Recent advancements in Large Language Models (LLMs) have created new opportunities to enhance performance on complex reasoning tasks by leveraging test-time computation. However, conventional approaches such as repeated sampling with majority voting or reward model scoring, often face diminishing returns as test-time compute scales, in addition to requiring costly task-specific reward model training. In this paper, we present Self-Enhanced Test-Time Scaling (SETS), a novel method that leverages the self-verification and self-correction capabilities of recent advanced LLMs to overcome these limitations. SETS integrates sampling, self-verification, and self-correction into a unified framework, enabling efficient and scalable test-time computation for improved capabilities at complex tasks. Through extensive experiments on challenging planning and reasoning benchmarks, compared to the alternatives, we demonstrate that SETS achieves significant performance improvements and more favorable test-time scaling laws.

Recent studies show that in supervised fine-tuning (SFT) of large language models (LLMs), data quality matters more than quantity. While most data cleaning methods concentrate on filtering entire samples, the quality of individual tokens within a sample can vary significantly. After pre-training, even in high-quality samples, patterns or phrases that are not task-related can be redundant or uninformative. Continuing to fine-tune on these patterns may offer limited benefit and even degrade downstream task performance. In this paper, we investigate token quality from a noisy-label perspective and propose a generic token cleaning pipeline for SFT tasks. Our method filters out uninformative tokens while preserving those carrying key task-specific information. Specifically, we first evaluate token quality by examining the influence of model updates on each token, then apply a threshold-based separation. The token influence can be measured in a single pass with a fixed reference model or iteratively with self-evolving reference models. The benefits and limitations of both methods are analyzed theoretically by error upper bounds. Extensive experiments show that our framework consistently improves performance across multiple downstream tasks.

Large Vision-Language Models (LVLMs) exhibit impressive multimodal reasoning capabilities but remain highly susceptible to object hallucination, where models generate responses that are not factually aligned with the visual content. Recent works attribute this issue to an inherent bias of LVLMs where vision token attention map has a fixed correlation with spatial position, and propose to mitigate this issue by reordering visual tokens. However, we find that different LVLMs exhibit different correlations between attention and spatial position, which makes the existing solution difficult to generalize to other LVLMs. To address this issue, we first introduce a training-free solution, Uniform Attention Calibration (UAC), that estimates the bias from single meaningless input image and applies a calibration matrix to rectify attention imbalances. To further alleviate the bias, we relax the assumption of single meaningless input in UAC and introduce a fine-tuning solution, Dynamic Attention Calibration (DAC), that enforces the consistent outputs wherever the object locates in the image via a plug-and-plays module. Comprehensive experiments across multiple benchmarks demonstrate that UAC and DAC significantly reduce object hallucination while improving general multimodal alignment. Our methods achieve state-of-the-art performance across diverse LVLM architectures on various metrics.

Large language models (LLMs) are prone to hallucinations in question-answering (QA) tasks when faced with ambiguous questions. Users often assume that LLMs share their cognitive alignment, a mutual understanding of context, intent, and implicit details, leading them to omit critical information in the queries. However, LLMs generate responses based on assumptions that can misalign with user intent, which may be perceived as hallucinations if they misalign with the user's intent. Therefore, identifying those implicit assumptions is crucial to resolve ambiguities in QA. Prior work, such as AmbigQA, reduces ambiguity in queries via human-annotated clarifications, which is not feasible in real application. Meanwhile, ASQA compiles AmbigQA's short answers into long-form responses but inherits human biases and fails capture explicit logical distinctions that differentiates the answers. We introduce Conditional Ambiguous Question-Answering (CondAmbigQA), a benchmark with 200 ambiguous queries and condition-aware evaluation metrics. Our study pioneers the concept of ``conditions'' in ambiguous QA tasks, where conditions stand for contextual constraints or assumptions that resolve ambiguities. The retrieval-based annotation strategy uses retrieved Wikipedia fragments to identify possible interpretations for a given query as its conditions and annotate the answers through those conditions. Such a strategy minimizes human bias introduced by different knowledge levels among annotators. By fixing retrieval results, CondAmbigQA evaluates how RAG systems leverage conditions to resolve ambiguities. Experiments show that models considering conditions before answering improve performance by $20\%$, with an additional $5\%$ gain when conditions are explicitly provided. These results underscore the value of conditional reasoning in QA, offering researchers tools to rigorously evaluate ambiguity resolution.

Large Language Models (LLMs) have gained unprecedented prominence, achieving widespread adoption across diverse domains and integrating deeply into society. The capability to fine-tune general-purpose LLMs, such as Generative Pre-trained Transformers (GPT), for specific tasks has facilitated the emergence of numerous Custom GPTs. These tailored models are increasingly made available through dedicated marketplaces, such as OpenAI's GPT Store. However, their black-box nature introduces significant safety and compliance risks. In this work, we present a scalable framework for the automated evaluation of Custom GPTs against OpenAI's usage policies, which define the permissible behaviors of these systems. Our framework integrates three core components: (1) automated discovery and data collection of models from the GPT store, (2) a red-teaming prompt generator tailored to specific policy categories and the characteristics of each target GPT, and (3) an LLM-as-a-judge technique to analyze each prompt-response pair for potential policy violations. We validate our framework with a manually annotated ground truth, and evaluate it through a large-scale study with 782 Custom GPTs across three categories: Romantic, Cybersecurity, and Academic GPTs. Our manual annotation process achieved an F1 score of 0.975 in identifying policy violations, confirming the reliability of the framework's assessments. The results reveal that 58.7% of the analyzed models exhibit indications of non-compliance, exposing weaknesses in the GPT store's review and approval processes. Furthermore, our findings indicate that a model's popularity does not correlate with compliance, and non-compliance issues largely stem from behaviors inherited from base models rather than user-driven customizations. We believe this approach is extendable to other chatbot platforms and policy domains, improving LLM-based systems safety.

