Prompt caching is critical for reducing latency and cost in LLM inference--OpenAI and Anthropic report up to 50-90% cost savings through prompt reuse. Despite its widespread success, little is known about what constitutes an optimal prompt caching policy, particularly when optimizing tail latency--a metric of central importance to practitioners. The widely used Least Recently Used (LRU) policy can perform arbitrarily poor on this metric, as it is oblivious to the heterogeneity of conversation lengths. To address this gap, we propose Tail-Optimized LRU, a simple two-line modification that reallocates KV cache capacity to prioritize high-latency conversations by evicting cache entries that are unlikely to affect future turns. Though the implementation is simple, we prove its optimality under a natural stochastic model of conversation dynamics, providing the first theoretical justification for LRU in this setting--a result that may be of independent interest to the caching community. Experimentally, on real conversation data WildChat [Zhao et al., 2024], Tail-Optimized LRU achieves up to 27.5% reduction in P90 tail Time to First Token latency and 23.9% in P95 tail latency compared to LRU, along with up to 38.9% decrease in SLO violations of 200ms. We believe this provides a practical and theoretically grounded option for practitioners seeking to optimize tail latency in real-world LLM deployments.

Parameter-efficient fine-tuning (PEFT) methods have shown promise in adapting large language models, yet existing approaches exhibit counter-intuitive phenomena: integrating either matrix decomposition or mixture-of-experts (MoE) individually decreases performance across tasks, though decomposition improves results on specific domains despite reducing parameters, while MoE increases parameter count without corresponding decrease in training efficiency. Motivated by these observations and the modular nature of PT, we propose PT-MoE, a novel framework that integrates matrix decomposition with MoE routing for efficient PT. Evaluation results across 17 datasets demonstrate that PT-MoE achieves state-of-the-art performance in both question answering (QA) and mathematical problem solving tasks, improving F1 score by 1.49 points over PT and 2.13 points over LoRA in QA tasks, while improving mathematical accuracy by 10.75 points over PT and 0.44 points over LoRA, all while using 25% fewer parameters than LoRA. Our analysis reveals that while PT methods generally excel in QA tasks and LoRA-based methods in math datasets, the integration of matrix decomposition and MoE in PT-MoE yields complementary benefits: decomposition enables efficient parameter sharing across experts while MoE provides dynamic adaptation, collectively enabling PT-MoE to demonstrate cross-task consistency and generalization abilities. These findings, along with ablation studies on routing mechanisms and architectural components, provide insights for future PEFT methods. 1

Prompt tuning has emerged as a powerful parameter-efficient fine-tuning technique, allowing large pretrained Transformers to adapt to downstream tasks by optimizing a small set of prompt embeddings. Despite its empirical success, the extent to which prompt tuning can memorize data remains poorly understood. In this paper, we provide both theoretical and empirical analyses of data memorization ability of prompt-tuned Transformers. Building on recent theoretical frameworks, we derive an upper bound on the required prompt length for exact memorization of finite datasets and establish a trade-off between prompt length and the number of autoregressive generation steps. Specifically, we show that a constant-size Transformer can memorize ninput-output pairs with prompts of length O( nN), where N denotes the sequence length. Empirical results further demonstrate that prompt-tuned, randomly initialized Transformers are able to effectively memorize finite datasets. These models also capture the intrinsic low-rank structure of the data, leading to a reduction in the required prompt length. Finally, we analyze how the initialization of the Transformer backbone affects the performance of prompt tuning. Our findings provide new insights into the expressivity, efficiency, and underlying mechanisms of prompt tuning, bridging theoretical memorization limits with observed empirical behaviors.

Prompt tuning has emerged as a powerful parameter-efficient fine-tuning technique, allowing large pretrained Transformers to adapt to downstream tasks by optimizing a small set of prompt embeddings. Despite its empirical success, the extent to which prompt tuning can memorize data remains poorly understood. In this paper, we provide both theoretical and empirical analyses of data memorization ability of prompt-tuned Transformers. Building on recent theoretical frameworks, we derive an upper bound on the required prompt length for exact memorization of finite datasets and establish a trade-off between prompt length and the number of autoregressive generation steps. Specifically, we show that a constant-size Transformer can memorize $n$ input-output pairs with prompts of length $\tilde{O}(\sqrt{nN})$, where $N$ denotes the sequence length. Empirical results further demonstrate that prompt-tuned, randomly initialized Transformers are able to effectively memorize finite datasets. These models also capture the intrinsic low-rank structure of the data, leading to a reduction in the required prompt length. Finally, we analyze how the initialization of the Transformer backbone affects the performance of prompt tuning. Our findings provide new insights into the expressivity, efficiency, and underlying mechanisms of prompt tuning, bridging theoretical memorization limits with observed empirical behaviors.

Diffusion language models (DLMs) have emerged as a promising alternative to the long-dominant autoregressive (AR) paradigm, offering a parallelable decoding process that could yield greater efficiency. Yet, in practice, current open-source DLMs often underperform their AR counterparts in speed, limiting their real-world utility. This work presents a systematic study of DLM efficiency, identifying key issues in prior evaluation methods. Through empirical benchmarking and a theoretical analysis, we demonstrate that AR models generally achieve higher throughput, while DLMs consistently lag. We also investigate acceleration strategies, finding that techniques like dual cache and parallel decoding mainly offer gains at small batch sizes, with their benefits diminishing upon scaling. Our findings underscore the necessity of robust evaluation methods and improved acceleration strategies to advance research on DLMs.

