Searcharxiv⌕ Search

arXiv subjects

Kwang-Ting Cheng

Publications and source records attributed to Kwang-Ting Cheng.

At least 19 recordsLinked to original sources
arXiv2026

FlagRot: Selective Rotation for Low-Bit KV Caches

Long-context inference is bound by the KV cache, which 2-bit codes could shrink about sixfold if their error stays in check. The usual remedy rotates keys before quantizing, assuming rotation always helps. However, we observe that it does not: rotation flattens a spiky piece of a key but can give a flat one a new peak, and both coexist within one key. We therefore propose FlagRot, which stores each 32-channel piece rotated or as is, whichever has the smaller calibrated attention error, and rotates the query instead when reading. We further show that how often to decide follows the quantization axis: per-channel caches such as KIVI need one decision per head, at no write cost, while per-token caches decide per key, with the choice stored in existing sign bits. At 2.5 bits per value, FlagRot lowers the twelve-model KL divergence of KIVI from 18.7 to 12.4 x 10^-3, below RotateKV on 9 of 12 models. Served as a vLLM plugin, it raises RULER at 128k tokens by +2.9 points and the LongBench v1 mean by +0.4 to +1.9 points on the three of four models where 2-bit KIVI loses accuracy. It also lowers the error of 4-bit integer and FP4 caches.

cs.CL↗
arXiv2026

PQ-HSA: Reusing Product-Quantized Scores for Hybrid Sparse-Approximate Attention

At each decoding step a language model attends over the key-value (KV) cache of every earlier token, so at long context the attention call is bounded by memory bandwidth. Sparse attention reads only a subset of keys chosen by a cheap score estimate, and most methods give the unread tokens zero weight. The output then draws on only a small fraction of the KV cache, and accuracy drops at small budgets, most on tasks that aggregate information across the context. An inverted-file product-quantization (IVF-PQ) index over the cached keys computes an approximate score for every indexed token in order to rank them; after ranking, those scores approximate the attention logits of the tokens left out. PQ-HSA (hybrid sparse-approximate attention) attends the selected tokens with their original keys and values, and the unselected tokens, the background, enter the same softmax through those scores, summed per inverted list and multiplied by the list's mean value. At 128K and a 1-2% retrieval budget, PQ-HSA is more accurate than Quest and SnapKV on Llama-3.1-8B and Qwen3-30B-A3B and stays close to full attention in macro accuracy; with the same selector, the background term raises macro accuracy on the 8B model from 0.71 to 0.83. In the same 128K setting, inside vLLM on one NVIDIA H20, the decode attention call runs 1.6x faster than the FlashAttention-3 kernel; the speedup grows with context length, and a cost model fitted on 8B to 30B models gives the context length at which it begins. A vLLM plugin runs PQ-HSA on two engine versions without changes to the engine source; code is available at https://github.com/KunmingSHAO/pqhsa_release.

cs.LG↗
arXiv2026

EfficientAgent: What Makes KV Cache Offloading Work for Concurrent Agents?

LLM agents resend their whole conversation on every turn, and most of it was already processed on the previous turn. Serving systems avoid recomputing it by caching its key-value (KV) state and, when GPU memory runs out, by offloading that state to host memory. For agents, offloading gives inconsistent results: on the same coding-agent workload it speeds up one deployment, slows down another, and changes nothing on a third, even where loading a token back is several times cheaper than recomputing it. The reason is that cached state must survive until it is used again. While one agent waits for its tool, the server processes the contexts of all other agents, so an agent's prefix is reused only if the host tier holds the reusable context of the whole agent pool, which we call the reuse working set. A smaller tier keeps writing state that is evicted before anyone reads it. We present EfficientAgent, which sizes and manages the host tier by this working set. A stack-distance model estimates the working set from agent histories to size the host tier; its predictions, made before the experiments, located the capacity at which offloading starts to pay. When the tier is too small, a runtime policy stops writing large refills of evicted context and keeps extending prefixes that are still cached; when the tier is large enough, it writes everything. On SWE-bench Verified coding agents, a host tier sized to the estimated working set cuts recomputed prompt tokens by 93% and end-to-end time by 39%. With a small fixed tier, the policy cuts recomputation by 35%; with a large tier, it avoids the 4.3-fold increase caused by always filtering writes. Across three GPU types and two models, offloading pays off when the GPU has little compute per byte of host bandwidth and the host tier holds the working set. Code is available at https://github.com/KunmingSHAO/efficientagent_release.

cs.DC↗
arXiv2026

RAEM: Robust Autonomous Exploration for Multi-Floor Environments with a Quadruped Robot

