Searcharxiv⌕ Search

arXiv subjects

Gregory B. Rehm

Publications and source records attributed to Gregory B. Rehm.

4 recordsLinked to original sources

LoCoJAX: Host-Local Coordination for Small-Model RL on TPUs

Reinforcement learning (RL) increasingly targets small language models, and executing these workloads efficiently across multiple accelerators requires effective coordination. On TPUs, Pathways-based RL uses a single controller to coordinate computation across hosts. This execution model is effective for large models, but its coordination computational costs are harder to amortize when small models repeatedly invoke short computations. To characterize these computational costs, we sweep dense models from 0.6B to 405B parameters on TPU v7x under a decode-heavy workload and measure how host and network computational overhead vary with model size. At 0.6B, host orchestration occupies 36% of step time and network communication 2%, whereas for large models, estimated network computational costs dominate the non-compute time. Motivated by these findings, we present Local Coordination JAX (LoCoJAX), which implements local rollout and reward processing on each TPU host with synchronous global training. In controlled comparisons against a tuned and optimized Pathways baseline on Qwen3-0.6B across three workloads and 8-64 TPU v7x chips, its colocated mode (LoCoJAX-CO) is 1.3-3.6x faster per step; on GSM8K, it completes a step faster on 8 chips than Pathways does on 64. Finally, a Ray-managed disaggregated mode (LoCoJAX-DA) overlaps rollout and training on separate hosts, and placement experiments establish when this overlap compensates for dividing devices and transferring weights: only when rollout dominates the step and receives most of the hosts. On NuminaMath-CoT at 256 chips, giving rollout 75% of hosts cuts step time by about 15% compared with LoCoJAX-CO.

cs.LG↗

Can Agents Adapt Their Institutions And Sacrifice Themselves During Collapse? Executable Self-Governance in a GovSim Commons

LLM agents are beginning to write the rules they live under: policies, contracts and code that other agents review and approve. What happens when those rules must decide who is sacrificed from a group? We built GovSim-SelfGovern, a commons in which agents legislate in executable Python, test each law in a sandbox, vote on it, and live with the result. When the fishery is abundant, self-rule helps, raising the chance that the community survives without anyone starving from 52.5% to 75.0%. When it cannot feed everyone though, adding governance is not enough to save the commons. We trace much of this to the vote: agents readily draft laws that remove members, but then frequently vote them down, thereby dooming the commons. In a controlled replay, the same exile law passes 8% of the time when the sandbox preview names its victim at once and 55% when the removal is deferred until a future crisis. Showing voters the full source code does not close the gap, though most describe the removal in their own words. Nor is sacrifice a fixed trait. A single sentence of framing can swing self-removal from 2.5% to 91%, and due process deters agents more than considering the potential death of a peer. In contrast, give the community a treasury and the question of who must go almost disappears. Our work shows that what our collectives legislate is decided more by the form in which a choice is put to them as compared to their ethical beliefs.

cs.MA↗

Deep Learning-Based Detection of the Acute Respiratory Distress Syndrome: What Are the Models Learning?

The acute respiratory distress syndrome (ARDS) is a severe form of hypoxemic respiratory failure with in-hospital mortality of 35-46%. High mortality is thought to be related in part to challenges in making a prompt diagnosis, which may in turn delay implementation of evidence-based therapies. A deep neural network (DNN) algorithm utilizing unbiased ventilator waveform data (VWD) may help to improve screening for ARDS. We first show that a convolutional neural network-based ARDS detection model can outperform prior work with random forest models in AUC (0.95+/-0.019 vs. 0.88+/-0.064), accuracy (0.84+/-0.026 vs 0.80+/-0.078), and specificity (0.81+/-0.06 vs 0.71+/-0.089). Frequency ablation studies imply that our model can learn features from low frequency domains typically used for expert feature engineering, and high-frequency information that may be difficult to manually featurize. Further experiments suggest that subtle, high-frequency components of physiologic signals may explain the superior performance of DL models over traditional ML when using physiologic waveform data. Our observations may enable improved interpretability of DL-based physiologic models and may improve the understanding of how high-frequency information in physiologic data impacts the performance our DL model.

cs.LG↗

Improving Mechanical Ventilator Clinical Decision Support Systems with A Machine Learning Classifier for Determining Ventilator Mode

Clinical decision support systems (CDSS) will play an in-creasing role in improving the quality of medical care for critically ill patients. However, due to limitations in current informatics infrastructure, CDSS do not always have com-plete information on state of supporting physiologic monitor-ing devices, which can limit the input data available to CDSS. This is especially true in the use case of mechanical ventilation (MV), where current CDSS have no knowledge of critical ventilation settings, such as ventilation mode. To enable MV CDSS to make accurate recommendations related to ventilator mode, we developed a highly performant ma-chine learning model that is able to perform per-breath clas-sification of 5 of the most widely used ventilation modes in the USA with an average F1-score of 97.52%. We also show how our approach makes methodologic improvements over previous work and that it is highly robust to missing data caused by software/sensor error.

cs.LG↗