Preprints and Publications
Preprints
* indicates equal contribution; + indicates corresponding author.
-
Haifeng Huang, Biyin Xu, Yang Li+.
"KiToke: Distribution-Aware Token Compression for Video Large Language Models".
arXiv preprint arXiv:2604.03414, 2026
We introduced KiToke, a training-free method that reduces inference costs for video large language models by compressing redundant visual tokens while preserving diverse visual information. On LLaVA-OneVision, its distribution-aware sampling and temporal aggregation retained 98.5% of the original four-benchmark mean accuracy with 10% of visual tokens, while providing a 6.8× speedup in initial input processing (prefill).
-
Haifeng Huang, Biyin Xu, Chunsheng Xin, Yang Li+.
"FORTE: Adaptive Scoring and Exact Keyframe Selection for Long-Video Question Answering".
arXiv preprint arXiv:2610.00573, 2026
We introduced FORTE to improve long-video question answering under limited compute by adaptively locating and selecting question-relevant keyframes. It combines Gaussian-process relevance prediction with exact optimization of relevance and temporal coverage, achieving the highest mean accuracy among compared selectors across four benchmarks at every tested scoring budget.
-
Daniel Agyei Asante, Yang Li+.
"TopoCompress: Long Context Compression via Graph-Wired Semantic Trajectories".
arXiv preprint arXiv:2608.30811, 2026
We introduced TopoCompress to reduce the cost of long-context inference in large language models while preserving the evidence needed to answer complex questions. Using training-free relevance propagation over a graph of coherent text passages, it achieved answer quality comparable to the strongest tested baseline with one-quarter of the context budget across five question-answering tasks.
-
Daniel Agyei Asante, Ernie Chang, Yang Li+.
"Importance-Guided Basis Selection for Low-Rank Decomposition of Large Language Models".
arXiv preprint arXiv:2605.01627, 2026
We introduced Basis Selection with Importance (BSI) to compress large language models while preserving the components most important for target-task performance. Combining gradient-based sensitivity with an efficient curvature estimate, BSI ranks low-rank components by the expected loss from removing them and outperformed the compared methods on mathematical-reasoning benchmarks, especially under aggressive compression.
-
Mokarram Chowdhury, Ernie Chang, Yang Li+.
"The Safety Relay in Roleplay Jailbreaks: A Component-Resolved Causal Analysis of Harm Recognition and Refusal".
arXiv preprint arxiv:2608.30585, 2026
We used mechanistic interpretability to investigate how roleplay jailbreaks cause large language models to answer harmful requests they would normally refuse. Controlled prompt decomposition and causal interventions revealed that harm-related signals can persist while their influence on refusal weakens; reversing selected internal changes restored refusal on held-out requests.
Publications
* indicates equal contribution; + indicates corresponding author.
-
Md Mokarram Chowdhury*, Daniel Agyei Asante, Ernie Chang, Yang Li*+.
"IMPACT: Importance-Aware Activation Space Reconstruction".
Annual Meeting of the Association for Computational Linguistics (ACL), 2026
(selected for oral presentation, 504 papers out of 12148 submissions = 4.1%).
We introduced IMPACT, an activation-aware low-rank compression method that makes large language models more compact while preserving task accuracy. By combining activation structure with gradient-based importance, it achieved up to 55.4% greater model-size reduction than the compared baselines at comparable or better accuracy.
-
Daniel Agyei Asante, Md Mokarram Chowdhury, Yang Li+.
"Decomposed Trust: Privacy, Adversarial Robustness, Ethics, and Fairness in Low-Rank LLMs".
Findings of the Association for Computational Linguistics: ACL 2026.
We investigated whether low-rank compression preserves the safety of large language models, evaluating privacy, adversarial robustness, ethics, and fairness. The study found improved adversarial robustness but weaker fairness and conversational privacy in the reported tests, highlighting safety trade-offs that conventional accuracy benchmarks can overlook.
