Ryan Solgi

Ryan Solgi, Jiayi Tian, Zheng Zhang

Implicit chain-of-thought (iCoT) methods aim to internalize reasoning in large language models, but often underperform explicit CoT prompting. We empirically find that hidden-state reasoning trajectories exhibit low-rank structure. Motivated by this observation, we propose a low-rank distillation framework that transfers reasoning by aligning teacher and student trajectories in a shared low-rank tensor subspace using first- and second-order statistics. The resulting formulation captures the global structure of reasoning while supporting a compact latent reasoning process. We evaluate the method across multiple model families, including LLaMA and Qwen, at different scales on mathematical reasoning benchmarks. Our approach consistently improves performance, especially on challenging multi-step tasks, approaching explicit CoT accuracy and outperforming prior iCoT distillation methods.

Ryan Solgi, Zichang He, William Jiahua Liang et al.

Various tensor decomposition methods have been proposed for data compression. In real world applications of the tensor decomposition, selecting the tensor shape for the given data poses a challenge and the shape of the tensor may affect the error and the compression ratio. In this work, we study the effect of the tensor shape on the tensor decomposition and propose an optimization model to find an optimum shape for the tensor train (TT) decomposition. The proposed optimization model maximizes the compression ratio of the TT decomposition given an error bound. We implement a genetic algorithm (GA) linked with the TT-SVD algorithm to solve the optimization model. We apply the proposed method for the compression of RGB images. The results demonstrate the effectiveness of the proposed evolutionary tensor shape search for the TT decomposition.

Jiayi Tian, Yupeng Su, Ryan Solgi et al.

Large reasoning models (LRMs) enhance problem-solving capabilities by generating explicit multi-step chains of thought (CoT) reasoning; however, they incur substantial inference latency and computational overhead. To mitigate this issue, recent works have explored model collaboration paradigms, where small reasoning models (SRMs) generate intermediate reasoning steps to achieve a better accuracy--latency trade-off. Despite recent progress, effectively and efficiently detecting and mitigating SRM failures in collaborative systems remains a key challenge. To address this issue, we analyze SRM inference in both the generated text and hidden-state spaces, and identify three types of failure modes: \textit{overconfidence}, \textit{uncertainty}, and \textit{heavy revalidation}. Building on these insights, we propose \textbf{RankGuide}, a framework that improves the efficiency and effectiveness of SRM--LRM collaboration through tensor-rank-guided routing and steering. Specifically, RankGuide leverages a routing signal that incorporates tensor-rank signals derived from consecutive hidden states to detect when SRMs are likely to fail and selectively invoke LRMs. In addition, we introduce a tensor-rank-filtered steering vector extraction method to modulate the reasoning trajectory of SRMs, thereby improving their generation quality. By improving both routing and steering through tensor-rank signals, RankGuide enables SRM--LRM collaborative systems to achieve more efficient reasoning with fewer steps and improved accuracy. Experiments on multiple reasoning benchmarks demonstrate the efficacy of RankGuide in reducing latency by up to $1.75\times$ compared to LRM, while maintaining competitive accuracy relative to prior methods.

Parsa Madinei, Ryan Solgi, Ziqi Wen et al.

We introduce INTERLACE, a novel framework that prunes redundant layers in VLMs while maintaining performance through sample-efficient finetuning. Existing layer pruning methods lead to significant performance drop when applied to VLMs. Instead, we analyze triplets of consecutive layers to identify local redundancy, removing the most redundant of the first two layers, finetune the remaining layer to compensate for the lost capacity, and freeze the third layer to serve as a stable anchor during finetuning. We found that this interleaved finetune-freeze design enables rapid convergence with minimal data after pruning. By finetuning only a subset of layers on just 1% of the FineVision dataset for one epoch, Interlace achieves 88.9% average performance retention after dropping 25% of the network, achieving SOTA performance. Our code is available at: https://github.com/pmadinei/Interlace.git

Subhadra Vadlamannati, Ryan Solgi

The transformer architecture has revolutionized Natural Language Processing (NLP) and other machine-learning tasks, due to its unprecedented accuracy. However, their extensive memory and parameter requirements often hinder their practical applications. In this work, we study the effect of tensor-train decomposition to improve the accuracy and compress transformer vision-language neural networks, namely BERT and ViT. We focus both on embedding-layer compression and partial tensorization of neural networks (PTNN) through an algorithmic approach. Our novel PTNN approach significantly improves the accuracy of existing models by up to 5%, all without the need for post-training adjustments, breaking new ground in the field of tensor decomposition.

