2026-03-16

Table of Contents

MoEKD Mixture-of-Experts Knowledge Distillation for Robust and High-Performing Compressed Code Models

针对代码大模型计算开销高的问题，提出专家混合知识蒸馏方法，提升压缩模型性能与鲁棒性。

Authors: Md. Abdul Awal, Mrigank Rochan, Chanchal K. Roy | Date: 2026-03-13

Link: http://arxiv.org/abs/2603.13213v1

Abstract Large language models for code have achieved strong performance across diverse software analytics tasks, yet their real-world adoption remains limited by high computational demands, slow inference speeds, significant energy consumption, and environmental impact. Knowledge distillation (KD) offers a practical solution by transferring knowledge from a large model to a smaller and more efficient model. Despite its effectiveness, recent studies show that models distilled from a single source often exhibit degraded adversarial robustness, even when robustness-aware distillation techniques are employed. These observations suggest a fundamental limitation of single-source distillation in simultaneously transferring high-quality and robust knowledge. To overcome this limitation, we propose Mixture of Experts Knowledge Distillation (MoEKD), a KD framework that leverages a Mixture of Experts (MoE) architecture to enable more effective and robust knowledge transfer from multiple specialized experts into a compact model. MoEKD decomposes the distillation process into expert and router training, aggregation of expert knowledge through a learned routing mechanism, and distillation from the aggregated knowledge. We evaluate MoEKD on the vulnerability detection task using CodeBERT and GraphCodeBERT models. Experimental results show that MoEKD not only improves adversarial robustness by up to 35.8%, but also enhances predictive performance by up to 13%, compared to state-of-the-art KD baselines, including Compressor and AVATAR. Furthermore, an ablation study demonstrates that aggregating expert knowledge enables ultra-compact models to maintain competitive performance even when their size is reduced by approximately half. Overall, these results highlight the effectiveness of multi-expert knowledge aggregation in addressing key limitations of existing single-source KD approaches.

AgentRM An OS-Inspired Resource Manager for LLM Agent Systems

针对LLM代理系统资源管理问题，提出OS启发的资源管理器，解决调度失败与系统无响应问题。

Authors: Jianshu She | Date: 2026-03-13

Link: http://arxiv.org/abs/2603.13110v1

Abstract Large Language Model (LLM) agent systems have experienced rapid adoption across diverse domains, yet they suffer from critical user experience problems that limit their practical deployment. Through an empirical analysis of over 40,000 GitHub issues from six major agent frameworks (OpenClaw, AutoGen, CrewAI, LangGraph, Codex, Claude Code), we identify two fundamental resource management challenges: (1) scheduling failures leading to system unresponsiveness due to blocking, zombie processes, and rate limit cascades, and (2) context degradation causing agent "amnesia" from unbounded memory growth and poor retention policies. Drawing inspiration from decades of operating systems research, we present AgentRM, a middleware resource manager that treats agent resources analogously to OS resources. AgentRM employs a Multi-Level Feedback Queue (MLFQ) scheduler with zombie reaping and rate-limit-aware admission control, coupled with a three-tier Context Lifecycle Manager that implements adaptive compaction and hibernation mechanisms. Our evaluation demonstrates significant improvements: AgentRM-MLFQ reduces P95 latency by 86%, decreases lane waste by 96%, and increases throughput by 168% while eliminating zombie agents (0 vs. 29 baseline). AgentRM-CLM achieves 100% key information retention with 95% quality score compared to 65.1% retention and 87% quality for existing approaches, albeit with higher compaction costs (34,330 vs. 17,212 tokens).

Structured Distillation for Personalized Agent Memory 11x Token Reduction with Retrieval Preservation

