Data-Free Quantization (DFQ) enables the quantization of Vision Transformers (ViTs) without requiring access to data, allowing for the deployment of ViTs on devices with limited resources. In DFQ, the quantization model must be calibrated using synthetic samples, making the quality of these synthetic samples crucial. Existing methods fail to fully capture and balance the global and local features within the samples, resulting in limited synthetic data quality. Moreover, we have found that during inference, there is a significant difference in the distributions of intermediate layer activations between the quantized and full-precision models. These issues lead to a severe performance degradation of the quantized model. To address these problems, we propose a pipeline for Data-Free Quantization for Vision Transformers (DFQ-ViT). Specifically, we synthesize samples in order of increasing difficulty, effectively enhancing the quality of synthetic data. During the calibration and inference stage, we introduce the activation correction matrix for the quantized model to align the intermediate layer activations with those of the full-precision model. Extensive experiments demonstrate that DFQ-ViT achieves remarkable superiority over existing DFQ methods and its performance is on par with models quantized through real data. For example, the performance of DeiT-T with 3-bit weights quantization is 4.29% higher than the state-of-the-art. Our method eliminates the need for fine-tuning, which not only reduces computational overhead but also lowers the deployment barriers for edge devices. This characteristic aligns with the principles of Green Learning by improving energy efficiency and facilitating real-world applications in resource-constrained environments.

Prior work proposed simple test-time scaling, a method for replicating this scaling behavior with models distilled from o1-like models by manually controlling test-time compute: either scaling down by enforcing a maximum length or scaling up by iteratively appending "Wait" when the model is about to terminate its generation. This paper presents an analysis of simple test-time scaling and finds that the scaling behavior is largely attributed to scaling down by enforcing a maximum length. In contrast, fine-tuning on long CoT data distilled from o1-like models has no significant impact on scaling behavior, and scaling up by appending "Wait" leads to inconsistencies, as the model may oscillate between solutions. A key distinction exists between scaling down by enforcing a maximum length and scaling up test-time compute in o1-like models, such as DeepSeek-R1\@. These models are typically allowed to utilize as much compute as needed, with the only constraint being the model's maximum supported length. By learning to naturally scale up test-time compute during reinforcement learning, o1-like models surpass their peak performance when scaling up. In contrast, simple test-time scaling progressively imposes a lower upper limit on model performance as it scales down. While replicating the test-time scaling behavior of o1 models can be straightforward by scaling down, it is crucial to recognize that the goal of scaling test-time compute is to unlock higher performance -- beyond what the model could originally achieve -- rather than merely reproducing the appearance of scaling behavior.

One challenge in technical interviews is the think-aloud process, where candidates verbalize their thought processes while solving coding tasks. Despite its importance, opportunities for structured practice remain limited. Conversational AI offers potential assistance, but limited research explores user perceptions of its role in think-aloud practice. To address this gap, we conducted a study with 17 participants using an LLM-based technical interview practice tool. Participants valued AI's role in simulation, feedback, and learning from generated examples. Key design recommendations include promoting social presence in conversational AI for technical interview simulation, providing feedback beyond verbal content analysis, and enabling crowdsourced think-aloud examples through human-AI collaboration. Beyond feature design, we examined broader considerations, including intersectional challenges and potential strategies to address them, how AI-driven interview preparation could promote equitable learning in computing careers, and the need to rethink AI's role in interview practice by suggesting a research direction that integrates human-AI collaboration.

Socially assistive robots (SARs) have shown great potential for supplementing well-being support. However, prior studies have found that existing dialogue pipelines for SARs remain limited in real-time latency, back-channeling, and personalized speech dialogue. Toward addressing these limitations, we propose using integrated end-to-end speech-language models (SLMs) with SARs. This work 1) evaluated the usability of an SLM-enabled SAR dialogue system through a small user study, and 2) identified remaining limitations through study user feedback to inform future improvements. We conducted a small within-participant user study with university students (N = 11) whose results showed that participants perceived an SLM-enabled SAR system as capable of providing empathetic feedback, natural turn-taking, back-channeling, and adaptive responses. We also found that participants reported the robot's nonverbal behaviors as lacking variability and synchronization with conversation, and the SLM's verbal feedback as generic and repetitive. These findings highlighted the need for real-time robot movement synchronized with conversation, improved prompting or fine-tuning to generate outputs better aligned with mental health practices, and more expressive, adaptive vocal generation.

State-of-the-art reasoning LLMs are powerful problem solvers, but they still occasionally make mistakes. However, adopting AI models in risk-sensitive domains often requires error rates near 0%. To address this gap, we propose collaboration between a reasoning model and a human expert who resolves queries the model cannot confidently answer. We find that quantifying the uncertainty of a reasoning model through the length of its reasoning trace yields an effective basis for deferral to a human, e.g., cutting the error rate of Qwen3 235B-A22B on difficult MATH problems from 3% to less than 1% when deferring 7.5% of queries. However, the high latency of reasoning models still makes them challenging to deploy on use cases with high query volume. To address this challenge, we explore fronting a reasoning model with a large non-reasoning model. We call this modified human-in-the-loop system "Fail Fast, or Ask", since the non-reasoning model may defer difficult queries to the human expert directly ("failing fast"), without incurring the reasoning model's higher latency. We show that this approach yields around 40% latency reduction and about 50% cost savings for DeepSeek R1 while maintaining 90+% area under the accuracy-rejection curve. However, we observe that latency savings are lower than expected because of "latency drag", the phenomenon that processing easier queries with a non-reasoning model pushes the reasoning model's latency distribution towards longer latencies. Broadly, our results suggest that the deficiencies of state-of-the-art reasoning models -- nontrivial error rates and high latency -- can be substantially mitigated through black-box systems engineering, without requiring access to LLM internals.

