Phi-4-reasoning Technical Report

We introduce Phi-4-reasoning, a 14-billion parameter reasoning model that achieves strong performance on complex reasoning tasks. Trained via supervised fine-tuning of Phi-4 on carefully curated set of "teachable" prompts-selected for the right level of complexity and diversity-and reasoning demonstrations generated using o3-mini, Phi-4-reasoning generates detailed reasoning chains that effectively leverage inference-time compute. We further develop Phi-4-reasoning-plus, a variant enhanced through a short phase of outcome-based reinforcement learning that offers higher performance by generating longer reasoning traces. Across a wide range of reasoning tasks, both models outperform significantly larger open-weight models such as DeepSeek-R1-Distill-Llama-70B model and approach the performance levels of full DeepSeek-R1 model. Our comprehensive evaluations span benchmarks in math and scientific reasoning, coding, algorithmic problem solving, planning, and spatial understanding. Interestingly, we observe a non-trivial transfer of improvements to general-purpose benchmarks as well. In this report, we provide insights into our training data, our training methodologies, and our evaluations. We show that the benefit of careful data curation for supervised fine-tuning (SFT) extends to reasoning language models, and can be further amplified by reinforcement learning (RL). Finally, our evaluation points to opportunities for improving how we assess the performance and robustness of reasoning models.

Platonic Grounding for Efficient Multimodal Language Models

The hyperscaling of data and parameter count in Transformer-based models is yielding diminishing performance improvement, especially when weighed against training costs. Such plateauing indicates the importance of methods for more efficient finetuning and inference, while retaining similar performance. This is especially relevant for multimodal learning paradigms, where inference costs of processing multimodal tokens can determine the model's practical viability. At the same time, research on representations and mechanistic interpretability has improved our understanding of the inner workings of Transformer-based models; one such line of work reveals an implicit alignment in the deeper layers of pretrained models, across modalities. Taking inspiration from this, we motivate and propose a simple modification to existing multimodal frameworks that rely on aligning pretrained models. We demonstrate that our approach maintains and, in some cases, even improves performance of baseline methods while achieving significant gains in both training and inference-time compute. Our work also has implications for combining pretrained models into larger systems efficiently.

EmoAgent: Assessing and Safeguarding Human-AI Interaction for Mental Health Safety

The rise of LLM-driven AI characters raises safety concerns, particularly for vulnerable human users with psychological disorders. To address these risks, we propose EmoAgent, a multi-agent AI framework designed to evaluate and mitigate mental health hazards in human-AI interactions. EmoAgent comprises two components: EmoEval simulates virtual users, including those portraying mentally vulnerable individuals, to assess mental health changes before and after interactions with AI characters. It uses clinically proven psychological and psychiatric assessment tools (PHQ-9, PDI, PANSS) to evaluate mental risks induced by LLM. EmoGuard serves as an intermediary, monitoring users' mental status, predicting potential harm, and providing corrective feedback to mitigate risks. Experiments conducted in popular character-based chatbots show that emotionally engaging dialogues can lead to psychological deterioration in vulnerable users, with mental state deterioration in more than 34.4% of the simulations. EmoGuard significantly reduces these deterioration rates, underscoring its role in ensuring safer AI-human interactions. Our code is available at: https://github.com/1akaman/EmoAgent

Inference-Time Scaling for Complex Tasks: Where We Stand and What Lies Ahead

Inference-time scaling can enhance the reasoning capabilities of large language models (LLMs) on complex problems that benefit from step-by-step problem solving. Although lengthening generated scratchpads has proven effective for mathematical tasks, the broader impact of this approach on other tasks remains less clear. In this work, we investigate the benefits and limitations of scaling methods across nine state-of-the-art models and eight challenging tasks, including math and STEM reasoning, calendar planning, NP-hard problems, navigation, and spatial reasoning. We compare conventional models (e.g., GPT-4o) with models fine-tuned for inference-time scaling (e.g., o1) through evaluation protocols that involve repeated model calls, either independently or sequentially with feedback. These evaluations approximate lower and upper performance bounds and potential for future performance improvements for each model, whether through enhanced training or multi-model inference systems. Our extensive empirical analysis reveals that the advantages of inference-time scaling vary across tasks and diminish as problem complexity increases. In addition, simply using more tokens does not necessarily translate to higher accuracy in these challenging regimes. Results from multiple independent runs with conventional models using perfect verifiers show that, for some tasks, these models can achieve performance close to the average performance of today's most advanced reasoning models. However, for other tasks, a significant performance gap remains, even in very high scaling regimes. Encouragingly, all models demonstrate significant gains when inference is further scaled with perfect verifiers or strong feedback, suggesting ample potential for future improvements.

