Robust fine-tuning aims to achieve competitive in-distribution (ID) performance while maintaining the out-of-distribution (OOD) robustness of a pre-trained model when transferring it to a downstream task. Recently, projected gradient descent has been successfully used in robust fine-tuning by constraining the deviation from the initialization of the fine-tuned model explicitly through projection. However, algorithmically, two limitations prevent this method from being adopted more widely, scalability and efficiency. In this paper, we propose a new projection-based fine-tuning algorithm, Fast Trainable Projection (FTP) for computationally efficient learning of per-layer projection constraints, resulting in an average 35% speedup on our benchmarks compared to prior works. FTP can be combined with existing optimizers such as AdamW, and be used in a plug-and-play fashion. Finally, we show that FTP is a special instance of hyper-optimizers that tune the hyper-parameters of optimizers in a learnable manner through nested differentiation. Empirically, we show superior robustness on OOD datasets, including domain shifts and natural corruptions, across four different vision tasks with five different pre-trained models. Additionally, we demonstrate that FTP is broadly applicable and beneficial to other learning scenarios such as low-label and continual learning settings thanks to its easy adaptability. The code will be available at https://github.com/GT-RIPL/FTP.git.

The Freeze-Tag Problem (FTP) involves activating a set of initially asleep robots as quickly as possible, starting from a single awake robot. Once activated, a robot can assist in waking up other robots. Each active robot moves at unit speed. The objective is to minimize the makespan, i.e., the time required to activate the last robot. A key performance measure is the wake-up ratio, defined as the maximum time needed to activate any number of robots in any primary positions. This work focuses on the geometric (Euclidean) version of FTP in $\mathbb{R}^d$ under the $\ell_p$ norm, where the initial distance between each asleep robot and the single active robot is at most 1. For $(\mathbb{R}^2, \ell_2)$, we improve the previous upper bound of 4.62 ([7], CCCG 2024) to 4.31. Note that it is known that 3.82 is a lower bound for the wake-up ratio. In $\mathbb{R}^3$, we propose a new strategy that achieves a wake-up ratio of 12 for $(\mathbb{R}^3, \ell_1)$ and 12.76 for $(\mathbb{R}^3, \ell_2)$, improving upon the previous bounds of 13 and $13\sqrt{3}$, respectively, reported in [2].

Large Language Models (LLMs) are increasingly evaluated on multiple-choice question answering (MCQA) tasks using *first-token probability* (FTP), which selects the answer option whose initial token has the highest likelihood. While efficient, FTP can be fragile: models may assign high probability to unrelated tokens (*misalignment*) or use a valid token merely as part of a generic preamble rather than as a clear answer choice (*misinterpretation*), undermining the reliability of symbolic evaluation. We propose a simple solution: the *prefilling attack*, a structured natural-language prefix (e.g., "*The correct option is:*") prepended to the model output. Originally explored in AI safety, we repurpose prefilling to steer the model to respond with a clean, valid option, without modifying its parameters. Empirically, the FTP with prefilling strategy substantially improves accuracy, calibration, and output consistency across a broad set of LLMs and MCQA benchmarks. It outperforms standard FTP and often matches the performance of open-ended generation approaches that require full decoding and external classifiers, while being significantly more efficient. Our findings suggest that prefilling is a simple, robust, and low-cost method to enhance the reliability of FTP-based evaluation in multiple-choice settings.

We study the Freeze-Tag Problem (FTP), introduced by Arkin et al. (SODA'02), where the objective is to activate a group of n robots, starting from a single initially active robot. Robots are positioned in $\mathbb{R}^d$, and once activated, they move at a constant speed to wake up others. The goal is to minimize the time required to activate the last robot, known as the makespan. We establish new upper bounds for the makespan under the $l_1$ and $l_2$ norms in $\mathbb{R}^2$ and $\mathbb{R}^3$. Specifically, we improve the previous upper bound for $(\mathbb{R}^2, l_2)$ from $7.07r$ (Bonichon et al., DISC'24) to $5.064r$. For $(\mathbb{R}^3, l_1)$, we derive a makespan bound of $13r$, which translates to $22.52r$ for $(\mathbb{R}^3, l_2)$. Here, $r$ denotes the maximum distance of any robot from the initially active robot under the given norm. To our knowledge, these are the first makespan bounds for FTP in $\mathbb{R}^3$. Additionally, we show that the maximum makespan for $n$ robots is not necessarily achieved when robots are equally distributed along the boundary in $(\mathbb{R}^2, l_2)$. We further investigate FTP in $(\mathbb{R}^3, l_2)$ for specific configurations where robots lie on a boundary, providing insights into practical scenarios.