-
Ernie Chang*, Yang Li*, Patrick Huber, Vish Vogeti, David Kant, Yangyang Shi, Vikas Chandra.
"AutoMixer: Checkpoint Artifacts as Automatic Data Mixers".
Annual Meeting of the Association for Computational Linguistics (ACL), 2025.
We introduced AutoMixer to improve language-model pretraining by selecting data mixtures aligned with the reasoning capabilities a model needs to develop. It repurposes saved model checkpoints and first-order data-influence estimates to guide data mixing, improving performance in evaluations spanning eight reasoning benchmarks.
-
Yang Li*+, Daniel Agyei Asante*, Changsheng Zhao, Ernie Chang, Yangyang Shi, Vikas Chandra.
"Streamlining Language Models via Semantic Basis Analysis".
Transactions on Machine Learning Research (TMLR), 2025.
We introduced Basel, a task-aware low-rank compression framework that reduces language-model size while preserving the capabilities needed for a target application. By reweighting, pruning, and augmenting semantic bases in model weights, Basel achieved greater compression than the compared low-rank methods while maintaining comparable or better downstream accuracy.
-
Yang Li*+, Yuan Shangguan*, Yuhao Wang, Liangzhen Lai, Ernie Chang, Changsheng Zhao, Yangyang Shi, and Vikas Chandra. "Breaking Down Power Barriers in On-Device Streaming ASR: Insights and Solutions”. Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics (NAACL), Industry Track, 2025 (selected for oral presentation).
We developed model–system co-design principles to reduce the energy demands of on-device streaming speech recognition by targeting the components that dominate power consumption. Optimizations guided by invocation frequency and memory behavior reduced energy use by up to 47% at comparable recognition accuracy while improving real-time performance.
-
Yang Li, Di Wang, and José M. F. Moura.
"Forecasting Graph-Based Time-Dependent Data with Graph Sequence Attention".
ACM Transactions on Knowledge Discovery from Data (TKDD), 2025.
We introduced GSA-Forecaster to improve forecasting for interconnected time series, such as demand across a transportation network. Its graph sequence attention mechanism integrates temporal patterns, graph-structured spatial dependencies, and auxiliary information, outperforming the compared spatiotemporal forecasting models on real-world datasets.
-
Ernie Chang, Matteo Paltenghi, Yang Li, Pin-Jie Lin, Changsheng Zhao, Patrick Huber, Zechun Liu, Rastislav Rabatin, Yangyang Shi, Vikas Chandra.
"Scaling Parameter-Constrained Language Models with Quality Data".
Conference on Empirical Methods in Natural Language Processing (EMNLP), Industry Track, 2024.
We extended language-model scaling laws to account for data quality, helping predict performance when model size is constrained. Our formulation of effective training tokens, evaluated using more than 200 trained models, linked data quality, model size, and training volume to downstream reasoning accuracy.
-
Ernie Chang, Pin-Jie Lin, Yang Li, Changsheng Zhao, Daeil Kim, Rastislav Rabatin, Zechun Liu, Yangyang Shi, Vikas Chandra.
"Target-Aware Language Modeling via Granular Data Sampling".
Conference on Empirical Methods in Natural Language Processing (EMNLP), 2024.
We developed target-aware data selection to focus language-model pretraining on text that supports desired downstream capabilities. Models pretrained on roughly 1% of RefinedWeb, chosen through multi-granular n-gram importance sampling, matched full-dataset performance across eight benchmarks in the reported evaluations.
-
Maximilian Lam, Jeff Johnson, Wenjie Xiong, Kiwan Maeng, Udit Gupta, Yang Li, Liangzhen Lai, Illias Leontiadis, Minsoo Rhu, Hsien-Hsin Lee, Vijay Janapa Reddi, Gu-Yeon Wei, David Brooks, and Edward Suh.
"GPU-based Private Information Retrieval for On-Device Machine Learning Inference".
ACM International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS), 2024.