In this paper, we propose RAEM, a robust autonomous exploration framework for quadruped robots operating in multi-floor environments. Most existing ground-robot exploration approaches rely on planar traversability representations, which cannot adequately represent the overlapping structures and cross-floor connectivity of multi-floor buildings. Although tomography-based representations provide effective traversability modeling for multi-floor navigation, maintaining a global tomography map incurs substantial computational overhead for online exploration with frequent replanning. Moreover, sparse and fragmented LiDAR observations in stairwells can degrade local traversability estimation, leading to irregular viewpoint placement and temporary topological disconnections. To address these challenges, RAEM adopts a hybrid local-global traversability representation, in which a local tomography map and an explicitly categorized local 3D grid map are used for online terrain analysis and connectivity evaluation, while an elevation-aware global topological graph is incrementally constructed from these local spatial representations for efficient cross-floor exploration planning. We further introduce a staircase center alignment strategy to reduce abrupt yaw variations during climbing and a dual path searching mechanism to recover guidance paths when the global topology is locally disconnected. Extensive simulation and real-world experiments demonstrate robust and computationally stable autonomous exploration across multi-floor structures, including continuous exploration of a five-floor stairwell.

cs.RO↗
arXiv2026

A Deployment-Friendly Foundational Framework for Efficient Computational Pathology

Pathology foundation models (PFMs) generalize well across computational pathology tasks but remain costly for gigapixel whole-slide image analysis. Here, we present LitePath, a deployment-friendly framework that addresses model over-parameterization and patch-level redundancy. LitePath combines LiteFM, a compact model distilled from Virchow2, H-Optimus-1 and UNI2 using 190 million patches, with the Adaptive Patch Selector for task-specific patch selection. Compared with Virchow2, LitePath uses 28x fewer parameters and 403.5x fewer FLOPs. On an NVIDIA Jetson Orin Nano Super, it processes 208 slides per hour, 104.5x faster than Virchow2, and consumes 0.36 kWh per 3,000 slides, 171x less energy than Virchow2 on an RTX 3090 GPU. We evaluated LitePath on 45 multicenter cohorts across four organs and 33 tasks, comprising 37 classification and 8 survival cohorts, including 33 internal, 10 external and 2 prospective cohorts, with 17,837 slides from 9,977 patients disjoint from the pretraining data. Among 22 PFMs, LitePath achieved the best average rank (6.56 vs. 6.58 for Virchow2), retained 99.71% of Virchow2's Macro-AUC across classification cohorts, and improved mean C-index by 2.14 percentage points across survival cohorts (71.91% vs. 69.77%). We further introduce the Deployability Score (D-Score), a weighted geometric mean of normalized task performance and FLOPs. LitePath achieved the highest D-Score (0.8455), outperforming H0-mini (0.7723) and Virchow2 (0.7297). In a randomized paired crossover study of four pathologists and 120 cases, LitePath increased diagnostic accuracy by 4.1-15.8 percentage points and reduced diagnostic time by 10.9-14.2%. These results demonstrate rapid, cost-effective and energy-efficient pathology image analysis on accessible hardware while maintaining competitive performance.

cs.CV↗
arXiv2026

Prompt-Guided Foundation Model Tuning for Pathology Image Classification

Foundation models have become pivotal in advancing computational pathology, particularly for whole slide image (WSI) classification. However, prevailing methodologies often rely on frozen, pre-trained models for feature extraction, overlooking the pronounced domain shift and task discrepancy between the pre-training and downstream tasks. To address this challenge, we propose PAMT, a novel Prompt-guided Adaptive Model Transformation framework that enables precise adaptation of general foundation models to the distinct domain of histopathology. To encapsulate the intricate distributions characteristic of histopathological data, we introduce Representative Patch Sampling (RPS) and Prototypical Visual Prompt (PVP), which reconstruct the input into compact yet highly informative representations. Further, to effectively bridge the domain gap, we incorporate Adaptive Model Transformation (AMT) via adapter modules within the feature extraction pipeline, facilitating the acquisition of domain-specific features by the foundation model. We conduct rigorous evaluation across 14 publicly available datasets and demonstrate consistent, substantial improvements in classification accuracy. These results establish PAMT as a compelling new benchmark for pathology image classification and underscore the critical value of targeted model adaptation within computational pathology.

cs.CV↗
arXiv2026

SR-LIO++: LiDAR-Inertial Odometry and Quantized Mapping with Caching-Aware Sweep Reconstruction