D3-Guard: Acoustic-based Drowsy Driving Detection Using Smartphones

Since the number of cars has grown rapidly in recent years, driving safety draws more and more public attention. Drowsy driving is one of the biggest threatens to driving safety. Therefore, a simple but robust system that can detect drowsy driving with commercial off-the-shelf devices (such as smartphones) is very necessary. With this motivation, we explore the feasibility of purely using acoustic sensors embedded in smartphones to detect drowsy driving. We first study characteristics of drowsy driving, and find some unique patterns of Doppler shift caused by three typical drowsy behaviors, i.e. nodding, yawning and operating steering wheel. We then validate our important findings through empirical analysis of the driving data collected from real driving environments. We further propose a real-time Drowsy Driving Detection system (D3-Guard) based on audio devices embedded in smartphones. In order to improve the performance of our system, we adopt an effective feature extraction method based on undersampling technique and FFT, and carefully design a high-accuracy detector based on LSTM networks for the early detection of drowsy driving. Through extensive experiments with 5 volunteer drivers in real driving environments, our system can distinguish drowsy driving actions with an average total accuracy of 93.31% in real-time. Over 80% drowsy driving actions can be detected within first 70% of action duration.

HearFit+: Personalized Fitness Monitoring via Audio Signals on Smart Speakers

Fitness can help to strengthen muscles, increase resistance to diseases, and improve body shape. Nowadays, a great number of people choose to exercise at home/office rather than at the gym due to lack of time. However, it is difficult for them to get good fitness effects without professional guidance. Motivated by this, we propose the first personalized fitness monitoring system, HearFit+, using smart speakers at home/office. We explore the feasibility of using acoustic sensing to monitor fitness. We design a fitness detection method based on Doppler shift and adopt the short time energy to segment fitness actions. Based on deep learning, HearFit+ can perform fitness classification and user identification at the same time. Combined with incremental learning, users can easily add new actions. We design 4 evaluation metrics (i.e., duration, intensity, continuity, and smoothness) to help users to improve fitness effects. Through extensive experiments including over 9,000 actions of 10 types of fitness from 12 volunteers, HearFit+ can achieve an average accuracy of 96.13% on fitness classification and 91% accuracy for user identification. All volunteers confirm that HearFit+ can help improve the fitness effect in various environments.

HearSmoking: Smoking Detection in Driving Environment via Acoustic Sensing on Smartphones

Driving safety has drawn much public attention in recent years due to the fast-growing number of cars. Smoking is one of the threats to driving safety but is often ignored by drivers. Existing works on smoking detection either work in contact manner or need additional devices. This motivates us to explore the practicability of using smartphones to detect smoking events in driving environment. In this paper, we propose a cigarette smoking detection system, named HearSmoking, which only uses acoustic sensors on smartphones to improve driving safety. After investigating typical smoking habits of drivers, including hand movement and chest fluctuation, we design an acoustic signal to be emitted by the speaker and received by the microphone. We calculate Relative Correlation Coefficient of received signals to obtain movement patterns of hands and chest. The processed data is sent into a trained Convolutional Neural Network for classification of hand movement. We also design a method to detect respiration at the same time. To improve system performance, we further analyse the periodicity of the composite smoking motion. Through extensive experiments in real driving environments, HearSmoking detects smoking events with an average total accuracy of 93.44 percent in real-time.

