KeepKV: Eliminating Output Perturbation in KV Cache Compression for Efficient LLMs Inference

Efficient inference of large language models (LLMs) is hindered by an ever-growing key-value (KV) cache, making KV cache compression a critical research direction. Traditional methods selectively evict less important KV cache entries based on attention scores or position heuristics, which leads to information loss and hallucinations. Recently, merging-based strategies have been explored to retain more information by merging KV pairs that would be discarded; however, these existing approaches inevitably introduce inconsistencies in attention distributions before and after merging, causing output perturbation and degraded generation quality. To overcome this challenge, we propose KeepKV, a novel adaptive KV cache merging method designed to eliminate output perturbation while preserving performance under strict memory constraints. KeepKV introduces the Electoral Votes mechanism that records merging history and adaptively adjusts attention scores. Moreover, it further leverages a novel Zero Inference-Perturbation Merging methods, keeping attention consistency and compensating for attention loss resulting from cache merging. KeepKV successfully retains essential context information within a significantly compressed cache. Extensive experiments on various benchmarks and LLM architectures demonstrate that KeepKV substantially reduces memory usage, enhances inference throughput by more than 2x and keeps superior generation quality even with 10% KV cache budgets.

TinyR1-32B-Preview: Boosting Accuracy with Branch-Merge Distillation

The challenge of reducing the size of Large Language Models (LLMs) while maintaining their performance has gained significant attention. However, existing methods, such as model distillation and transfer learning, often fail to achieve high accuracy. To address this limitation, we introduce the Branch-Merge distillation approach, which enhances model compression through two phases: (1) the Branch Phase, where knowledge from a large teacher model is \textit{selectively distilled} into specialized student models via domain-specific supervised fine-tuning (SFT); And (2) the Merge Phase, where these student models are merged to enable cross-domain knowledge transfer and improve generalization. We validate our distillation approach using DeepSeek-R1 as the teacher and DeepSeek-R1-Distill-Qwen-32B as the student. The resulting merged model, TinyR1-32B-Preview, outperforms its counterpart DeepSeek-R1-Distill-Qwen-32B across multiple benchmarks, including Mathematics (+5.5 points), Coding (+4.4 points) and Science (+2.9 points), while achieving near-equal performance to DeepSeek-R1 on AIME 2024. The Branch-Merge distillation approach provides a scalable solution for creating smaller, high-performing LLMs with reduced computational cost and time.

ReFocus: Reinforcing Mid-Frequency and Key-Frequency Modeling for Multivariate Time Series Forecasting

Recent advancements have progressively incorporated frequency-based techniques into deep learning models, leading to notable improvements in accuracy and efficiency for time series analysis tasks. However, the Mid-Frequency Spectrum Gap in the real-world time series, where the energy is concentrated at the low-frequency region while the middle-frequency band is negligible, hinders the ability of existing deep learning models to extract the crucial frequency information. Additionally, the shared Key-Frequency in multivariate time series, where different time series share indistinguishable frequency patterns, is rarely exploited by existing literature. This work introduces a novel module, Adaptive Mid-Frequency Energy Optimizer, based on convolution and residual learning, to emphasize the significance of mid-frequency bands. We also propose an Energy-based Key-Frequency Picking Block to capture shared Key-Frequency, which achieves superior inter-series modeling performance with fewer parameters. A novel Key-Frequency Enhanced Training strategy is employed to further enhance Key-Frequency modeling, where spectral information from other channels is randomly introduced into each channel. Our approach advanced multivariate time series forecasting on the challenging Traffic, ECL, and Solar benchmarks, reducing MSE by 4%, 6%, and 5% compared to the previous SOTA iTransformer. Code is available at this GitHub Repository: https://github.com/Levi-Ackman/ReFocus.

FairKV: Balancing Per-Head KV Cache for Fast Multi-GPU Inference

KV cache techniques in Transformer models aim to reduce redundant computations at the expense of substantially increased memory usage, making KV cache compression an important and popular research topic. Recently, state-of-the-art KV cache compression methods implement imbalanced, per-head allocation algorithms that dynamically adjust the KV cache budget for each attention head, achieving excellent performance in single-GPU scenarios. However, we observe that such imbalanced compression leads to significant load imbalance when deploying multi-GPU inference, as some GPUs become overburdened while others remain underutilized. In this paper, we propose FairKV, a method designed to ensure fair memory usage among attention heads in systems employing imbalanced KV cache compression. The core technique of FairKV is Fair-Copying, which replicates a small subset of memory-intensive attention heads across GPUs using data parallelism to mitigate load imbalance. Our experiments on popular models, including LLaMA 70b and Mistral 24b model, demonstrate that FairKV increases throughput by 1.66x compared to standard tensor parallelism inference. Our code will be released as open source upon acceptance.