The rise of Large Reasoning Models (LRMs) promises a significant leap forward in language model capabilities, aiming to tackle increasingly sophisticated tasks with unprecedented efficiency and accuracy. However, despite their impressive performance, recent studies have highlighted how current reasoning models frequently fail to generalize to novel, unseen problems, often resorting to memorized solutions rather than genuine inferential reasoning. Such behavior underscores a critical limitation in modern LRMs, i.e., their tendency toward overfitting, which in turn results in poor generalization in problem-solving capabilities. In this paper, we introduce Nexus Architect, an enhanced iteration of our multi-agent system framework, Nexus, equipped with a novel automated workflow synthesis mechanism. Given a user's prompt and a small set of representative examples, the Architect autonomously generates a tailored reasoning workflow by selecting suitable strategies, tool integrations, and adversarial techniques for a specific problem class. Furthermore, the Architect includes an iterative prompt refinement mechanism that fine-tunes agents' system prompts to maximize performance and improve the generalization capabilities of the system. We empirically evaluate Nexus Architect by employing an off-the-shelf, non-reasoning model on a custom dataset of challenging logical questions and compare its performance against state-of-the-art LRMs. Results show that Nexus Architect consistently outperforms existing solutions, achieving up to a 66% increase in pass rate over Gemini 2.5 Flash Preview, nearly 2.5$\times$ against Claude Sonnet 4 and DeepSeek-R1, and over 3$\times$ w.r.t. Llama 4 Scout.

In this study, we investigate the use of large language models (LLMs), specifically ChatGPT, for structured deductive qualitative coding. While most current research emphasizes inductive coding applications, we address the underexplored potential of LLMs to perform deductive classification tasks aligned with established human-coded schemes. Using the Comparative Agendas Project (CAP) Master Codebook, we classified U.S. Supreme Court case summaries into 21 major policy domains. We tested four intervention methods: zero-shot, few-shot, definition-based, and a novel Step-by-Step Task Decomposition strategy, across repeated samples. Performance was evaluated using standard classification metrics (accuracy, F1-score, Cohen's kappa, Krippendorff's alpha), and construct validity was assessed using chi-squared tests and Cramer's V. Chi-squared and effect size analyses confirmed that intervention strategies significantly influenced classification behavior, with Cramer's V values ranging from 0.359 to 0.613, indicating moderate to strong shifts in classification patterns. The Step-by-Step Task Decomposition strategy achieved the strongest reliability (accuracy = 0.775, kappa = 0.744, alpha = 0.746), achieving thresholds for substantial agreement. Despite the semantic ambiguity within case summaries, ChatGPT displayed stable agreement across samples, including high F1 scores in low-support subclasses. These findings demonstrate that with targeted, custom-tailored interventions, LLMs can achieve reliability levels suitable for integration into rigorous qualitative coding workflows.

Schema matching is essential for integrating heterogeneous data sources and enhancing dataset discovery, yet it remains a complex and resource-intensive problem. We introduce SCHEMORA, a schema matching framework that combines large language models with hybrid retrieval techniques in a prompt-based approach, enabling efficient identification of candidate matches without relying on labeled training data or exhaustive pairwise comparisons. By enriching schema metadata and leveraging both vector-based and lexical retrieval, SCHEMORA improves matching accuracy and scalability. Evaluated on the MIMIC-OMOP benchmark, it establishes new state-of-the-art performance, with gains of 7.49% in HitRate@5 and 3.75% in HitRate@3 over previous best results. To our knowledge, this is the first LLM-based schema matching method with an open-source implementation, accompanied by analysis that underscores the critical role of retrieval and provides practical guidance on model selection.

Relation Classification (RC) in biomedical texts is essential for constructing knowledge graphs and enabling applications such as drug repurposing and clinical decision-making. We propose an error-aware teacher--student framework that improves RC through structured guidance from a large language model (GPT-4o). Prediction failures from a baseline student model are analyzed by the teacher to classify error types, assign difficulty scores, and generate targeted remediations, including sentence rewrites and suggestions for KG-based enrichment. These enriched annotations are used to train a first student model via instruction tuning. This model then annotates a broader dataset with difficulty scores and remediation-enhanced inputs. A second student is subsequently trained via curriculum learning on this dataset, ordered by difficulty, to promote robust and progressive learning. We also construct a heterogeneous biomedical knowledge graph from PubMed abstracts to support context-aware RC. Our approach achieves new state-of-the-art performance on 4 of 5 PPI datasets and the DDI dataset, while remaining competitive on ChemProt.

The introduction of large language models has brought rapid progress on Text-to-SQL benchmarks, but it is not yet easy to build a working enterprise solution. In this paper, we present insights from building an internal chatbot that enables LinkedIn's product managers, engineers, and operations teams to self-serve data insights from a large, dynamic data lake. Our approach features three components. First, we construct a knowledge graph that captures up-to-date semantics by indexing database metadata, historical query logs, wikis, and code. We apply clustering to identify relevant tables for each team or product area. Second, we build a Text-to-SQL agent that retrieves and ranks context from the knowledge graph, writes a query, and automatically corrects hallucinations and syntax errors. Third, we build an interactive chatbot that supports various user intents, from data discovery to query writing to debugging, and displays responses in rich UI elements to encourage follow-up chats. Our chatbot has over 300 weekly users. Expert review shows that 53% of its responses are correct or close to correct on an internal benchmark set. Through ablation studies, we identify the most important knowledge graph and modeling components, offering a practical path for developing enterprise Text-to-SQL solutions.