Recently, large language models (LLMs) have demonstrated superior performance across various tasks by adhering to scaling laws, which significantly increase model size. However, the huge computation overhead during inference hinders the deployment in industrial applications. Many works leverage traditional compression approaches to boost model inference, but these always introduce additional training costs to restore the performance and the pruning results typically show noticeable performance drops compared to the original model when aiming for a specific level of acceleration. To address these issues, we propose a fine-grained token-wise pruning approach for the LLMs, which presents a learnable router to adaptively identify the less important tokens and skip them across model blocks to reduce computational cost during inference. To construct the router efficiently, we present a search-based sparsity scheduler for pruning sparsity allocation, a trainable router combined with our proposed four low-dimensional factors as input and three proposed losses. We conduct extensive experiments across different benchmarks on different LLMs to demonstrate the superiority of our method. Our approach achieves state-of-the-art (SOTA) pruning results, surpassing other existing pruning methods. For instance, our method outperforms BlockPruner and ShortGPT by approximately 10 points on both LLaMA2-7B and Qwen1.5-7B in accuracy retention at comparable token sparsity levels.

Robust fine-tuning aims to achieve competitive in-distribution (ID) performance while maintaining the out-of-distribution (OOD) robustness of a pre-trained model when transferring it to a downstream task. Recently, projected gradient descent has been successfully used in robust fine-tuning by constraining the deviation from the initialization of the fine-tuned model explicitly through projection. However, algorithmically, two limitations prevent this method from being adopted more widely, scalability and efficiency. In this paper, we propose a new projection-based fine-tuning algorithm, Fast Trainable Projection (FTP) for computationally efficient learning of per-layer projection constraints, resulting in an average 35% speedup on our benchmarks compared to prior works. FTP can be combined with existing optimizers such as AdamW, and be used in a plug-and-play fashion.

Estimating forest AGB at large scales and fine spatial resolutions has become increasingly important for greenhouse gas accounting, monitoring, and verification efforts to mitigate climate change. Airborne LiDAR is highly valuable for modeling attributes of forest structure including AGB, yet most LiDAR collections take place at local or regional scales covering irregular, non-contiguous footprints, resulting in a patchwork of different landscape segments at various points in time. Here, as part of a statewide forest carbon assessment for New York State (USA), we addressed common obstacles in leveraging a LiDAR patchwork for AGB mapping at landscape scales, including selection of training data, the investigation of regional or coverage specific patterns in prediction error, and map agreement with field inventory across multiple scales. Three machine learning algorithms and an ensemble model were trained with FIA field measurements, airborne LiDAR, and topographic, climatic and cadastral geodata. Using a strict set of plot selection criteria, 801 FIA plots were selected with co-located point clouds drawn from a patchwork of 17 leaf-off LiDAR coverages (2014-2019). Our ensemble model was used to produce 30 m AGB prediction surfaces within a predictor-defined area of applicability (98% of LiDAR coverage), and the resulting AGB maps were compared with FIA plot-level and areal estimates at multiple scales of aggregation. Our model was overall accurate (% RMSE 22-45%; MAE 11.6-29.4 Mg ha$^{-1}$; ME 2.4-6.3 Mg ha$^{-1}$), explained 73-80% of field-observed variation, and yielded estimates that were consistent with FIA's design-based estimates (89% of estimates within FIA's 95% CI). We share practical solutions to challenges faced in using spatiotemporal patchworks of LiDAR to meet growing needs for AGB mapping in support of applications in forest carbon accounting and ecosystem.

Temporal Pattern Mining (TPM) is the problem of mining predictive complex temporal patterns from multivariate time series in a supervised setting. We develop a new method called the Fast Temporal Pattern Mining with Extended Vertical Lists. This method utilizes an extension of the Apriori property which requires a more complex pattern to appear within records only at places where all of its subpatterns are detected as well. The approach is based on a novel data structure called the Extended Vertical List that tracks positions of the first state of the pattern inside records. Extensive computational results indicate that the new method performs significantly faster than the previous version of the algorithm for TMP. However, the speed-up comes at the expense of memory usage.