Addressing the inherent low acquisition frequency limitation of 3D LiDAR to achieve high-frequency output has become a critical research focus in the LiDAR-Inertial Odometry (LIO) domain. To ensure real-time performance, frequency-enhanced LIO systems must process each sweep within significantly reduced timeframe, which presents substantial challenges for deployment on resource-constrained platforms. To address these limitations, we introduce SR-LIO++, an innovative LIO system capable of achieving doubled output frequency relative to input frequency on resource-constrained hardware platforms, including the Raspberry Pi 4B. Our system employs the previously proposed sweep reconstruction methodology to enhance LiDAR sweep frequency, generating high-frequency reconstructed sweeps. Building upon this foundation, we propose a caching mechanism for intermediate results (i.e., surface parameters) of the most recent segments, effectively minimizing redundant processing of common segments in adjacent reconstructed sweeps. This method decouples processing time from the traditionally linear dependence on reconstructed sweep frequency. Furthermore, we present a quantized map point management based on index table mapping, significantly reducing memory usage by converting global 3D point storage from 64-bit double precision to 8-bit char representation. This method also converts the computationally intensive Euclidean distance calculations in nearest neighbor searches from 64-bit double precision to 16-bit short and 32-bit integer formats, reducing computational cost. Extensive experimental evaluations across three distinct computing platforms and four public datasets demonstrate that SR-LIO++ maintains state-of-the-art accuracy while substantially enhancing efficiency. Notably, our system successfully achieves 20 Hz state output on Raspberry Pi 4B hardware.

cs.RO↗
arXiv2026

31.1 A 14.08-to-135.69Token/s ReRAM-on-Logic Stacked Outlier-Free Large-Language-Model Accelerator with Block-Clustered Weight-Compression and Adaptive Parallel-Speculative-Decoding

This work presents a 55nm speculative decoding-based LLM accelerator with bumping-based face-to-face ReRAM-on-logic stacking technology. It features a local rotation unit for outlier-free low-bit quantization, a stacking-aware PNM architecture co-designed with blockwise vector quantization to reduce weight EMA overheads, and an adaptive parallel speculative decoding scheme with an out-of-order scheduler for high resource and bandwidth utilization. Our chip achieves 14.08-to-135.69token/s and 4.46-to-7.17x speedup over vanilla speculative decoding.

cs.AR↗
arXiv2026

Expert Streaming: Accelerating Low-Batch MoE Inference via Multi-chiplet Architecture and Dynamic Expert Trajectory Scheduling

Mixture-of-Experts is a promising approach for edge AI with low-batch inference. Yet, on-device deployments often face limited on-chip memory and severe workload imbalance; the prevalent use of offloading further incurs off-chip memory access bottlenecks. Moreover, MoE sparsity and dynamic gating shift distributed strategies toward much finer granularity and introduce runtime scheduling considerations. Recently, high die-to-die bandwidth chiplet interconnects have created new opportunities for multi-chiplet systems to address workload imbalance and offloading bottlenecks with fine-grained scheduling. In this paper, we propose Fully Sharded Expert Data Parallelism, a parallelization paradigm specifically architected for low-batch MoE inference on multi-chiplet accelerators. FSE-DP attains adaptive computation-communication overlap and balanced load by orchestrating fine-grained, complementary expert streams along dynamic trajectories across high-bandwidth D2D links. The attendant dataflow complexity is tamed by a minimal, hardware-amenable set of virtualization rules and a lightweight scheduling algorithm. Our approach achieves 1.22 to 2.00 times speedup over state-of-the-art baselines and saves up to 78.8 percent on-chip memory.

cs.AR↗
arXiv2026

EoRA: Fine-tuning-free Compensation for Compressed LLM with Eigenspace Low-Rank Approximation

While post-training compression techniques effectively reduce the memory footprint, latency, and power consumption of Large Language Models (LLMs), they often result in noticeable accuracy degradation and remain limited by hardware and kernel constraints that restrict supported compression formats - ultimately reducing flexibility across a wide range of deployment scenarios. In this work, we propose EoRA - a novel $\textbf{fine-tuning-free}$ method that augments compressed LLMs with low-rank matrices, allowing users to rapidly enhance task-specific performance and freely balance the trade-off between accuracy and computational overhead beyond the constraints of compression formats. EoRA consistently outperforms prior fine-tuning-free low rank methods in recovering the accuracy of compressed LLMs, achieving notable accuracy improvements (e.g., $\mathbf{10.84\%}$ on ARC-Challenge, $\mathbf{6.74\%}$ on MathQA, and $\mathbf{11.45\%}$ on GSM8K for LLaMA3-8B compressed to 3-bit). We also introduce an optimized CUDA kernel, accelerating inference by up to 1.4x and reducing memory overhead through quantizing EoRA. Overall, EoRA offers a prompt solution for improving the accuracy of compressed models under varying user requirements, enabling more efficient and flexible deployment of LLMs. Code is available at https://github.com/NVlabs/EoRA.

