February 2026

Hello, I wanted to ask how do you restrict the output vocabulary/ possible actions of VLAs. Specifically I am reading currently the research papers of RT-2 and OpenVLA. OpenVLA references RT-2 and RT-2 says nothing specifically, it just says in the fine-tuning phase: “Thus, to ensure that RT-2 outputs valid action tokens during decoding, we constrain its output vocabulary via only sampling valid action tokens when the model is prompted with a robot-action task …” So do you […]

Ver mais

Long before “engineering” became a formal career path, it was already part of Will Jiang’s daily life. Growing up in China, Jiang was the student teacher whom a classmate turned to when a laptop froze minutes before class. By high school, he was maintaining his school’s network infrastructure by diagnosing connectivity issues, keeping systems online, and supporting hundreds of users. What stayed with him wasn’t just the technical challenge, but the idea that thoughtfully built systems could quietly […]

Ver mais

Machine learning models often perform well during experimentation. Offline metrics improve, prototypes demonstrate potential, and early validation builds confidence across teams. In controlled environments, systems behave predictably and progress feels steady. The transition to production introduces a different set of pressures. Training jobs fail intermittently. Features arrive outside expected time windows. Historical data changes without notice. Deployments slow as teams hesitate, unsure of downstream consequences. What worked in isolation begins to strain under operational reality. The cause is […]

Ver mais

Designing aligned and robust rewards for open-ended generation remains a key barrier to RL post-training. Rubrics provide structured, interpretable supervision, but scaling rubric construction is difficult: expert rubrics are costly, prompted rubrics are often superficial or inconsistent, and fixed-pool discriminative rubrics can saturate and drift, enabling reward hacking. We present SibylSense, an inference-time learning approach that adapts a frozen rubric generator through a tunable memory bank of validated rubric items. Memory is updated via verifier-based item rewards measured […]

Ver mais

Safe Reinforcement Learning (RL) is crucial for achieving high performance while ensuring safety in real-world applications. However, the complex interplay of multiple uncertainty sources in real environments poses significant challenges for interpretable risk assessment and robust decision-making. To address these challenges, we propose Fuz-RL, a fuzzy measure-guided robust framework for safe RL. Specifically, our framework develops a novel fuzzy Bellman operator for estimating robust value functions using Choquet integrals. Theoretically, we prove that solving the Fuz-RL problem (in […]

Ver mais

How Blitz scaled their game coaching app with lower latency and leaner operations Blitz is a fast-growing startup that provides personalized coaching for games such as League of Legends, Valorant, and Fortnite. They aim to help gamers become League of Legends legends through real-time insights and post-match analysis. While players play, the app does quite a lot of work. It captures live match data, analyzes it quickly, and uses it for real-time game screen overlays plus personalized post-game coaching. […]

Ver mais

In modern applications such as ECG monitoring, neuroimaging, wearable sensing, and industrial equipment diagnostics, complex and continuously structured data are ubiquitous, presenting both challenges and opportunities for functional data analysis. However, existing methods face a critical trade-off: conventional functional models are limited by linearity, whereas deep learning approaches lack interpretable region selection for sparse effects. To bridge these gaps, we propose a sparse Bayesian functional deep neural network (sBayFDNN). It learns adaptive functional embeddings through a deep Bayesian […]

Ver mais

Mobility trajectories are essential for understanding urban dynamics and enhancing urban planning, yet access to such data is frequently hindered by privacy concerns. This research introduces a transformative framework for generating large-scale urban mobility trajectories, employing a novel application of a transformer-based model pre-trained and fine-tuned through a two-phase process. Initially, trajectory generation is conceptualized as an offline reinforcement learning (RL) problem, with a significant reduction in vocabulary space achieved during tokenization. The integration of Inverse Reinforcement Learning […]

Ver mais