We enabled on-device machine learning models to access large server-side embedding tables without revealing which entries they request. GPU-accelerated private information retrieval, co-designed with downstream ML applications, delivered over 100× higher throughput than the CPU-based baseline while preserving model accuracy in the evaluated applications.
-
Yang Li*, Liangzhen Lai*, Yuan Shangguan, Forrest N. Iandola, Zhaoheng Ni, Ernie Chang, Yangyang Shi, and Vikas Chandra.
"Folding Attention: Memory and Power Optimization for On-Device Transformer-based Streaming Speech Recognition".
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024.
We introduced Folding Attention to reduce the memory footprint and power consumption of Transformer-based on-device streaming speech recognition. By restructuring attention and feedforward layers around lower-dimensional sub-tokens, it reduced model size by up to 24% and power consumption by up to 23% without sacrificing recognition accuracy.
-
Ernie Chang*, Pin-Jie Lin*, Yang Li, Sidd Srinivasan, Gael Le Lan, David Kant, Yangyang Shi, Forrest Iandola, and Vikas Chandra.
"In-Context Prompt Editing for Conditional Audio Generation".
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024.
We improved text-to-audio generation for real-world user prompts that differ from the descriptions seen during training. Our retrieval-based in-context prompt editing method rewrites requests using relevant training captions as examples, reducing the impact of distribution shift and improving generated audio quality in the reported evaluations.
-
Duc Le, Frank Seide, Yuhao Wang, Yang Li, Kjell Schubert, Ozlem Kalinli, and Michael L. Seltzer.
"Factorized Blank Thresholding for Improved Runtime Efficiency of Neural Transducers".
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2023.
We reduced the computation and power demands of on-device speech recognition by optimizing the neural transducer’s output layer. Factorized blank thresholding skips expensive token computations when no-output predictions are sufficiently confident, delivering 26–30% faster decoding and 43–53% lower power consumption without accuracy loss in the reported tests.
-
Tong Shen, Yang Li, and José M. F. Moura.
"Forecasting COVID-19 Dynamics: Clustering, Generalized Spatiotemporal Attention, and Impacts of Mobility and Geographic Proximity".
IEEE International Conference on Data Engineering (ICDE), 2023.
We introduced COVID-Forecaster to improve county-level epidemic forecasting by modeling how disease trends depend on local conditions and connections between regions. Combining data-driven clustering with generalized spatiotemporal attention, it incorporated mobility and geographic relationships and reduced weighted absolute percentage error by 13.5% across 3,141 U.S. counties relative to the compared baselines.
-
Yang Li and José M. F. Moura.
"Forecaster: A Graph Transformer for Forecasting Spatial and Time-Dependent Data".
European Conference on Artificial Intelligence (ECAI), 2020.
We introduced Forecaster, a graph Transformer that improves demand prediction by jointly modeling relationships across locations and long-range patterns over time. It learns a graph of spatial dependencies and uses that structure to build a sparse Transformer, outperforming the compared demand forecasting models while accounting for heterogeneous inputs and nonstationary patterns.
-
Yang Li, Charles R. Lefurgy, Karthick Rajamani, Malcolm S. Allen-Ware, Guillermo J. Silva, Daniel D. Heimsoth, Saugata Ghose, and Onur Mutlu.
"A Scalable Priority-Aware Approach to Managing Data Center Server Power".
IEEE International Symposium on High-Performance Computer Architecture (HPCA), 2019.
Technical report: "CapMaestro: Exploiting power redundancy, data center-wide priorities, and stranded power for boosting data center performance," in IBM Research Report, 2018.
We designed CapMaestro to support more servers on existing power infrastructure in cloud data centers while protecting high-priority workloads. Its scalable, priority-aware power-capping architecture reclaims unused capacity across redundant power feeds and was validated in a real cloud control plane and large-scale simulations.
-
Cong Xu, Karthick Rajamani, Alexandre Ferreira, Wesley Felter, Juan Rubio, and Yang Li.
"dCat: Dynamic Cache Management for Efficient, Performance-Sensitive Infrastructure-as-a-Service".