cs.CL↗
arXiv2026

Path-Decoupled Hyperbolic Flow Matching for Few-Shot Adaptation

Recent advances in cross-modal few-shot adaptation treat visual-semantic alignment as a continuous feature transport problem via Flow Matching (FM). However, we argue that Euclidean-based FM overlooks fundamental limitations of flat geometry, where polynomial volume growth fails to accommodate diverse feature distributions, leading to severe path entanglement. To this end, we propose path-decoupled Hyperbolic Flow Matching (HFM), leveraging the Lorentz manifold's exponential expansion for trajectory decoupling. HFM structures the transport via two key designs: 1) Centripetal hyperbolic alignment: It constructs a centripetal hierarchy by anchoring textual roots, which pushes visual leaves to the boundary to initialize orderly flows. 2) Path-decoupled objective: It acts as a ``semantic guardrail'' rigidly confining trajectories within isolated class-specific geodesic corridors via step-wise supervision. Furthermore, we devise an adaptive diameter-based stopping to prevent over-transportation into the crowded origin based on the intrinsic semantic scale. Extensive ablations on 11 benchmarks have shown that HFM establishes a new state-of-the-art, consistently outperforming its Euclidean counterparts. Our codes and models will be released.

cs.CV↗
arXiv2026

GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization

As language models become increasingly capable, users expect them to provide not only accurate responses but also behaviors aligned with diverse human preferences across a variety of scenarios. To achieve this, Reinforcement learning (RL) pipelines have begun incorporating multiple rewards, each capturing a distinct preference, to guide models toward these desired behaviors. However, recent work has defaulted to apply Group Relative Policy Optimization (GRPO) under multi-reward setting without examining its suitability. In this paper, we demonstrate that directly applying GRPO to normalize distinct rollout reward combinations causes them to collapse into identical advantage values, reducing the resolution of the training signal and resulting in suboptimal convergence and, in some cases, early training failure. We then introduce Group reward-Decoupled Normalization Policy Optimization (GDPO), a new policy optimization method to resolve these issues by decoupling the normalization of individual rewards, more faithfully preserving their relative differences and enabling more accurate multi-reward optimization, along with substantially improved training stability. We compare GDPO with GRPO across three tasks: tool calling, math reasoning, and coding reasoning, evaluating both correctness metrics (accuracy, bug ratio) and constraint adherence metrics (format, length). Across all settings, GDPO consistently outperforms GRPO, demonstrating its effectiveness and generalizability for multi-reward reinforcement learning optimization.

cs.CL↗
arXiv2025

A Study of Finetuning Video Transformers for Multi-view Geometry Tasks

This paper presents an investigation of vision transformer learning for multi-view geometry tasks, such as optical flow estimation, by fine-tuning video foundation models. Unlike previous methods that involve custom architectural designs and task-specific pretraining, our research finds that general-purpose models pretrained on videos can be readily transferred to multi-view problems with minimal adaptation. The core insight is that general-purpose attention between patches learns temporal and spatial information for geometric reasoning. We demonstrate that appending a linear decoder to the Transformer backbone produces satisfactory results, and iterative refinement can further elevate performance to stateof-the-art levels. This conceptually simple approach achieves top cross-dataset generalization results for optical flow estimation with end-point error (EPE) of 0.69, 1.78, and 3.15 on the Sintel clean, Sintel final, and KITTI datasets, respectively. Our method additionally establishes a new record on the online test benchmark with EPE values of 0.79, 1.88, and F1 value of 3.79. Applications to 3D depth estimation and stereo matching also show strong performance, illustrating the versatility of video-pretrained models in addressing geometric vision tasks.

cs.CV↗
arXiv2025

Relation-R1: Progressively Cognitive Chain-of-Thought Guided Reinforcement Learning for Unified Relation Comprehension

