Rumi Nakagawa

Rumi Nakagawa, Mary Chau, John Calzaretta et al.

Reliable estimates of Gross Primary Productivity (GPP), crucial for evaluating climate change initiatives, are currently only available from sparsely distributed eddy covariance tower sites. This limitation hampers access to reliable GPP quantification at regional to global scales. Prior machine learning studies on upscaling \textit{in situ} GPP to global wall-to-wall maps at sub-daily time steps faced limitations such as lack of input features at higher temporal resolutions and significant missing values. This research explored a novel upscaling solution using Temporal Fusion Transformer (TFT) without relying on past GPP time series. Model development was supplemented by Random Forest Regressor (RFR) and XGBoost, followed by the hybrid model of TFT and tree algorithms. The best preforming model yielded to model performance of 0.704 NSE and 3.54 RMSE. Another contribution of the study was the breakdown analysis of encoder feature importance based on time and flux tower sites. Such analysis enhanced the interpretability of the multi-head attention layer as well as the visual understanding of temporal dynamics of influential features.

Nitya Nadgir, Sayash Kapoor, Kangheng Liu et al.

When a benchmark's accuracy saturates, it is often retired and replaced with a more challenging version. We show that this approach privileges accuracy and misses the opportunity to study six other key dimensions of agent performance: construct validity issues such as shortcuts, out-of-distribution generalizability, efficiency, reliability, the relative importance of the model versus the scaffold, and uplift from human-agent collaboration. We use CORE-Bench Hard, a benchmark for computational reproducibility of scientific code, as a case study to demonstrate that measuring agents along these dimensions yields meaningful insights into agent performance even after accuracy saturates. First, we surface threats to construct validity in CORE-Bench Hard that are difficult to anticipate with less capable agents. We introduce an improved benchmark, CORE-Bench v1.1, and an out-of-distribution task suite, CORE-Bench OOD. Second, we find that despite accuracy saturation, CORE-Bench v1.1 remains useful for measuring efficiency, reliability, model performance, and scaffold performance. Finally, we conduct a small-scale randomized experiment to measure uplift from human-agent collaboration on real-world computational reproducibility tasks. We find a statistically significant speedup by about a factor of two -- likely underestimated due to one-fifth of human-only reproductions reaching the time limit before completing -- and describe various other findings. Together, our contributions present a more rigorous alternative to the dominant accuracy-centric evaluation paradigm.