针对长对话历史存储开销问题，提出结构化蒸馏方法，实现11倍令牌压缩并保留检索能力。

Authors: Sydney Lewis | Date: 2026-03-13

Link: http://arxiv.org/abs/2603.13017v1

Abstract Long conversations with an AI agent create a simple problem for one user: the history is useful, but carrying it verbatim is expensive. We study personalized agent memory: one user's conversation history with an agent, distilled into a compact retrieval layer for later search. Each exchange is compressed into a compound object with four fields (exchange_core, specific_context, thematic room_assignments, and regex-extracted files_touched). The searchable distilled text averages 38 tokens per exchange. Applied to 4,182 conversations (14,340 exchanges) from 6 software engineering projects, the method reduces average exchange length from 371 to 38 tokens, yielding 11x compression. We evaluate whether personalized recall survives that compression using 201 recall-oriented queries, 107 configurations spanning 5 pure and 5 cross-layer search modes, and 5 LLM graders (214,519 consensus-graded query-result pairs). The best pure distilled configuration reaches 96% of the best verbatim MRR (0.717 vs 0.745). Results are mechanism-dependent. All 20 vector search configurations remain non-significant after Bonferroni correction, while all 20 BM25 configurations degrade significantly (effect sizes |d|=0.031-0.756). The best cross-layer setup slightly exceeds the best pure verbatim baseline (MRR 0.759). Structured distillation compresses single-user agent memory without uniformly sacrificing retrieval quality. At 1/11 the context cost, thousands of exchanges fit within a single prompt while the verbatim source remains available for drill-down. We release the implementation and analysis pipeline as open-source software.

Dependency-Aware Parallel Decoding via Attention for Diffusion LLMs

针对扩散LLM并行解码依赖问题，提出注意力感知依赖解码方法，提升并行效率与准确性。

Authors: Bumjun Kim, Dongjae Jeon, Moongyu Jeon, Albert No | Date: 2026-03-13

Link: http://arxiv.org/abs/2603.12996v1

Abstract Parallel decoding for diffusion LLMs (dLLMs) is difficult because each denoising step provides only token-wise marginal distributions, while unmasking multiple tokens simultaneously requires accounting for inter-token dependencies. We propose Dependency-Aware Parallel Decoding (DAPD), a simple, training-free decoding method that uses self-attention to induce a conditional dependency graph over masked tokens. At each iteration, edges in this graph capture strong token interactions, while non-edges indicate weak dependence. Parallel decoding is then reduced to selecting an independent set on the graph and unmasking the selected tokens in parallel. This avoids co-updating strongly coupled tokens without auxiliary models or retraining. Experiments on LLaDA and Dream show that DAPD improves the accuracy-steps trade-off over existing methods and enables more globally distributed parallel updates that better exploit the any-order generation capability of dLLMs.

Serving Hybrid LLM Loads with SLO Guarantees Using CPU-GPU Attention Piggybacking

针对混合LLM负载干扰问题，提出CPU-GPU注意力共享方法，保障SLO并提升服务容量。

Authors: Zizhao Mo, Junlin Chen, Huanle Xu, Chengzhong Xu | Date: 2026-03-13

Link: http://arxiv.org/abs/2603.12831v1

Abstract Nowadays, service providers often deploy multiple types of LLM services within shared clusters. While the service colocation improves resource utilization, it introduces significant interference risks for latency-sensitive (LS) services-which have strict SLO requirements for inference latency-and severely constrain the service capacity of best-effort (BE) services due to limited available memory. To address interference, existing systems typically rely on reserving headroom to constrain BE resource usage. However, this approach's coarse granularity compromises the SLO compliance of the latency-sensitive service and unnecessarily restricts the generation potential of the best effort service. In this paper, we propose OmniServe, a novel LLM serving system that efficiently harnesses both CPU and GPU resources to mitigate interference and improve throughput. Central to OmniServe is the Attention Piggybacking mechanism, which effectively offloads the Attention computation of BE services to CPUs on the fly. This mechanism also facilitates asynchronous communication between CPU and GPU streams, preventing GPUs from being blocked while aggregating Attention results. Additionally, OmniServe incorporates a dynamic batching control policy to adapt to fluctuating request arrivals, facilitating Dense module computation using layer-wise batching. Experimental results show that OmniServe improves the SLO attainment rate for LS services by up to $1.48\times$ while enhancing BE serving throughput by up to $9.85\times$ compared to state-of-the-art systems.

NanoVDR Distilling a 2B Vision-Language Retriever into a 70M Text-Only Encoder for Visual Document Retrieval