While recent text-to-image (T2I) models show impressive capabilities in synthesizing images from brief descriptions, their performance significantly degrades when confronted with long, detail-intensive prompts required in professional applications. We present DetailMaster, the first comprehensive benchmark specifically designed to evaluate T2I models' systematic abilities to handle extended textual inputs that contain complex compositional requirements. Our benchmark introduces four critical evaluation dimensions: Character Attributes, Structured Character Locations, Multi-Dimensional Scene Attributes, and Spatial/Interactive Relationships. The benchmark comprises long and detail-rich prompts averaging 284.89 tokens, with high quality validated by expert annotators. Evaluation on 7 general-purpose and 5 long-prompt-optimized T2I models reveals critical performance limitations: state-of-the-art models achieve merely $\sim$50\% accuracy in key dimensions like attribute binding and spatial reasoning, while all models showing progressive performance degradation as prompt length increases. Our analysis reveals fundamental limitations in compositional reasoning, demonstrating that current encoders flatten complex grammatical structures and that diffusion models suffer from attribute leakage under detail-intensive conditions. We open-source our dataset, data curation code, and evaluation tools to advance detail-rich T2I generation and enable applications previously hindered by the lack of a dedicated benchmark.

Large-scale multilingual ASR models like Whisper excel in high-resource settings but face challenges in low-resource scenarios, such as rare languages and code-switching (CS), due to computational costs and catastrophic forgetting. We explore Soft Prompt Tuning (SPT), a parameter-efficient method to enhance CS ASR while preserving prior knowledge. We evaluate two strategies: (1) full fine-tuning (FFT) of both soft prompts and the entire Whisper model, demonstrating improved cross-lingual capabilities compared to traditional methods, and (2) adhering to SPT's original design by freezing model parameters and only training soft prompts. Additionally, we introduce SPT4ASR, a combination of different SPT variants. Experiments on the SEAME and ASRU2019 datasets show that deep prompt tuning is the most effective SPT approach, and our SPT4ASR methods achieve further error reductions in CS ASR, maintaining parameter efficiency similar to LoRA, without degrading performance on existing languages.

Serving disaggregated large language models has been widely adopted in industrial practice for enhanced performance. However, too many tokens generated in decoding phase, i.e., occupying the resources for a long time, essentially hamper the cloud from achieving a higher throughput. Meanwhile, due to limited on-device resources, the time to first token (TTFT), i.e., the latency of prefill phase, increases dramatically with the growth on prompt length. In order to concur with such a bottleneck on resources, i.e., long occupation in cloud and limited on-device computing capacity, we propose to separate large language model between cloud and devices. That is, the cloud helps a portion of the content for each device, only in its prefill phase. Specifically, after receiving the first token from the cloud, decoupling with its own prefill, the device responds to the user immediately for a lower TTFT. Then, the following tokens from cloud are presented via a speed controller for smoothed TPOT (the time per output token), until the device catches up with the progress. On-device prefill is then amortized using received tokens while the resource usage in cloud is controlled. Moreover, during cloud prefill, the prompt can be refined, using those intermediate data already generated, to further speed up on-device inference. We implement such a scheme P/D-Device, and confirm its superiority over other alternatives. We further propose an algorithm to decide the best settings. Real-trace experiments show that TTFT decreases at least 60%, maximum TPOT is about tens of milliseconds, and cloud throughput increases by up to 15x.

Large language models (LLMs) have been increasingly deployed as local agents on personal devices with CPUs, NPUs and integrated GPUs. However, forecasting inference performance on devices with such heterogeneity remains challenging due to the dynamic compute and memory demands. Existing approaches rely on GPU benchmarking or machine learning-based latency predictors, which are often hardware-specific and lack generalizability. To this end, we introduce LIFE, a lightweight and modular analytical framework that is comprised of modular analytical model of operators, configurable to characterize LLM inference workloads in a hardware and dataset-agnostic manner. LIFE characterizes the influence of software and model optimizations, such as quantization, KV cache compression, LoRA adapters, chunked prefill, different attentions, and operator fusion, on performance metrics such as time-to-first-token (TTFT), time-per-output-token (TPOT) and tokens-per-second (TPS). LIFE enables performance forecasting using only hardware specifications, such as TOPS and memory bandwidth, without requiring extensive dataset benchmarking. We validate LIFE's forecasting with inference on AMD Ryzen CPUs, NPUs, iGPUs and NVIDIA V100 GPUs, with Llama2-7B variants, demonstrating the utility of LIFE in forecasting LLM performance through lens of system efficiency to enable efficient LLM deployment across different hardware platforms.

Real-world fraud detection applications benefits from graph learning techniques that jointly exploit node features--often rich in textual data--and graph structural information. Recently, Graph-Enhanced LLMs emerge as a promising graph learning approach that converts graph information into prompts, exploiting LLMs' ability to reason over both textual and structural information. Among them, text-only prompting, which converts graph information to prompts consisting solely of text tokens, offers a solution that relies only on LLM tuning without requiring additional graph-specific encoders. However, text-only prompting struggles on heterogeneous fraud-detection graphs: multi-hop relations expand exponentially with each additional hop, leading to rapidly growing neighborhoods associated with dense textual information. These neighborhoods may overwhelm the model with long, irrelevant content in the prompt and suppress key signals from the target node, thereby degrading performance. To address this challenge, we propose Dual Granularity Prompting (DGP), which mitigates information overload by preserving fine-grained textual details for the target node while summarizing neighbor information into coarse-grained text prompts. DGP introduces tailored summarization strategies for different data modalities--bi-level semantic abstraction for textual fields and statistical aggregation for numerical features-- enabling effective compression of verbose neighbor content into concise, informative prompts. Experiments across public and industrial datasets demonstrate that DGP operates within a manageable token budget while improving fraud detection performance by up to 6.8% (AUPRC) over state-of-the-art methods, showing the potential of Graph-Enhanced LLMs for fraud detection.