Recent advances in multi-modal large language models (MLLMs) have significantly improved object-level grounding and region captioning. However, they remain limited in visual relation understanding, struggling even with binary relation detection, let alone \textit{N}-ary relations involving multiple semantic roles. The core reason is the lack of modeling for \textit{structural semantic dependencies} among multi-entities, leading to unreliable outputs, hallucinations, and over-reliance on language priors (\eg, defaulting to ``person drinks a milk'' if a person is merely holding it). To this end, we propose Relation-R1, the \textit{first unified} relation comprehension framework that explicitly integrates cognitive chain-of-thought (CoT)-guided supervised fine-tuning (SFT) and group relative policy optimization (GRPO) within a reinforcement learning (RL) paradigm. Specifically, we first establish foundational reasoning capabilities via SFT, enforcing structured outputs with thinking processes. Then, GRPO is utilized to refine these outputs via multi-rewards optimization, prioritizing visual-semantic grounding over language-induced biases, thereby improving generalization capability. Furthermore, we investigate the impact of various CoT strategies within this framework, demonstrating that a specific-to-general progressive approach in CoT guidance further improves generalization, especially in capturing synonymous \textit{N}-ary relations. Extensive experiments on widely-used PSG and SWiG datasets demonstrate that Relation-R1 achieves state-of-the-art performance in both binary and \textit{N}-ary relation understanding.

cs.CV↗
arXiv2025

DLER: Doing Length pEnalty Right - Incentivizing More Intelligence per Token via Reinforcement Learning

Reasoning language models such as OpenAI-o1, DeepSeek-R1, and Qwen achieve strong performance via extended chains of thought but often generate unnecessarily long outputs. Maximizing intelligence per token--accuracy relative to response length--remains an open problem. We revisit reinforcement learning (RL) with the simplest length penalty--truncation--and show that accuracy degradation arises not from the lack of sophisticated penalties but from inadequate RL optimization. We identify three key challenges: (i) large bias in advantage estimation, (ii) entropy collapse, and (iii) sparse reward signal. We address them with Doing Length pEnalty Right (DLER), a training recipe combining batch-wise reward normalization, higher clipping, dynamic sampling, and a simple truncation length penalty. DLER achieves state-of-the-art accuracy--efficiency trade-offs, cutting output length by over 70 percent while surpassing all previous baseline accuracy. It also improves test-time scaling: compared to DeepSeek-R1-7B, DLER-7B generates multiple concise responses in parallel with 28 percent higher accuracy and lower latency. We further introduce Difficulty-Aware DLER, which adaptively tightens truncation on easier questions for additional efficiency gains. We also propose an update-selective merging method that preserves baseline accuracy while retaining the concise reasoning ability of the DLER model, which is useful for scenarios where RL training data is scarce.

cs.LG↗
arXiv2025

Token Merging via Spatiotemporal Information Mining for Surgical Video Understanding

Vision Transformer models have shown impressive effectiveness in the surgical video understanding tasks through long-range dependency modeling. However, current methods suffer from prohibitive computational costs due to processing massive spatiotemporal tokens across video frames. While prior work on token merging has advanced model efficiency, they fail to adequately consider the inherent spatiotemporal structure of video data and overlook the heterogeneous nature of information distribution, leading to suboptimal performance. In this paper, we propose a spatiotemporal information mining token merging (STIM-TM) method, representing the first dedicated approach for surgical video understanding. STIM-TM introduces a decoupled strategy that reduces token redundancy along temporal and spatial dimensions independently. Specifically, the temporal component merges spatially corresponding tokens from consecutive frames using saliency weighting, preserving critical sequential information and maintaining continuity. Meanwhile, the spatial component prioritizes merging static tokens through temporal stability analysis, protecting dynamic regions containing essential surgical information. Operating in a training-free manner, STIM-TM achieves significant efficiency gains with over $65\%$ GFLOPs reduction while preserving competitive accuracy across comprehensive surgical video tasks. Our method also supports efficient training of long-sequence surgical videos, addressing computational bottlenecks in surgical applications.

cs.CV↗
arXiv2025

Towards Customized Knowledge Distillation for Chip-Level Dense Image Predictions

It has been revealed that efficient dense image prediction (EDIP) models designed for AI chips, trained using the knowledge distillation (KD) framework, encounter two key challenges, including \emph{maintaining boundary region completeness} and \emph{ensuring target region connectivity}, despite their favorable real-time capacity to recognize the main object regions. In this work, we propose a customized boundary and context knowledge distillation (BCKD) method for EDIPs, which facilitates the targeted KD from large accurate teacher models to compact small student models. Specifically, the \emph{boundary distillation} focuses on extracting explicit object-level boundaries from the hierarchical feature maps to enhance the student model's mask quality in boundary regions. Meanwhile, the \emph{context distillation} leverages self-relations as a bridge to transfer implicit pixel-level contexts from the teacher model to the student model, ensuring strong connectivity in target regions. Our proposed method is specifically designed for the EDIP tasks and is characterized by its simplicity and efficiency. Theoretical analysis and extensive experimental results across semantic segmentation, object detection, and instance segmentation on five representative datasets demonstrate the effectiveness of BCKD, resulting in well-defined object boundaries and smooth connecting regions.

cs.CV↗