针对视觉文档检索高延迟问题，提出不对称蒸馏方法，将2B视觉语言检索器压缩为70M文本编码器。

Authors: Zhuchenyang Liu, Yao Zhang, Yu Xiao | Date: 2026-03-13

Link: http://arxiv.org/abs/2603.12824v1

Abstract Vision-Language Model (VLM) based retrievers have advanced visual document retrieval (VDR) to impressive quality. They require the same multi-billion parameter encoder for both document indexing and query encoding, incurring high latency and GPU dependence even for plain-text queries. We observe that this design is unnecessarily symmetric: documents are visually complex and demand strong visual understanding, whereas queries are just short text strings. NanoVDR exploits this query--document asymmetry by decoupling the two encoding paths: a frozen 2B VLM teacher indexes documents offline, while a distilled text-only student as small as 69M parameters encodes queries at inference. The key design choice is the distillation objective. Through systematic comparison of six objectives across three backbones and 22 ViDoRe benchmark datasets, we find that pointwise cosine alignment on query text consistently outperforms ranking-based and contrastive alternatives, while requiring only pre-cached teacher query embeddings and no document processing during training. Furthermore, we identify cross-lingual transfer as the primary performance bottleneck, and resolve it cheaply by augmenting training data with machine-translated queries. The resulting NanoVDR-S-Multi (DistilBERT, 69M) retains 95.1\% of teacher quality and outperforms DSE-Qwen2 (2B) on v2 and v3 with 32$\times$ fewer parameters and 50$\times$ lower CPU query latency, at a total training cost under 13 GPU-hours.

ToolTree Efficient LLM Agent Tool Planning via Dual-Feedback Monte Carlo Tree Search and Bidirectional Pruning

针对LLM代理工具规划缺乏 foresight 问题，提出双反馈蒙特卡洛树搜索方法，提升规划效率与准确性。

Authors: Shuo Yang, Soyeon Caren Han, Yihao Ding, Shuhe Wang, Eduard Hoy | Date: 2026-03-13

Link: http://arxiv.org/abs/2603.12740v1

Abstract Large Language Model (LLM) agents are increasingly applied to complex, multi-step tasks that require interaction with diverse external tools across various domains. However, current LLM agent tool planning methods typically rely on greedy, reactive tool selection strategies that lack foresight and fail to account for inter-tool dependencies. In this paper, we present ToolTree, a novel Monte Carlo tree search-inspired planning paradigm for tool planning. ToolTree explores possible tool usage trajectories using a dual-stage LLM evaluation and bidirectional pruning mechanism that enables the agent to make informed, adaptive decisions over extended tool-use sequences while pruning less promising branches before and after the tool execution. Empirical evaluations across both open-set and closed-set tool planning tasks on 4 benchmarks demonstrate that ToolTree consistently improves performance while keeping the highest efficiency, achieving an average gain of around 10\% compared to the state-of-the-art planning paradigm.

Cost-Efficient Multimodal LLM Inference via Cross-Tier GPU Heterogeneity

针对多模态LLM推理硬件需求冲突问题，提出跨层级GPU异构方法，降低传输开销与成本。

Authors: Donglin Yu | Date: 2026-03-13

Link: http://arxiv.org/abs/2603.12707v1

Abstract Multimodal large language model (MLLM) inference splits into two phases with opposing hardware demands: vision encoding is compute-bound, while language generation is memory-bandwidth-bound. We show that under standard transformer KV caching, the modality boundary (between vision encoder and language model) minimizes cross-device transfer among all partition points that preserve standard stage-based execution. Partitioning here reduces transfer complexity from $O(L * s_ctx)$ bytes (GB-scale KV caches under stage-level disaggregation) to $O(N_v * d)$ bytes (MB-scale embeddings), an O(L) reduction where L is the transformer depth. The result holds across attention mechanisms (MHA/GQA), dynamic vision resolutions, and model scales, and the advantage grows as models deepen. A direct implication is that existing stage-level disaggregation systems are constrained to high-bandwidth interconnects (e.g., NVLink), whereas modality-level disaggregation enables cross-tier heterogeneous serving over commodity PCIe. A closed-form cost model shows that heterogeneous deployment is cost-optimal under phase-separable workloads (predicts 31.4% savings; observed 40.6%). We build HeteroServe, a phase-aware runtime with modality-level partitioning and cross-tier scheduling, and evaluate it on LLaVA-1.5-7B and Qwen2.5-VL against vLLM v0.3.0. On identical 4xA100 hardware, engine optimizations raise throughput by up to 54%. Under a fixed budget, a heterogeneous cluster (\$38k) improves Tokens/\$ by 37% over a homogeneous baseline (\$64k) without degrading latency.