Jiayi Tian, Ryan Solgi, Jinming Lu et al.

Large Language Models (LLMs) have enabled remarkable progress in natural language processing, yet their high computational and memory demands pose challenges for deployment in resource-constrained environments. Although recent low-rank decomposition methods offer a promising path for structural compression, they often suffer from accuracy degradation, expensive calibration procedures, and result in inefficient model architectures that hinder real-world inference speedups. In this paper, we propose FLAT-LLM, a fast and accurate, training-free structural compression method based on fine-grained low-rank transformations in the activation space. Specifically, we reduce the hidden dimension by transforming the weights using truncated eigenvectors computed via head-wise Principal Component Analysis, and employ a greedy budget redistribution strategy to adaptively allocate ranks across decoders. FLAT-LLM achieves efficient and effective weight compression without recovery fine-tuning, which could complete the calibration within a few minutes. Evaluated across 5 models and 11 datasets, FLAT-LLM outperforms structural pruning baselines in generalization and downstream performance, while delivering inference speedups over decomposition-based methods.

Ryan Solgi, Parsa Madinei, Jiayi Tian et al.

Large language models (LLM) and vision-language models (VLM) have achieved state-of-the-art performance, but they impose significant memory and computing challenges in deployment. We present a novel low-rank compression framework to address this challenge. First, we upper bound the change of network loss via layer-wise activation-based compression errors, filling a theoretical gap in the literature. We then formulate low-rank model compression as a bi-objective optimization and prove that a single uniform tolerance yields surrogate Pareto-optimal heterogeneous ranks. Based on our theoretical insights, we propose Pareto-Guided Singular Value Decomposition (PGSVD), a zero-shot pipeline that improves activation-aware compression via Pareto-guided rank selection and alternating least-squares implementation. We apply PGSVD to both LLM and VLM, showing better accuracy at the same compression levels and inference speedup.

Ryan Solgi, Seyedali Mousavinezhad, Hugo A. Loaiciga

In this study, we investigate for the first time the low-rank properties of a tensorized large-scale spatio-temporal dynamic atmospheric variable. We focus on the Sentinel-5P tropospheric NO2 product (S5P-TN) over a four-year period in an area that encompasses the contiguous United States (CONUS). Here, it is demonstrated that a low-rank approximation of such a dynamic variable is feasible. We apply the low-rank properties of the S5P-TN data to inpaint gaps in the Sentinel-5P product by adopting a low-rank tensor model (LRTM) based on the CANDECOMP / PARAFAC (CP) decomposition and alternating least squares (ALS). Furthermore, we evaluate the LRTM's results by comparing them with spatial interpolation using geostatistics, and conduct a comprehensive spatial statistical and temporal analysis of the S5P-TN product. The results of this study demonstrated that the tensor completion successfully reconstructs the missing values in the S5P-TN product, particularly in the presence of extended cloud obscuration, predicting outliers and identifying hotspots, when the data is tensorized over extended spatial and temporal scales.

Ryan Solgi, Kai Zhen, Rupak Vignesh Swaminathan et al.

The efficient implementation of large language models (LLMs) is crucial for deployment on resource-constrained devices. Low-rank tensor compression techniques, such as tensor-train (TT) networks, have been widely studied for over-parameterized neural networks. However, their applications to compress pre-trained large language models (LLMs) for downstream tasks (post-training) remains challenging due to the high-rank nature of pre-trained LLMs and the lack of access to pretraining data. In this study, we investigate low-rank tensorized LLMs during fine-tuning and propose sparse augmented tensor networks (Saten) to enhance their performance. The proposed Saten framework enables full model compression. Experimental results demonstrate that Saten enhances both accuracy and compression efficiency in tensorized language models, achieving state-of-the-art performance.