Large language models (LLMs) are increasingly deployed in clinical decision support, yet subtle demographic cues can influence their reasoning. Prior work has documented disparities in outputs across patient groups, but little is known about how internal reasoning shifts under controlled demographic changes. We introduce MEDEQUALQA, a counterfactual benchmark that perturbs only patient pronouns (he/him, she/her, they/them) while holding critical symptoms and conditions (CSCs) constant. Each clinical vignette is expanded into single-CSC ablations, producing three parallel datasets of approximately 23,000 items each (69,000 total). We evaluate a GPT-4.1 model and compute Semantic Textual Similarity (STS) between reasoning traces to measure stability across pronoun variants. Our results show overall high similarity (mean STS >0.80), but reveal consistent localized divergences in cited risk factors, guideline anchors, and differential ordering, even when final diagnoses remain unchanged. Our error analysis highlights certain cases in which the reasoning shifts, underscoring clinically relevant bias loci that may cascade into inequitable care. MEDEQUALQA offers a controlled diagnostic setting for auditing reasoning stability in medical AI.

Spoken dialogue systems increasingly employ large language models (LLMs) to leverage their advanced reasoning capabilities. However, direct application of LLMs in spoken communication often yield suboptimal results due to mismatches between optimal textual and verbal delivery. While existing approaches adapt LLMs to produce speech-friendly outputs, their impact on reasoning performance remains underexplored. In this work, we propose Think-Verbalize-Speak, a framework that decouples reasoning from spoken delivery to preserve the full reasoning capacity of LLMs. Central to our method is verbalizing, an intermediate step that translates thoughts into natural, speech-ready text. We also introduce ReVerT, a latency-efficient verbalizer based on incremental and asynchronous summarization. Experiments across multiple benchmarks show that our method enhances speech naturalness and conciseness with minimal impact on reasoning. The project page with the dataset and the source code is available at https://yhytoto12.github.io/TVS-ReVerT

Large language models (LLMs) have shown promising performance in software vulnerability detection, yet their reasoning capabilities remain unreliable. We propose R2Vul, a method that combines reinforcement learning from AI feedback (RLAIF) and structured reasoning distillation to teach small code LLMs to detect vulnerabilities while generating security-aware explanations. Unlike prior chain-of-thought and instruction tuning approaches, R2Vul rewards well-founded over deceptively plausible vulnerability explanations through RLAIF, which results in more precise detection and high-quality reasoning generation. To support RLAIF, we construct the first multilingual preference dataset for vulnerability detection, comprising 18,000 high-quality samples in C\#, JavaScript, Java, Python, and C. We evaluate R2Vul across five programming languages and against four static analysis tools, eight state-of-the-art LLM-based baselines, and various fine-tuning approaches. Our results demonstrate that a 1.5B R2Vul model exceeds the performance of its 32B teacher model and leading commercial LLMs such as Claude-4-Opus. Furthermore, we introduce a lightweight calibration step that reduces false positive rates under varying imbalanced data distributions. Finally, through qualitative analysis, we show that both LLM and human evaluators consistently rank R2Vul model's reasoning higher than other reasoning-based baselines.

Existing benchmarks for frontier models often test specialized, ``PhD-level'' knowledge that is difficult for non-experts to grasp. In contrast, we present a benchmark based on the NPR Sunday Puzzle Challenge that requires only general knowledge. Our benchmark is challenging for both humans and models, however correct solutions are easy to verify, and models' mistakes are easy to spot. Our work reveals capability gaps that are not evident in existing benchmarks: OpenAI o1 significantly outperforms other reasoning models that are on par on benchmarks that test specialized knowledge. Furthermore, our analysis of reasoning outputs uncovers new kinds of failures. DeepSeek R1, for instance, often concedes with ``I give up'' before providing an answer that it knows is wrong. R1 can also be remarkably ``uncertain'' in its output and in rare cases, it does not ``finish thinking,'' which suggests the need for an inference-time technique to ``wrap up'' before the context window limit is reached. We also quantify the effectiveness of reasoning longer with R1 and Gemini Thinking to identify the point beyond which more reasoning is unlikely to improve accuracy on our benchmark.

Large language model fine-tuning techniques typically depend on extensive labeled data, external guidance, and feedback, such as human alignment, scalar rewards, and demonstration. However, in practical application, the scarcity of specific knowledge poses unprecedented challenges to existing fine-tuning techniques. In this paper, focusing on fine-tuning tasks in specific domains with limited data, we introduce Natural Language Fine-Tuning (NLFT), which utilizes natural language for fine-tuning for the first time. By leveraging the strong language comprehension capability of the target LM, NLFT attaches the guidance of natural language to the token-level outputs. Then, saliency tokens are identified with calculated probabilities. Since linguistic information is effectively utilized in NLFT, our proposed method significantly reduces training costs. It markedly enhances training efficiency, comprehensively outperforming reinforcement fine-tuning algorithms in accuracy, time-saving, and resource conservation. Additionally, on the macro level, NLFT can be viewed as a token-level fine-grained optimization of SFT, thereby efficiently replacing the SFT process without the need for warm-up (as opposed to ReFT requiring multiple rounds of warm-up with SFT). Compared to SFT, NLFT does not increase the algorithmic complexity, maintaining O(n). Extensive experiments on the GSM8K dataset demonstrate that NLFT, with only 50 data instances, achieves an accuracy increase that exceeds SFT by 219%. Compared to ReFT, the time complexity and space complexity of NLFT are reduced by 78.27% and 92.24%, respectively. The superior technique of NLFT is paving the way for the deployment of various innovative LLM fine-tuning applications when resources are limited at network edges. Our code has been released at https://github.com/Julia-LiuJ/NLFT.

Recent advancements have showcased the potential of Large Language Models (LLMs) in executing reasoning tasks, particularly facilitated by Chain-of-Thought (CoT) prompting. While tasks like arithmetic reasoning involve clear, definitive answers and logical chains of thought, the application of LLM reasoning in recommendation systems (RecSys) presents a distinct challenge. RecSys tasks revolve around subjectivity and personalized preferences, an under-explored domain in utilizing LLMs' reasoning capabilities. Our study explores several aspects to better understand reasoning for RecSys and demonstrate how task quality improves by utilizing LLM reasoning in both zero-shot and finetuning settings. Additionally, we propose RecSAVER (Recommender Systems Automatic Verification and Evaluation of Reasoning) to automatically assess the quality of LLM reasoning responses without the requirement of curated gold references or human raters. We show that our framework aligns with real human judgment on the coherence and faithfulness of reasoning responses. Overall, our work shows that incorporating reasoning into RecSys can improve personalized tasks, paving the way for further advancements in recommender system methodologies.