HyGra Accelerating Network-State Simulation for LLM Training in DCNs via Adaptive Packet-Flow Granularity

针对LLM训练网络通信瓶颈问题，提出自适应包流粒度模拟方法，加速网络状态仿真。

Authors: Wenyi Wang, Zheng Wu, Yanmeng Wang, Haolin Mao, Lei Han, Gaogang Xie, Fu Xiao | Date: 2026-03-13

Link: http://arxiv.org/abs/2603.12671v1

Abstract In recent years, large language models (LLMs) have driven substantial intelligent transformation across diverse industries. Commercial LLM training is typically performed over data center networks (DCNs) comprising hundreds to thousands of GPUs, with multiple devices collocated per node. As network scale expands, inter-node communication becomes a primary bottleneck to training efficiency. Network-state simulators therefore play a crucial role by enabling cost-effective evaluation of network configurations and parallelization strategies through faithful emulation of DCN dynamics during LLM training. However, existing simulators are constrained by a efficiency-fidelity tradeoff, as packet-level simulators (PLSs) incur prohibitive runtime overhead, whereas flow-level simulators (FLSs) compromise essential modeling accuracy. In this paper, we develop \texttt{HyGra}, a hybrid-granularity network-state simulator that exploits intrinsic network dynamics in LLM training to adaptively switch simulation granularity. Specifically, \texttt{HyGra} employs packet-level simulation during non-steady phases with transient fluctuations and flow-level simulation during steady phases with periodic patterns, thereby accelerating execution while preserving high fidelity. Moreover, it requires no specialized hardware, supports single-machine deployment, and is compatible with existing simulators. Experiments based representative commercial LLM workloads, including ChatGPT, DeepSeek, and Qwen, show that \texttt{HyGra} achieves up to 15.4$\times$ speedup under single parallelization strategy and 7.8$\times$ under hybrid parallelization strategies while maintaining high accuracy.

98 Faster LLM Routing Without a Dedicated GPU Flash Attention, Prompt Compression, and Near-Streaming for the vLLM Semantic Router

针对LLM路由长上下文分类内存问题，提出闪存注意力与提示压缩方法，实现98倍加速且无需专用GPU。

Authors: Xunzhuo Liu, Bowei He, Xue Liu, Andy Luo, Haichen Zhang, Huamin Chen | Date: 2026-03-13

Link: http://arxiv.org/abs/2603.12646v1

Abstract System-level routers that intercept LLM requests for safety classification, domain routing, and PII detection must be both fast and operationally lightweight: they should add minimal latency to every request, yet not require a dedicated GPU -- an expensive resource better used for LLM inference itself. When the router co-locates on the same GPU as vLLM serving instances, standard attention's $O(n^2)$ memory makes long-context classification (8K--32K tokens) impossible: at 8K tokens, three concurrent classifiers need ${\sim}$4.5\,GB for attention masks alone, far exceeding the memory left by vLLM. We present three staged optimizations for the vLLM Semantic Router, benchmarked on AMD Instinct MI300X, that solve both the latency and the memory problem. \emph{Stage~1}: a custom CK Flash Attention operator for ONNX Runtime on ROCm reduces attention memory from $O(n^2)$ to $O(n)$ and end-to-end (E2E) latency from 4{,}918\,ms to 127\,ms (\textbf{38.7$\times$}), enabling 8K--32K tokens where SDPA OOMs. \emph{Stage~2}: classical NLP prompt compression (TextRank, position weighting, TF-IDF, and novelty scoring) reduces all inputs to ${\sim}$512 tokens without neural inference, capping both latency and GPU memory at a constant regardless of original prompt length (E2E 127$\to$62\,ms, \textbf{2.0$\times$}). \emph{Stage~3}: near-streaming body processing with adaptive chunking and zero-copy JSON eliminates serialization overhead (E2E 62$\to$50\,ms, \textbf{1.2$\times$}). Cumulatively: \textbf{98$\times$} improvement (4{,}918\,ms to 50\,ms), 16K-token routing in 108\,ms, and a total router GPU footprint under 800\,MB -- small enough to share a GPU with LLM serving and removing the need for a dedicated accelerator. Stage~1 targets AMD ROCm (NVIDIA GPUs already have FlashAttention via cuDNN); Stages~2 and~3 are hardware-agnostic.