Large language models (LLMs) have achieved remarkable success in contextual knowledge understanding. In this paper, we show that these concentrated massive values consistently emerge in specific regions of attention queries (Q) and keys (K) while not having such patterns in values (V) in various modern transformer-based LLMs (Q, K, and V mean the representations output by the query, key, and value layers respectively). Through extensive experiments, we further demonstrate that these massive values play a critical role in interpreting contextual knowledge (i.e., knowledge obtained from the current context window) rather than in retrieving parametric knowledge stored within the model's parameters. Our further investigation of quantization strategies reveals that ignoring these massive values leads to a pronounced drop in performance on tasks requiring rich contextual understanding, aligning with our analysis. Finally, we trace the emergence of concentrated massive values and find that such concentration is caused by Rotary Positional Encoding (RoPE), which has appeared since the first layers. These findings shed new light on how Q and K operate in LLMs and offer practical insights for model design and optimization.

Large Language Models (LLMs) are becoming increasingly competent across various domains, educators are showing a growing interest in integrating these LLMs into the learning process. Especially in software engineering, LLMs have demonstrated qualitatively better capabilities in code summarization, code generation, and debugging. Despite various research on LLMs for software engineering tasks in practice, limited research captures the benefits of LLMs for pedagogical advancements and their impact on the student learning process. To this extent, we analyze 126 undergraduate students' interaction with an AI assistant during a 13-week semester to understand the benefits of AI for software engineering learning. We analyze the conversations, code generated, code utilized, and the human intervention levels to integrate the code into the code base. Our findings suggest that students prefer ChatGPT over CoPilot. Our analysis also finds that ChatGPT generates responses with lower computational complexity compared to CoPilot. Furthermore, conversational-based interaction helps improve the quality of the code generated compared to auto-generated code. Early adoption of LLMs in software engineering is crucial to remain competitive in the rapidly developing landscape. Hence, the next generation of software engineers must acquire the necessary skills to interact with AI to improve productivity.

Large language models (LLMs) have seen rapid development in recent years, revolutionizing various applications and significantly enhancing convenience and productivity. However, alongside their impressive capabilities, ethical concerns and new types of attacks, such as jailbreaking, have emerged. While most prompting techniques focus on optimizing adversarial inputs for individual cases, resulting in higher computational costs when dealing with large datasets. Less research has addressed the more general setting of training a universal attacker that can transfer to unseen tasks. In this paper, we introduce JUMP, a prompt-based method designed to jailbreak LLMs using universal multi-prompts. We also adapt our approach for defense, which we term DUMP. Experimental results demonstrate that our method for optimizing universal multi-prompts outperforms existing techniques.

Thematic Analysis (TA) is a fundamental method in healthcare research for analyzing transcript data, but it is resource-intensive and difficult to scale for large, complex datasets. This study investigates the potential of large language models (LLMs) to augment the inductive TA process in high-stakes healthcare settings. Focusing on interview transcripts from parents of children with Anomalous Aortic Origin of a Coronary Artery (AAOCA), a rare congenital heart disease, we propose an LLM-Enhanced Thematic Analysis (LLM-TA) pipeline. Our pipeline integrates an affordable state-of-the-art LLM (GPT-4o mini), LangChain, and prompt engineering with chunking techniques to analyze nine detailed transcripts following the inductive TA framework. We evaluate the LLM-generated themes against human-generated results using thematic similarity metrics, LLM-assisted assessments, and expert reviews. Results demonstrate that our pipeline outperforms existing LLM-assisted TA methods significantly. While the pipeline alone has not yet reached human-level quality in inductive TA, it shows great potential to improve scalability, efficiency, and accuracy while reducing analyst workload when working collaboratively with domain experts. We provide practical recommendations for incorporating LLMs into high-stakes TA workflows and emphasize the importance of close collaboration with domain experts to address challenges related to real-world applicability and dataset complexity. https://github.com/jiaweixu98/LLM-TA

Chinese literary classics hold significant cultural and educational value, offering deep insights into morality, history, and human nature. These works often include classical Chinese and complex narratives, making them difficult for children to read. To bridge this gap, we introduce a child-friendly literary adaptation (CLA) task to adapt the Chinese literary classic into engaging and accessible text for children. However, recent large language models (LLMs) overlook children's reading preferences (\ie, vivid character portrayals, concise narrative structures, and appropriate readability), which poses challenges in CLA. In this paper, we propose a method called InstructChild, which augments the LLM with these preferences for adaptation. Specifically, we first obtain the characters' personalities and narrative structure as additional information for fine-grained instruction tuning. Then, we devise a readability metric as the reward to align the LLM with the children's reading level. Finally, a lookahead decoding strategy is applied to improve the readability of the generated text during inference. To support the evaluation of CLA task, we construct the Classic4Children dataset, which comprises both the original and child-friendly versions of the Four Great Classical Novels of Chinese literature. Experimental results show that our InstructChild significantly improves automatic and human evaluation performance.