Michael Lin

Huihua Zhao, Rafael Cathomen, Lionel Gulich et al.

Recent advances in reinforcement learning (RL) have enabled impressive humanoid behaviors in simulation, yet transferring these results to new robots remains challenging. In many real deployments, the primary bottleneck is no longer simulation throughput or algorithm design, but the absence of systematic infrastructure that links environment verification, training, evaluation, and deployment in a coherent loop. To address this gap, we present AGILE, an end-to-end workflow for humanoid RL that standardizes the policy-development lifecycle to mitigate common sim-to-real failure modes. AGILE comprises four stages: (1) interactive environment verification, (2) reproducible training, (3) unified evaluation, and (4) descriptor-driven deployment via robot/task configuration descriptors. For evaluation stage, AGILE supports both scenario-based tests and randomized rollouts under a shared suite of motion-quality diagnostics, enabling automated regression testing and principled robustness assessment. AGILE also incorporates a set of training stabilizations and algorithmic enhancements in training stage to improve optimization stability and sim-to-real transfer. With this pipeline in place, we validate AGILE across five representative humanoid skills spanning locomotion, recovery, motion imitation, and loco-manipulation on two hardware platforms (Unitree G1 and Booster T1), achieving consistent sim-to-real transfer. Overall, AGILE shows that a standardized, end-to-end workflow can substantially improve the reliability and reproducibility of humanoid RL development.

Shyam Venkatasubramanian, Sean Moushegian, Michael Lin et al.

Standard attention-based transformers are known to exhibit instability under learning rate overspecification during training, particularly at high learning rates. While various methods have been proposed to improve resilience to such overspecification by modifying the optimization procedure, fundamental architectural innovations to this end remain underexplored. In this work, we illustrate that the consensus mechanism, a drop-in replacement for attention, stabilizes transformer training across a wider effective range of learning rates. We formulate consensus as a graphical model and provide extensive empirical analysis demonstrating improved stability across learning rate sweeps on text, DNA, and protein modalities. We further propose a hybrid consensus-attention framework that preserves performance while improving stability. We provide theoretical analysis characterizing the properties of consensus.

Boyu Gou, Zanming Huang, Yuting Ning et al. · microsoft-research

Agentic search such as Deep Research systems-where agents autonomously browse the web, synthesize information, and return comprehensive citation-backed answers-represents a major shift in how users interact with web-scale information. While promising greater efficiency and cognitive offloading, the growing complexity and open-endedness of agentic search have outpaced existing evaluation benchmarks and methodologies, which largely assume short search horizons and static answers. In this paper, we introduce Mind2Web 2, a benchmark of 130 realistic, high-quality, and long-horizon tasks that require real-time web browsing and extensive information synthesis, constructed with over 1000 hours of human labor. To address the challenge of evaluating time-varying and complex answers, we propose a novel Agent-as-a-Judge framework. Our method constructs task-specific judge agents based on a tree-structured rubric design to automatically assess both answer correctness and source attribution. We conduct a comprehensive evaluation of ten frontier agentic search systems and human performance, along with a detailed error analysis to draw insights for future development. The best-performing system, OpenAI Deep Research, can already achieve 50-70% of human performance while spending half the time, highlighting its great potential. Altogether, Mind2Web 2 provides a rigorous foundation for developing and benchmarking the next generation of agentic search systems.

Boyuan Zheng, Zeyi Liao, Scott Salisbury et al. · microsoft-research

The rapid development of autonomous web agents powered by Large Language Models (LLMs), while greatly elevating efficiency, exposes the frontier risk of taking unintended or harmful actions. This situation underscores an urgent need for effective safety measures, akin to access controls for human users. To address this critical challenge, we introduce WebGuard, the first comprehensive dataset designed to support the assessment of web agent action risks and facilitate the development of guardrails for real-world online environments. In doing so, WebGuard specifically focuses on predicting the outcome of state-changing actions and contains 4,939 human-annotated actions from 193 websites across 22 diverse domains, including often-overlooked long-tail websites. These actions are categorized using a novel three-tier risk schema: SAFE, LOW, and HIGH. The dataset includes designated training and test splits to support evaluation under diverse generalization settings. Our initial evaluations reveal a concerning deficiency: even frontier LLMs achieve less than 60% accuracy in predicting action outcomes and less than 60% recall in lagging HIGH-risk actions, highlighting the risks of deploying current-generation agents without dedicated safeguards. We therefore investigate fine-tuning specialized guardrail models using WebGuard. We conduct comprehensive evaluations across multiple generalization settings and find that a fine-tuned Qwen2.5VL-7B model yields a substantial improvement in performance, boosting accuracy from 37% to 80% and HIGH-risk action recall from 20% to 76%. Despite these improvements, the performance still falls short of the reliability required for high-stakes deployment, where guardrails must approach near-perfect accuracy and recall.

Srijith Nair, Michael Lin, Peizhong Ju et al.

Collaborative training methods like Federated Learning (FL) and Split Learning (SL) enable distributed machine learning without sharing raw data. However, FL assumes clients can train entire models, which is infeasible for large-scale models. In contrast, while SL alleviates the client memory constraint in FL by offloading most training to the server, it increases network latency due to its sequential nature. Other methods address the conundrum by using local loss functions for parallel client-side training to improve efficiency, but they lack server feedback and potentially suffer poor accuracy. We propose FSL-SAGE (Federated Split Learning via Smashed Activation Gradient Estimation), a new federated split learning algorithm that estimates server-side gradient feedback via auxiliary models. These auxiliary models periodically adapt to emulate server behavior on local datasets. We show that FSL-SAGE achieves a convergence rate of $\mathcal{O}(1/\sqrt{T})$, where $T$ is the number of communication rounds. This result matches FedAvg, while significantly reducing communication costs and client memory requirements. Our empirical results also verify that it outperforms existing state-of-the-art FSL methods, offering both communication efficiency and accuracy.

