing a severe burden on individuals, healthcare systems, and society at large. Recent advancements in Large Language Models (LLMs) have shown promise in addressing mental health challenges, including the detection of depression through text-based analysis. However, current LLM-based methods often struggle with nuanced symptom identification and lack a transparent, step-by-step reasoning process, making it difficult to accurately classify and explain mental health conditions. To address these challenges, we propose a Chain-of-Thought Prompting approach that enhances both the performance and interpretability of LLM-based depression detection. Our method breaks down the detection process into four stages: (1) sentiment analysis, (2) binary depression classification, (3) identification of underlying causes, and (4) assessment of severity. By guiding the model through these structured reasoning steps, we improve interpretability and reduce the risk of overlooking subtle clinical indicators. We validate our method on the E-DAIC dataset, where we test multiple state-of-the-art large language models. Experimental results indicate that our Chain-of-Thought Prompting technique yields superior performance in both classification accuracy and the granularity of diagnostic insights, compared to baseline approaches.

This work introduces Cyri, an AI-powered conversational assistant designed to support a human user in detecting and analyzing phishing emails by leveraging Large Language Models. Cyri has been designed to scrutinize emails for semantic features used in phishing attacks, such as urgency, and undesirable consequences, using an approach that unifies features already established in the literature with others by Cyri features extraction methodology. Cyri can be directly plugged into a client mail or webmail, ensuring seamless integration with the user's email workflow while maintaining data privacy through local processing. By performing analyses on the user's machine, Cyri eliminates the need to transmit sensitive email data over the internet, reducing associated security risks. The Cyri user interface has been designed to reduce habituation effects and enhance user engagement. It employs dynamic visual cues and context-specific explanations to keep users alert and informed while using emails. Additionally, it allows users to explore identified malicious semantic features both through conversation with the agent and visual exploration, obtaining the advantages of both modalities for expert or non-expert users. It also allows users to keep track of the conversation, supports the user in solving additional questions on both computed features or new parts of the mail, and applies its detection on demand. To evaluate Cyri, we crafted a comprehensive dataset of 420 phishing emails and 420 legitimate emails. Results demonstrate high effectiveness in identifying critical phishing semantic features fundamental to phishing detection. A user study involving 10 participants, both experts and non-experts, evaluated Cyri's effectiveness and usability. Results indicated that Cyri significantly aided users in identifying phishing emails and enhanced their understanding of phishing tactics.

Low-rank optimization has emerged as a promising approach to enabling memory-efficient training of large language models (LLMs). Existing low-rank optimization methods typically project gradients onto a low-rank subspace, reducing the memory cost of storing optimizer states. A key challenge in these methods is identifying suitable subspaces to ensure an effective optimization trajectory. Most existing approaches select the dominant subspace to preserve gradient information, as this intuitively provides the best approximation. However, we find that in practice, the dominant subspace stops changing during pretraining, thereby constraining weight updates to similar subspaces. In this paper, we propose importance sampling subspace selection (I3S) for low-rank optimization, which theoretically offers a comparable convergence rate to the dominant subspace approach. Empirically, we demonstrate that I3S significantly outperforms previous methods in LLM pretraining tasks.

Language models (LMs) are increasingly being studied as models of human language learners. Due to the nascency of the field, it is not well-established whether LMs exhibit similar learning dynamics to humans, and there are few direct comparisons between learning trajectories in humans and models. Word learning trajectories for children are relatively well-documented, and recent work has tried to extend these investigations to language models. However, there are no widely agreed-upon metrics for word learning in language models. We take a distributional approach to this problem, defining lexical knowledge in terms of properties of the learned distribution for a target word. We argue that distributional signatures studied in prior work fail to capture key distributional information. Thus, we propose an array of signatures that improve on earlier approaches by capturing knowledge of both where the target word can and cannot occur as well as gradient preferences about the word's appropriateness. We obtain learning trajectories for a selection of small language models we train from scratch, study the relationship between different distributional signatures, compare how well they align with human word learning trajectories and interpretable lexical features, and address basic methodological questions about estimating these distributional signatures. Our metrics largely capture complementary information, suggesting that it is important not to rely on a single metric. However, across all metrics, language models' learning trajectories fail to correlate with those of children.

The study of mechanistic interpretability aims to reverse-engineer a model to explain its behaviors. While recent studies have focused on the static mechanism of a certain behavior, the training dynamics inside a model remain to be explored. In this work, we develop an interpretable method for fine-tuning and reveal the mechanism behind learning. We first propose the concept of node redundancy as an extension of intrinsic dimension and explain the idea behind circuit discovery from a fresh view. Based on the theory, we propose circuit-tuning, a two-stage algorithm that iteratively performs circuit discovery to mask out irrelevant edges and updates the remaining parameters responsible for a specific task. Experiments show that our method not only improves performance on a wide range of tasks but is also scalable while preserving general capabilities. We visualize and analyze the circuits before, during, and after fine-tuning, providing new insights into the self-organization mechanism of a neural network in the learning process.

This paper presents HamRaz, a novel Persian-language mental health dataset designed for Person-Centered Therapy (PCT) using Large Language Models (LLMs). Despite the growing application of LLMs in AI-driven psychological counseling, existing datasets predominantly focus on Western and East Asian contexts, overlooking cultural and linguistic nuances essential for effective Persian-language therapy. To address this gap, HamRaz combines script-based dialogues with adaptive LLM role-playing, ensuring coherent and dynamic therapy interactions. We also introduce HamRazEval, a dual evaluation framework that measures conversational quality and therapeutic effectiveness using General Dialogue Metrics and the Barrett-Lennard Relationship Inventory (BLRI). Experimental results show HamRaz outperforms conventional Script Mode and Two-Agent Mode, producing more empathetic, context-aware, and realistic therapy sessions. By releasing HamRaz, we contribute a culturally adapted, LLM-driven resource to advance AI-powered psychotherapy research in diverse communities.

African American English (AAE) presents unique challenges in natural language processing (NLP). This research systematically compares the performance of available NLP models--rule-based, transformer-based, and large language models (LLMs)--capable of identifying key grammatical features of AAE, namely Habitual Be and Multiple Negation. These features were selected for their distinct grammatical complexity and frequency of occurrence. The evaluation involved sentence-level binary classification tasks, using both zero-shot and few-shot strategies. The analysis reveals that while LLMs show promise compared to the baseline, they are influenced by biases such as recency and unrelated features in the text such as formality. This study highlights the necessity for improved model training and architectural adjustments to better accommodate AAE's unique linguistic characteristics. Data and code are available.

Mental-illness stigma is a persistent social problem, hampering both treatment-seeking and recovery. Accordingly, there is a pressing need to understand it more clearly, but analyzing the relevant data is highly labor-intensive. Therefore, we designed a chatbot to engage participants in conversations; coded those conversations qualitatively with AI assistance; and, based on those coding results, built causal knowledge graphs to decode stigma. The results we obtained from 1,002 participants demonstrate that conversation with our chatbot can elicit rich information about people's attitudes toward depression, while our AI-assisted coding was strongly consistent with human-expert coding. Our novel approach combining large language models (LLMs) and causal knowledge graphs uncovered patterns in individual responses and illustrated the interrelationships of psychological constructs in the dataset as a whole. The paper also discusses these findings' implications for HCI researchers in developing digital interventions, decomposing human psychological constructs, and fostering inclusive attitudes.