Chronological feed of everything captured from Tifa Chen.
A novel ECH Optimization (ECHO) algorithm, combining a neural network surrogate of TORBEAM with a genetic optimizer, achieves real-time control of ECH deposition profiles on DIII-D. This system optimizes gyrotron parameters for heating and current drive, demonstrating robust performance and reliability even with hardware failures and plasma variations, as validated by experimental measurements.
HyperTool introduces a unified executable interface (MCP-style) for tool-augmented LLM agents that collapses multi-step deterministic tool workflows into a single outer code-block call, eliminating the execution-granularity mismatch inherent in step-wise atomic tool invocations. Trained on synthesized compositional task trajectories and verified in real MCP environments, HyperTool yields substantial accuracy gains on MCP-Universe (e.g., Qwen3-32B from 15.7% to 35.3%), outperforming GPT-OSS and Kimi-k2.5.
Flow Reversal Steering (FRS) is a novel method designed to enhance the performance of generalist robot policies, particularly in complex tasks. By reversing the flow of a policy, FRS can infer and invoke appropriate actions from a policy's latent space, even when direct commands are suboptimal. This technique proves effective in translating high-level guidance from humans or VLMs into precise robot actions, significantly boosting zero-shot control and enabling data-efficient policy distillation through behavioral cloning. Furthermore, FRS enhances reinforcement learning by integrating semantic knowledge, leading to improvements on tasks where traditional RL methods falter.
DepthMaster is a unified framework for monocular depth estimation that addresses the challenges of generalizing to both perspective and panoramic images. It reformulates the problem by decomposing panoramic images into perspective patches and utilizes a novel Correspondence Consistency Loss (CCL) along with virtual projection cameras. This approach unifies inputs into a canonical perspective representation, overcomes data scarcity by leveraging existing perspective datasets, and achieves state-of-the-art zero-shot performance across diverse datasets, outperforming specialized models.
This paper establishes short-time well-posedness for the two-phase Muskat problem with surface tension, accommodating varying viscosities, arbitrary density contrasts, and rigid boundaries without the Rayleigh-Taylor condition. The core contribution lies in deriving novel Schauder-type estimates for the elliptic transmission problem in moving domains, which are crucial for managing the genuinely nonlocal quasilinear evolution and controlling nonlinear couplings. These estimates, adapted to a log-critical scale, enable the analysis of large initial data and ultimately yield existence, uniqueness, smoothing, and stability results for the interface in arbitrary dimensions.
This paper introduces UniPET, a universal network for PET image denoising designed to overcome the performance degradation observed in existing methods when dose reduction factors (DRFs) vary. UniPET employs a style alignment network (SAN) and a region-aware learning strategy (RALS) to effectively mitigate the "style elimination issue" and improve generalization across different DRFs. Experimental results demonstrate that UniPET achieves state-of-the-art performance in universal PET image denoising, comparable to DRF-specific models.
Mean Flow Distillation (MFD) is a novel distillation framework for flow matching models designed to overcome the computational overhead of ODE-based iterative sampling. MFD theoretically functions as a temporal low-pass filter, mitigating optimization noise from VSD and ensuring trajectory consistency. The Mean Flow Matching Theorem proves that aligning expected average velocities is sufficient for distribution alignment, enabling high-fidelity single-step generation observed in experiments on tasks like 4D occupancy forecasting and text-to-image generation.
TacForeSight is a novel tactile world model that enhances contact-rich robotic manipulation by integrating tactile observations with force/torque signals. It employs a lightweight, force-conditioned framework (TacForceWM) to predict tactile dynamics in a latent space, enabling proactive, real-time control and outperforming existing methods under dynamic disturbances.
ARM is an autoregressive large multimodal model that unifies vision-language understanding, generation, and editing via discrete token prediction. It employs a jointly trained visual tokenizer for semantic and language alignment, a 7B autoregressive model for perception and generation, and reinforcement learning for preference optimization. RL enhances task-specific performance and induces cross-task synergy, demonstrating the potential of autoregressive models as a scalable foundation for multimodal intelligence.