ACM European Conference on Computer Systems (EuroSys), 2018.
We developed dCat to protect cloud workloads from performance interference when applications share a processor’s last-level cache. Built on Intel Cache Allocation Technology, its dynamic cache partitioning provides isolation and reallocates unused capacity to workloads that benefit, improving multi-tenant performance without modifying applications.
-
Yang Li, Saugata Ghose, Jongmoo Choi, Jin Sun, Hui Wang, and Onur Mutlu. "Utility-based Hybrid Memory Management". IEEE International Conference on Cluster Computing (Cluster), 2017.
We introduced UH-MEM to improve application performance in hybrid memory systems by placing data where it delivers the greatest system-wide benefit. Its utility-based page-migration policy accounts for access frequency, locality, and memory-level parallelism, improving average performance by 14% over the strongest compared policy in the evaluated DRAM/non-volatile-memory configuration.
-
Yang Li, Di Wang, Saugata Ghose, Jie Liu, Sriram Govindan, Sean James, Eric Peterson, John Siegler, Rachata Ausavarungnirun, and Onur Mutlu. "SizeCap: Efficiently Handling Power Surges for Fuel Cell Powered Data Centers". IEEE International Symposium on High-Performance Computer Architecture (HPCA), 2016.
We designed SizeCap to reduce energy-storage requirements in data centers whose fuel cells respond slowly to sudden increases in power demand. Coordinating storage sizing with server power capping reduced required storage capacity by 50–85% in evaluations using Microsoft production traces, without violating service-level agreements in the tested scenarios.
-
Zuochang Ye, Tianshi Wang, and Yang Li. "Domain-Alternated Optimization for Passive Macromodeling". IEEE Transactions on Very Large Scale Integration Systems (TVLSI), 2015.
We developed an optimization framework that recovers accuracy lost during passivity enforcement, enabling more accurate models for stable circuit simulation. Domain-alternated optimization preserves passivity by construction and scales to problems with tens of thousands of variables.
-
Yang Li and David Z. Pan. "An Accurate Semi-Analytical Framework for Full-Chip TSV-induced Stress modeling". IEEE/ACM Design Automation Conference (DAC), 2013.
We developed a scalable method for predicting thermomechanical stress in 3D chips, helping designers assess reliability risks from vertical interconnects. Its semi-analytical model captures interactions between through-silicon vias, reducing average error from 36.8% to 14.3% with 1% additional runtime over linear superposition in one reported test.
-
Zuochang Ye, Bichen Wu, Song Han, and Yang Li. "Time-Domain Segmentation based Massively Parallel Simulation for ADCs". IEEE/ACM Design Automation Conference (DAC), 2013.
We accelerated transistor-level simulation of analog-to-digital converters, addressing a computational bottleneck in chip verification. Our time-domain segmentation method runs simulation intervals in parallel without modifying the underlying simulator, achieving 52–78× speedups on 100 CPU cores across the evaluated converter designs.
-
Tianshi Wang, Yang Li, and Zuochang Ye. "Robust Passive Macro-Model Generation with
Local Compensation". IEEE Transactions on Microwave Theory and Techniques (TMTT), 2012 (see "Corrections to ‘Robust Passive Macro-Model Generation with Local Compensation'" in TMTT’13 for authorship information).
We developed a robust passivity-enforcement method for circuit macromodels, enabling stable circuit simulation while limiting accuracy loss. The method uses local compensation to correct passivity violations without creating new ones, avoiding convergence problems in the enforcement process with only a modest increase in model size.
-
Zuochang Ye, Yang Li, Mingzhi Gao, and Zhiping Yu. "A Novel Framework for Passive Macromodeling". IEEE/ACM Design Automation Conference (DAC), 2011.
We introduced a framework for generating accurate circuit macromodels while maintaining the passivity needed for stable circuit simulation. By adding post-enforcement optimization to conventional model fitting and passivity enforcement, the framework recovers accuracy lost during enforcement without compromising passivity.
|