Abhishek Goswami

Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson et al. · microsoft-research

We introduce Phi-4-Mini and Phi-4-Multimodal, compact yet highly capable language and multimodal models. Phi-4-Mini is a 3.8-billion-parameter language model trained on high-quality web and synthetic data, significantly outperforming recent open-source models of similar size and matching the performance of models twice its size on math and coding tasks requiring complex reasoning. This achievement is driven by a carefully curated synthetic data recipe emphasizing high-quality math and coding datasets. Compared to its predecessor, Phi-3.5-Mini, Phi-4-Mini features an expanded vocabulary size of 200K tokens to better support multilingual applications, as well as group query attention for more efficient long-sequence generation. Phi-4-Multimodal is a multimodal model that integrates text, vision, and speech/audio input modalities into a single model. Its novel modality extension approach leverages LoRA adapters and modality-specific routers to allow multiple inference modes combining various modalities without interference. For example, it now ranks first in the OpenASR leaderboard to date, although the LoRA component of the speech/audio modality has just 460 million parameters. Phi-4-Multimodal supports scenarios involving (vision + language), (vision + speech), and (speech/audio) inputs, outperforming larger vision-language and speech-language models on a wide range of tasks. Additionally, we experiment to further train Phi-4-Mini to enhance its reasoning capabilities. Despite its compact 3.8-billion-parameter size, this experimental version achieves reasoning performance on par with or surpassing significantly larger models, including DeepSeek-R1-Distill-Qwen-7B and DeepSeek-R1-Distill-Llama-8B.

Marah Abdin, Jyoti Aneja, Hany Awadalla et al. · microsoft-research, stanford

We introduce phi-3-mini, a 3.8 billion parameter language model trained on 3.3 trillion tokens, whose overall performance, as measured by both academic benchmarks and internal testing, rivals that of models such as Mixtral 8x7B and GPT-3.5 (e.g., phi-3-mini achieves 69% on MMLU and 8.38 on MT-bench), despite being small enough to be deployed on a phone. Our training dataset is a scaled-up version of the one used for phi-2, composed of heavily filtered publicly available web data and synthetic data. The model is also further aligned for robustness, safety, and chat format. We also provide parameter-scaling results with a 7B, 14B models trained for 4.8T tokens, called phi-3-small, phi-3-medium, both significantly more capable than phi-3-mini (e.g., respectively 75%, 78% on MMLU, and 8.7, 8.9 on MT-bench). To enhance multilingual, multimodal, and long-context capabilities, we introduce three models in the phi-3.5 series: phi-3.5-mini, phi-3.5-MoE, and phi-3.5-Vision. The phi-3.5-MoE, a 16 x 3.8B MoE model with 6.6 billion active parameters, achieves superior performance in language reasoning, math, and code tasks compared to other open-source models of similar scale, such as Llama 3.1 and the Mixtral series, and on par with Gemini-1.5-Flash and GPT-4o-mini. Meanwhile, phi-3.5-Vision, a 4.2 billion parameter model derived from phi-3.5-mini, excels in reasoning tasks and is adept at handling both single-image and text prompts, as well as multi-image and text prompts.

Abhishek Goswami, Aru Ranjan Singh, Francesco Banterle et al.

The range of real-world scene luminance is larger than the capture capability of many digital camera sensors which leads to details being lost in captured images, most typically in bright regions. Inverse tone mapping attempts to boost these captured Standard Dynamic Range (SDR) images back to High Dynamic Range (HDR) by creating a mapping that linearizes the well exposed values from the SDR image, and provides a luminance boost to the clipped content. However, in most cases, the details in the clipped regions cannot be recovered or estimated. In this paper, we present a novel inverse tone mapping approach for mapping SDR images to HDR that generates lost details in clipped regions through a semantic-aware diffusion based inpainting approach. Our method proposes two major contributions - first, we propose to use a semantic graph to guide SDR diffusion based inpainting in masked regions in a saturated image. Second, drawing inspiration from traditional HDR imaging and bracketing methods, we propose a principled formulation to lift the SDR inpainted regions to HDR that is compatible with generative inpainting methods. Results show that our method demonstrates superior performance across different datasets on objective metrics, and subjective experiments show that the proposed method matches (and in most cases outperforms) state-of-art inverse tone mapping operators in terms of objective metrics and outperforms them for visual fidelity.

