Michael Linghzi Li

Ailin Huang, Ang Li, Aobo Kong et al.

We introduce Step 3.5 Flash, a sparse Mixture-of-Experts (MoE) model that bridges frontier-level agentic intelligence and computational efficiency. We focus on what matters most when building agents: sharp reasoning and fast, reliable execution. Step 3.5 Flash pairs a 196B-parameter foundation with 11B active parameters for efficient inference. It is optimized with interleaved 3:1 sliding-window/full attention and Multi-Token Prediction (MTP-3) to reduce the latency and cost of multi-round agentic interactions. To reach frontier-level intelligence, we design a scalable reinforcement learning framework that combines verifiable signals with preference feedback, while remaining stable under large-scale off-policy training, enabling consistent self-improvement across mathematics, code, and tool use. Step 3.5 Flash demonstrates strong performance across agent, coding, and math tasks, achieving 85.4% on IMO-AnswerBench, 86.4% on LiveCodeBench-v6 (2024.08-2025.05), 88.2% on tau2-Bench, 69.0% on BrowseComp (with context management), and 51.0% on Terminal-Bench 2.0, comparable to frontier models such as GPT-5.2 xHigh and Gemini 3.0 Pro. By redefining the efficiency frontier, Step 3.5 Flash provides a high-density foundation for deploying sophisticated agents in real-world industrial environments.

Dimitris Bertsimas, Vassilis Digalakis, Michael Linghzi Li et al.

We present the framework of slowly varying regression under sparsity, allowing sparse regression models to exhibit slow and sparse variations. The problem of parameter estimation is formulated as a mixed-integer optimization problem. We demonstrate that it can be precisely reformulated as a binary convex optimization problem through a novel relaxation technique. This relaxation involves a new equality on Moore-Penrose inverses, convexifying the non-convex objective function while matching the original objective on all feasible binary points. This enables us to efficiently solve the problem to provable optimality using a cutting plane-type algorithm. We develop a highly optimized implementation of this algorithm, substantially improving upon the asymptotic computational complexity of a straightforward implementation. Additionally, we propose a fast heuristic method that guarantees a feasible solution and, as empirically illustrated, produces high-quality warm-start solutions for the binary optimization problem. To tune the framework's hyperparameters, we suggest a practical procedure relying on binary search that, under certain assumptions, is guaranteed to recover the true model parameters. On both synthetic and real-world datasets, we demonstrate that the resulting algorithm outperforms competing formulations in comparable times across various metrics, including estimation accuracy, predictive power, and computational time. The algorithm is highly scalable, allowing us to train models with thousands of parameters. Our implementation is available open-source at https://github.com/vvdigalakis/SSVRegression.git.