Baohua Sun, Michael Lin, Hao Sha et al.

Optical Character Recognition (OCR) has many real world applications. The existing methods normally detect where the characters are, and then recognize the character for each detected location. Thus the accuracy of characters recognition is impacted by the performance of characters detection. In this paper, we propose a method for recognizing characters without detecting the location of each character. This is done by converting the OCR task into an image captioning task. One advantage of the proposed method is that the labeled bounding boxes for the characters are not needed during training. The experimental results show the proposed method outperforms the existing methods on both the license plate recognition and the watermeter character recognition tasks. The proposed method is also deployed into a low-power (300mW) CNN accelerator chip connected to a Raspberry Pi 3 for on-device applications.

Baohua Sun, Lin Yang, Hao Sha et al.

Recent years NLP research has witnessed the record-breaking accuracy improvement by DNN models. However, power consumption is one of the practical concerns for deploying NLP systems. Most of the current state-of-the-art algorithms are implemented on GPUs, which is not power-efficient and the deployment cost is also very high. On the other hand, CNN Domain Specific Accelerator (CNN-DSA) has been in mass production providing low-power and low cost computation power. In this paper, we will implement the Super Characters method on the CNN-DSA. In addition, we modify the Super Characters method to utilize the multi-modal data, i.e. text plus tabular data in the CL-Aff sharedtask.

Baohua Sun, Lin Yang, Michael Lin et al.

Power-efficient CNN Domain Specific Accelerator (CNN-DSA) chips are currently available for wide use in mobile devices. These chips are mainly used in computer vision applications. However, the recent work of Super Characters method for text classification and sentiment analysis tasks using two-dimensional CNN models has also achieved state-of-the-art results through the method of transfer learning from vision to text. In this paper, we implemented the text classification and sentiment analysis applications on mobile devices using CNN-DSA chips. Compact network representations using one-bit and three-bits precision for coefficients and five-bits for activations are used in the CNN-DSA chip with power consumption less than 300mW. For edge devices under memory and compute constraints, the network is further compressed by approximating the external Fully Connected (FC) layers within the CNN-DSA chip. At the workshop, we have two system demonstrations for NLP tasks. The first demo classifies the input English Wikipedia sentence into one of the 14 ontologies. The second demo classifies the Chinese online-shopping review into positive or negative.

Baohua Sun, Lin Yang, Michael Lin et al.

Language and vision are processed as two different modal in current work for image captioning. However, recent work on Super Characters method shows the effectiveness of two-dimensional word embedding, which converts text classification problem into image classification problem. In this paper, we propose the SuperCaptioning method, which borrows the idea of two-dimensional word embedding from Super Characters method, and processes the information of language and vision together in one single CNN model. The experimental results on Flickr30k data shows the proposed method gives high quality image captions. An interactive demo is ready to show at the workshop.

Baohua Sun, Lin Yang, Michael Lin et al.

The recent work of Super Characters method using two-dimensional word embedding achieved state-of-the-art results in text classification tasks, showcasing the promise of this new approach. This paper borrows the idea of Super Characters method and two-dimensional embedding, and proposes a method of generating conversational response for open domain dialogues. The experimental results on a public dataset shows that the proposed SuperChat method generates high quality responses. An interactive demo is ready to show at the workshop.

Baohua Sun, Lin Yang, Wenhan Zhang et al.

Tabular data is the most commonly used form of data in industry. Gradient Boosting Trees, Support Vector Machine, Random Forest, and Logistic Regression are typically used for classification tasks on tabular data. DNN models using categorical embeddings are also applied in this task, but all attempts thus far have used one-dimensional embeddings. The recent work of Super Characters method using two-dimensional word embeddings achieved the state of art result in text classification tasks, showcasing the promise of this new approach. In this paper, we propose the SuperTML method, which borrows the idea of Super Characters method and two-dimensional embeddings to address the problem of classification on tabular data. For each input of tabular data, the features are first projected into two-dimensional embeddings like an image, and then this image is fed into fine-tuned two-dimensional CNN models for classification. Experimental results have shown that the proposed SuperTML method had achieved state-of-the-art results on both large and small datasets.

Baohua Sun, Lin Yang, Catherine Chi et al.

The Super Characters method addresses sentiment analysis problems by first converting the input text into images and then applying 2D-CNN models to classify the sentiment. It achieves state of the art performance on many benchmark datasets. However, it is not as straightforward to apply in Latin languages as in Asian languages. Because the 2D-CNN model is designed to recognize two-dimensional images, it is better if the inputs are in the form of glyphs. In this paper, we propose SEW (Squared English Word) method generating a squared glyph for each English word by drawing Super Characters images of each English word at the alphabet level, combining the squared glyph together into a whole Super Characters image at the sentence level, and then applying the CNN model to classify the sentiment within the sentence. We applied the SEW method to Wikipedia dataset and obtained a 2.1% accuracy gain compared to the original Super Characters method. For multi-modal data with both structured tabular data and unstructured natural language text, the modified SEW method integrates the data into a single image and classifies sentiment with one unified CNN model.