Abhishek Goswami

Autonomous LLM agents can issue thousands of API calls per hour without human oversight. OAuth 2.0 assumes deterministic clients, but in agentic settings stochastic reasoning, prompt injection, or multi-agent orchestration can silently expand privileges. We introduce Agentic JWT (A-JWT), a dual-faceted intent token that binds each agent's action to verifiable user intent and, optionally, to a specific workflow step. A-JWT carries an agent's identity as a one-way checksum hash derived from its prompt, tools and configuration, and a chained delegation assertion to prove which downstream agent may execute a given task, and per-agent proof-of-possession keys to prevent replay and in-process impersonation. We define a new authorization mechanism and add a lightweight client shim library that self-verifies code at run time, mints intent tokens, tracks workflow steps and derives keys, thus enabling secure agent identity and separation even within a single process. We illustrate a comprehensive threat model for agentic applications, implement a Python proof-of-concept and show functional blocking of scope-violating requests, replay, impersonation, and prompt-injection pathways with sub-millisecond overhead on commodity hardware. The design aligns with ongoing OAuth agent discussions and offers a drop-in path toward zero-trust guarantees for agentic applications. A comprehensive performance and security evaluation with experimental results will appear in our forthcoming journal publication

Jingchao Peng, Thomas Bashford-Rogers, Zhuang Shao et al.

Infrared (IR) imaging offers advantages in several fields due to its unique ability of capturing content in extreme light conditions. However, the demanding hardware requirements of high-resolution IR sensors limit its widespread application. As an alternative, visible light can be used to synthesize IR images but this causes a loss of fidelity in image details and introduces inconsistencies due to lack of contextual awareness of the scene. This stems from a combination of using visible light with a standard dynamic range, especially under extreme lighting, and a lack of contextual awareness can result in pseudo-thermal-crossover artifacts. This occurs when multiple objects with similar temperatures appear indistinguishable in the training data, further exacerbating the loss of fidelity. To solve this challenge, this paper proposes CapHDR2IR, a novel framework incorporating vision-language models using high dynamic range (HDR) images as inputs to generate IR images. HDR images capture a wider range of luminance variations, ensuring reliable IR image generation in different light conditions. Additionally, a dense caption branch integrates semantic understanding, resulting in more meaningful and discernible IR outputs. Extensive experiments on the HDRT dataset show that the proposed CapHDR2IR achieves state-of-the-art performance compared with existing general domain transfer methods and those tailored for visible-to-infrared image translation.

Morteza Ziyadi, Yuting Sun, Abhishek Goswami et al.

We present a novel approach to named entity recognition (NER) in the presence of scarce data that we call example-based NER. Our train-free few-shot learning approach takes inspiration from question-answering to identify entity spans in a new and unseen domain. In comparison with the current state-of-the-art, the proposed method performs significantly better, especially when using a low number of support examples.

Zeeshan Ahmed, Saeed Amizadeh, Mikhail Bilenko et al.

Machine Learning is transitioning from an art and science into a technology available to every developer. In the near future, every application on every platform will incorporate trained models to encode data-based decisions that would be impossible for developers to author. This presents a significant engineering challenge, since currently data science and modeling are largely decoupled from standard software development processes. This separation makes incorporating machine learning capabilities inside applications unnecessarily costly and difficult, and furthermore discourage developers from embracing ML in first place. In this paper we present ML .NET, a framework developed at Microsoft over the last decade in response to the challenge of making it easy to ship machine learning models in large software applications. We present its architecture, and illuminate the application demands that shaped it. Specifically, we introduce DataView, the core data abstraction of ML .NET which allows it to capture full predictive pipelines efficiently and consistently across training and inference lifecycles. We close the paper with a surprisingly favorable performance study of ML .NET compared to more recent entrants, and a discussion of some lessons learned.