LightMoE Reducing Mixture-of-Experts Redundancy through Expert Replacing

针对MoE模型内存开销高问题，提出专家替换压缩方法，减少冗余并降低训练开销。

Authors: Jiawei Hao, Zhiwei Hao, Jianyuan Guo, Li Shen, Yong Luo, Han Hu, Dan Zeng | Date: 2026-03-13

Link: http://arxiv.org/abs/2603.12645v1

Abstract Mixture-of-Experts (MoE) based Large Language Models (LLMs) have demonstrated impressive performance and computational efficiency. However, their deployment is often constrained by substantial memory demands, primarily due to the need to load numerous expert modules. While existing expert compression techniques like pruning or merging attempt to mitigate this, they often suffer from irreversible knowledge loss or high training overhead. In this paper, we propose a novel expert compression paradigm termed expert replacing, which replaces redundant experts with parameter-efficient modules and recovers their capabilities with low training costs. We find that even a straightforward baseline of this paradigm yields promising performance. Building on this foundation, we introduce LightMoE, a framework that enhances the paradigm by introducing adaptive expert selection, hierarchical expert construction, and an annealed recovery strategy. Experimental results show that LightMoE matches the performance of LoRA fine-tuning at a 30% compression ratio. Even under a more aggressive 50% compression rate, it outperforms existing methods and achieves average performance improvements of 5.6% across five diverse tasks. These findings demonstrate that LightMoE strikes a superior balance among memory efficiency, training efficiency, and model performance.

When Drafts Evolve Speculative Decoding Meets Online Learning

针对推测解码草稿模型容量不足问题，提出在线学习进化草稿方法，提升接受长度与加速比。

Authors: Yu-Yang Qian, Hao-Cong Wu, Yichao Fu, Hao Zhang, Peng Zhao | Date: 2026-03-13

Link: http://arxiv.org/abs/2603.12617v1

Abstract Speculative decoding has emerged as a widely adopted paradigm for accelerating large language model inference, where a lightweight draft model rapidly generates candidate tokens that are then verified in parallel by a larger target model. However, due to limited model capacity, drafts often struggle to approximate the target distribution, resulting in shorter acceptance lengths and diminished speedup. A key yet under-explored observation is that speculative decoding inherently provides verification feedback that quantifies the deviation between the draft and target models at no additional cost. This process naturally forms an iterative "draft commits-feedback provides-draft adapts" evolving loop, which precisely matches the online learning paradigm. Motivated by this connection, we propose OnlineSpec, a unified framework that systematically leverages interactive feedback to continuously evolve draft models. Grounded in dynamic regret minimization, we establish a formal link between online learning performance and speculative system's acceleration rate, and develop novel algorithms via modern online learning techniques, including optimistic online learning that adaptively reuses historical gradients as predictive update hints, and online ensemble learning that dynamically maintains multiple draft models. Our algorithms are equipped with theoretical justifications and improved acceleration rates, achieving up to 24% speedup over seven benchmarks and three foundation models.

CA-HFP Curvature-Aware Heterogeneous Federated Pruning with Model Reconstruction

针对异构联邦学习压缩兼容问题，提出曲率感知剪枝与模型重建方法，保障收敛与个性化。

Authors: Gang Hu, Yinglei Teng, Pengfei Wu, Shijun Ma | Date: 2026-03-13

Link: http://arxiv.org/abs/2603.12591v1

Abstract Federated learning on heterogeneous edge devices requires personalized compression while preserving aggregation compatibility and stable convergence. We present Curvature-Aware Heterogeneous Federated Pruning (CA-HFP), a practical framework that enables each client perform structured, device-specific pruning guided by a curvature-informed significance score, and subsequently maps its compact submodel back into a common global parameter space via a lightweight reconstruction. We derive a convergence bound for federated optimization with multiple local SGD steps that explicitly accounts for local computation, data heterogeneity, and pruning-induced perturbations; from which a principled loss-based pruning criterion is derived. Extensive experiments on FMNIST, CIFAR-10, and CIFAR-100 using VGG and ResNet architectures under varying degrees of data heterogeneity demonstrate that CA-HFP preserves model accuracy while significantly reducing per-client computation and communication costs, outperforming standard federated training and existing pruning-based baselines.

Expert Pyramid Tuning Efficient Parameter Fine-Tuning for Expertise-Driven Task Allocation

针对MoE微调任务分配不均问题，提出专家金字塔调优方法，提升参数效率与任务适应性。

Authors: Jia-Chen Zhang, Zhen-Wei Yan, Yu-Jie Xiong, Chun-Ming Xia | Date: 2026-03-13

