Rob Miles

TRACERS Reveals IMF-Driven Reconnection Geometry as Primary Driver of Storm-Time Cusp Morphology

Using multi-crossing low-altitude ion and electron measurements from the TRACERS mission during the 30 September 2025 geomagnetic storm, this study traces cusp precipitation evolution across storm phases and links observed morphological changes directly to IMF-driven dayside reconnection geometry. The cusp expands poleward and broadens during storm-time, with enhanced electron energy flux, but the recovery-phase enhancement is attributed to sustained IMF solar wind driving rather than intrinsic storm dynamics. A key methodological finding is that accurate field-line transit distances from the reconnection X-line to the spacecraft — computed via the Tsyganenko model — are essential for reliable quantitative reconnection-rate estimates.

VLA Pretraining Produces Directly Executable Robot Behavior, Not Just Better Fine-Tuning Initialization

Wall-OSS-0.5 is an open-source 4B Vision-Language-Action model that challenges the prevailing assumption that VLA pretraining only matters as an initialization for downstream fine-tuning. Pretrained across 20+ embodiments on over one million robot trajectories per epoch, the model achieves non-trivial zero-shot real-robot behavior before any task-specific fine-tuning — including on a held-out deformable manipulation task. A gradient-bridged co-training recipe combining discrete action prediction, multimodal prediction, and continuous flow matching enables this capability while preserving vision-language grounding. Post fine-tuning, the model reaches 60.5% average task progress on 15 real-robot tasks, outperforming π0.5 by 17.5%.

Physics Foundation Model Generalizes from Simulation to Real Lab Turbulence — Zero-Shot

The Walrus foundation model for continuum dynamics, finetuned on as few as three DNS realizations of Rayleigh-Taylor instability (RTI), successfully transfers zero-shot to noisy laboratory experimental data it never encountered during training. The finetuned model exits the DNS-like regime and enters the experimentally observed mixing growth band, providing data-driven evidence that initial conditions — not fundamental physics gaps — are the primary driver of the long-standing ~3× discrepancy between simulated (α ≈ 0.02) and experimentally measured (α ≈ 0.06–0.07) late-time mixing growth rates. The model also generalizes zero-shot to stable stratification, a buoyancy regime entirely absent from training, correctly predicting reduced mixing-layer growth.

PEFT as Persistent Personal Model Infrastructure, Not Just a Fine-Tuning Budget Hack

This paper reframes Parameter-Efficient Fine-Tuning (PEFT) beyond its conventional role as a cheaper substitute for full fine-tuning, positioning small trainable adapters as persistent local state layered atop large shared foundation models. The architecture separates concerns cleanly: the base model supplies general competence while adapters encode instance-specific behavior — preferences, skills, tool habits, and memory-like updates. The authors organize the design space along three axes (Scale Up, Scale Down, Scale Out) and introduce MinT as a concrete infrastructure example for managing adapter lifecycle at scale, suggesting a path toward millions of coexisting personal models built on trillion-parameter foundations.

Logarithmic Time Grids + Implicit Stepping Slash MPS Imaginary Time Evolution Cost by Orders of Magnitude

This paper introduces a two-pronged acceleration of imaginary time evolution for matrix product state (MPS) wavefunctions: replacing uniform time grids with logarithmic grids (sufficient to resolve long-time dynamics) and using A-stable implicit ODE integrators that permit arbitrarily large time steps without instability. Together, these yield an exponential reduction in the number of required time steps versus the standard time-dependent variational principle (TDVP). The method relies only on matrix-vector products and linear solves — operations already native to the MPS toolbox — making it a drop-in upgrade for existing quantum many-body solvers.

Mars as an Exoplanet: Lessons from a Planet at the Edge of Habitability

Bayesian Inference of Mixing and Transmission Heterogeneity in Stratified Disease Surveillance Models

AcuityBench: LLMs Fail to Reliably Match Clinical Urgency Judgments, Especially Under Ambiguity

AcuityBench is a new benchmark from Columbia University that evaluates LLMs on clinical acuity identification — the ability to assign appropriate urgency levels to patient presentations — across 914 cases drawn from five harmonized public datasets. Tested across 12 frontier models, results reveal significant variance in accuracy and a systematic format-driven tradeoff: conversational responses reduce over-triage but increase under-triage compared to structured QA, a safety-critical asymmetry in high-acuity cases. Critically, no model closely replicates the distribution of physician uncertainty on ambiguous cases, with model predictions being more concentrated (less calibrated) than expert clinical judgment.

Graph-Informed Semiparametric Estimation Sharpens Causal Bounds Under Unmeasured Confounding

When causal effects cannot be point-identified due to unmeasured confounding, sensitivity analysis produces bounds rather than point estimates. This paper introduces an influence function projection method that leverages conditional independence constraints encoded in a known causal graph to improve the semiparametric efficiency of estimators for those bounds. The approach is framework-agnostic, applying across multiple sensitivity analysis models and causal estimands, effectively bridging graphical causal modeling with the partial-identification literature. Empirical validation covers both labor economics (job training on earnings) and clinical cardiology (ejection fraction on heart failure mortality) datasets.

Flow-Adaptive Ergodic Coverage Enables Formal Guarantees for Robot Exploration in Dynamic Environments

This paper reformulates adaptive robotic exploration as an ergodic coverage problem over time-varying, flow-induced domains — directly addressing the failure of static-environment assumptions in real-world settings like oceanography. The authors extend maximum mean discrepancy (MMD) as an ergodic metric into a flow-adaptive formulation that explicitly incorporates domain evolution, preserving coverage guarantees even in underactuated or open-loop planning regimes. Validation spans simulated spatiotemporal processes (ocean currents, human/cattle movement) and physical experiments on aerial and legged robots in non-convex, flow-restricted environments.