Temporal Consistency for LLM Reasoning Process Error Identification

Verification is crucial for effective mathematical reasoning. We present a new temporal consistency method where verifiers iteratively refine their judgments based on the previous assessment. Unlike one-round verification or multi-model debate approaches, our method leverages consistency in a sequence of self-reflection actions to improve verification accuracy. Empirical evaluations across diverse mathematical process error identification benchmarks (Mathcheck, ProcessBench, and PRM800K) show consistent performance improvements over baseline methods. When applied to the recent DeepSeek R1 distilled models, our method demonstrates strong performance, enabling 7B/8B distilled models to outperform all 70B/72B models and GPT-4o on ProcessBench. Notably, the distilled 14B model with our method achieves performance comparable to Deepseek-R1. Our codes are available at https://github.com/jcguo123/Temporal-Consistency

DeepSeek-R1 Outperforms Gemini 2.0 Pro, OpenAI o1, and o3-mini in Bilingual Complex Ophthalmology Reasoning

Purpose: To evaluate the accuracy and reasoning ability of DeepSeek-R1 and three other recently released large language models (LLMs) in bilingual complex ophthalmology cases. Methods: A total of 130 multiple-choice questions (MCQs) related to diagnosis (n = 39) and management (n = 91) were collected from the Chinese ophthalmology senior professional title examination and categorized into six topics. These MCQs were translated into English using DeepSeek-R1. The responses of DeepSeek-R1, Gemini 2.0 Pro, OpenAI o1 and o3-mini were generated under default configurations between February 15 and February 20, 2025. Accuracy was calculated as the proportion of correctly answered questions, with omissions and extra answers considered incorrect. Reasoning ability was evaluated through analyzing reasoning logic and the causes of reasoning error. Results: DeepSeek-R1 demonstrated the highest overall accuracy, achieving 0.862 in Chinese MCQs and 0.808 in English MCQs. Gemini 2.0 Pro, OpenAI o1, and OpenAI o3-mini attained accuracies of 0.715, 0.685, and 0.692 in Chinese MCQs (all P<0.001 compared with DeepSeek-R1), and 0.746 (P=0.115), 0.723 (P=0.027), and 0.577 (P<0.001) in English MCQs, respectively. DeepSeek-R1 achieved the highest accuracy across five topics in both Chinese and English MCQs. It also excelled in management questions conducted in Chinese (all P<0.05). Reasoning ability analysis showed that the four LLMs shared similar reasoning logic. Ignoring key positive history, ignoring key positive signs, misinterpretation medical data, and too aggressive were the most common causes of reasoning errors. Conclusion: DeepSeek-R1 demonstrated superior performance in bilingual complex ophthalmology reasoning tasks than three other state-of-the-art LLMs. While its clinical applicability remains challenging, it shows promise for supporting diagnosis and clinical decision-making.

Instance-Dependent Regret Bounds for Learning Two-Player Zero-Sum Games with Bandit Feedback

No-regret self-play learning dynamics have become one of the premier ways to solve large-scale games in practice. Accelerating their convergence via improving the regret of the players over the naive $O(\sqrt{T})$ bound after $T$ rounds has been extensively studied in recent years, but almost all studies assume access to exact gradient feedback. We address the question of whether acceleration is possible under bandit feedback only and provide an affirmative answer for two-player zero-sum normal-form games. Specifically, we show that if both players apply the Tsallis-INF algorithm of Zimmert and Seldin (2018, arXiv:1807.07623), then their regret is at most $O(c_1 \log T + \sqrt{c_2 T})$, where $c_1$ and $c_2$ are game-dependent constants that characterize the difficulty of learning -- $c_1$ resembles the complexity of learning a stochastic multi-armed bandit instance and depends inversely on some gap measures, while $c_2$ can be much smaller than the number of actions when the Nash equilibria have a small support or are close to the boundary. In particular, for the case when a pure strategy Nash equilibrium exists, $c_2$ becomes zero, leading to an optimal instance-dependent regret bound as we show. We additionally prove that in this case, our algorithm also enjoys last-iterate convergence and can identify the pure strategy Nash equilibrium with near-optimal sample complexity.