LLM-Sketch: Enhancing Network Sketches with LLM

Network stream mining is fundamental to many network operations. Sketches, as compact data structures that offer low memory overhead with bounded accuracy, have emerged as a promising solution for network stream mining. Recent studies attempt to optimize sketches using machine learning; however, these approaches face the challenges of lacking adaptivity to dynamic networks and incurring high training costs. In this paper, we propose LLM-Sketch, based on the insight that fields beyond the flow IDs in packet headers can also help infer flow sizes. By using a two-tier data structure and separately recording large and small flows, LLM-Sketch improves accuracy while minimizing memory usage. Furthermore, it leverages fine-tuned large language models (LLMs) to reliably estimate flow sizes. We evaluate LLM-Sketch on three representative tasks, and the results demonstrate that LLM-Sketch outperforms state-of-the-art methods by achieving a $7.5\times$ accuracy improvement.

Rank Also Matters: Hierarchical Configuration for Mixture of Adapter Experts in LLM Fine-Tuning

Large language models (LLMs) have demonstrated remarkable success across various tasks, accompanied by a continuous increase in their parameter size. Parameter-efficient fine-tuning (PEFT) methods, such as Low-Rank Adaptation (LoRA), address the challenges of fine-tuning LLMs by significantly reducing the number of trainable parameters. Recent studies have integrated LoRA with Mixture of Experts (MoE) architectures, leveraging multiple adapter experts and gating mechanisms to further improve fine-tuning performance. However, existing approaches primarily focus on adjusting the allocations of adapter experts per layer to optimize the introduced trainable parameter size, while neglecting a critical factor of adapters' rank. To this end, we propose a hierarchical scheme for expert allocation and rank configuration, HILO, which dynamically adjusts the number and rank of adapter experts across layers, matching the varying representational complexity of model layers in adapter-granularity. Extensive experiments on multiple benchmark tasks demonstrate that HILO outperforms existing methods in accuracy while introducing fewer trainable parameters, providing an efficient and practical solution for fine-tuning LLMs.

Predicting 3D representations for Dynamic Scenes

We present a novel framework for dynamic radiance field prediction given monocular video streams. Unlike previous methods that primarily focus on predicting future frames, our method goes a step further by generating explicit 3D representations of the dynamic scene. The framework builds on two core designs. First, we adopt an ego-centric unbounded triplane to explicitly represent the dynamic physical world. Second, we develop a 4D-aware transformer to aggregate features from monocular videos to update the triplane. Coupling these two designs enables us to train the proposed model with large-scale monocular videos in a self-supervised manner. Our model achieves top results in dynamic radiance field prediction on NVIDIA dynamic scenes, demonstrating its strong performance on 4D physical world modeling. Besides, our model shows a superior generalizability to unseen scenarios. Notably, we find that our approach emerges capabilities for geometry and semantic learning.

CFT-RAG: An Entity Tree Based Retrieval Augmented Generation Algorithm With Cuckoo Filter

Although retrieval-augmented generation(RAG) significantly improves generation quality by retrieving external knowledge bases and integrating generated content, it faces computational efficiency bottlenecks, particularly in knowledge retrieval tasks involving hierarchical structures for Tree-RAG. This paper proposes a Tree-RAG acceleration method based on the improved Cuckoo Filter, which optimizes entity localization during the retrieval process to achieve significant performance improvements. Tree-RAG effectively organizes entities through the introduction of a hierarchical tree structure, while the Cuckoo Filter serves as an efficient data structure that supports rapid membership queries and dynamic updates. The experiment results demonstrate that our method is much faster than naive Tree-RAG while maintaining high levels of generative quality. When the number of trees is large, our method is hundreds of times faster than naive Tree-RAG. Our work is available at https://github.com/TUPYP7180/CFT-RAG-2025.

Humanity's Last Exam

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 3,000 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Inference-to-complete: A High-performance and Programmable Data-plane Co-processor for Neural-network-driven Traffic Analysis

Neural-networks-driven intelligent data-plane (NN-driven IDP) is becoming an emerging topic for excellent accuracy and high performance. Meanwhile we argue that NN-driven IDP should satisfy three design goals: the flexibility to support various NNs models, the low-latency-high-throughput inference performance, and the data-plane-unawareness harming no performance and functionality. Unfortunately, existing work either over-modify NNs for IDP, or insert inline pipelined accelerators into the data-plane, failing to meet the flexibility and unawareness goals. In this paper, we propose Kaleidoscope, a flexible and high-performance co-processor located at the bypass of the data-plane. To address the challenge of meeting three design goals, three key techniques are presented. The programmable run-to-completion accelerators are developed for flexible inference. To further improve performance, we design a scalable inference engine which completes low-latency and low-cost inference for the mouse flows, and perform complex NNs with high-accuracy for the elephant flows. Finally, raw-bytes-based NNs are introduced, which help to achieve unawareness. We prototype Kaleidoscope on both FPGA and ASIC library. In evaluation on six NNs models, Kaleidoscope reaches 256-352 ns inference latency and 100 Gbps throughput with negligible influence on the data-plane. The on-board tested NNs perform state-of-the-art accuracy among other NN-driven IDP, exhibiting the the significant impact of flexibility on enhancing traffic analysis accuracy.