Link: http://arxiv.org/abs/2603.12577v1

Abstract Parameter-Efficient Fine-Tuning (PEFT) has become a dominant paradigm for deploying LLMs in multi-task scenarios due to its extreme parameter efficiency. While Mixture-of-Experts (MoE) based LoRA variants have achieved promising results by dynamically routing tokens to different low-rank experts, they largely overlook the hierarchical nature of task complexity. Existing methods typically employ experts with uniform architectures, limiting their ability to capture diverse feature granularities required by distinct tasks--where some tasks demand high-level semantic abstraction while others require fine-grained syntactic manipulation. To bridge this gap, we propose Expert Pyramid Tuning (EPT), a novel architecture that integrates the multi-scale feature pyramid concept from computer vision into the realm of PEFT. Unlike standard LoRA, EPT decomposes task adaptation into two stages: (1) A shared meta-knowledge Subspace that encodes universal linguistic patterns in low dimensions; (2) A Pyramid Projection Mechanism that utilizes learnable up-projection operators to reconstruct high-dimensional features at varying scales. A task-aware router then dynamically selects the optimal combination of these multi-scale features. Extensive experiments across multiple multi-task benchmarks demonstrate that EPT significantly outperforms SOTA MoE-LoRA variants. Crucially, thanks to the re-parameterization capability of our design, EPT achieves this performance improvement while simultaneously reducing the number of training parameters.

TaxBreak Unmasking the Hidden Costs of LLM Inference Through Overhead Decomposition

针对LLM推理开销不透明问题，提出TaxBreak分解方法，识别主机端开销以指导优化。

Authors: Prabhu Vellaisamy, Shreesh Tripathi, Vignesh Natarajan, Surya Santhan Thenarasu, Shawn Blanton, John P. Shen | Date: 2026-03-12

Link: http://arxiv.org/abs/2603.12465v1

Abstract Large Language Model (LLM) inference is widely used in interactive assistants and agentic systems. In latency-sensitive deployments, inference time can become dominated by host-side overheads. Existing approaches typically expose this cost only as an aggregate residual or a launch/queue metric, which is often insufficient to identify which execution layer should be optimized. This work presents TaxBreak, a trace-driven methodology for decomposing host-visible orchestration overhead into three components: framework translation time, CUDA library translation time, and kernel launch-path time. We validate TaxBreak on NVIDIA H100 and H200 systems and use it to derive our proposed Host-Device Balance Index (HDBI), a boundedness summary index that relates device-active execution to host-visible orchestration. Across representative dense and mixture-of-experts workloads in both prefill and decode, we show that aggregate latency, GPU inactivity, or boundedness ratios alone can obscure the dominant optimization target. TaxBreak instead distinguishes cases where optimization should reduce software-stack overhead from cases where the primary win comes from reducing device-side work. We further show that MoE models dispatch 8-11x more kernels per output token than dense models, and that for such host-bound workloads, CPU single-thread performance is a first-order parameter: a faster host CPU reduces orchestration overhead by 10-29% and improves end-to-end latency by up to 14%, even when paired with a slower-clocked GPU. These results position TaxBreak as a diagnostic tool for assessing whether optimization effort should target the software stack or the device-side workload execution.

Test-Time Strategies for More Efficient and Accurate Agentic RAG

针对多跳问答RAG效率问题，提出测试时策略优化方法，减少重复检索并提升生成准确性。

Authors: Brian Zhang, Deepti Guntur, Zhiyang Zuo, Abhinav Sharma, Shreyas Chaudhari, Wenlong Zhao, Franck Dernoncourt, Puneet Mathur, Ryan Rossi, Nedim Lipka | Date: 2026-03-12

Link: http://arxiv.org/abs/2603.12396v1

Abstract Retrieval-Augmented Generation (RAG) systems face challenges with complex, multihop questions, and agentic frameworks such as Search-R1 (Jin et al., 2025), which operates iteratively, have been proposed to address these complexities. However, such approaches can introduce inefficiencies, including repetitive retrieval of previously processed information and challenges in contextualizing retrieved results effectively within the current generation prompt. Such issues can lead to unnecessary retrieval turns, suboptimal reasoning, inaccurate answers, and increased token consumption. In this paper, we investigate test-time modifications to the Search-R1 pipeline to mitigate these identified shortcomings. Specifically, we explore the integration of two components and their combination: a contextualization module to better integrate relevant information from retrieved documents into reasoning, and a de-duplication module that replaces previously retrieved documents with the next most relevant ones. We evaluate our approaches using the HotpotQA (Yang et al., 2018) and the Natural Questions (Kwiatkowski et al., 2019) datasets, reporting the exact match (EM) score, an LLM-as-a-Judge assessment of answer correctness, and the average number of turns. Our best-performing variant, utilizing GPT-4.1-mini for contextualization, achieves a 5.6% increase in EM score and reduces the number of turns by 10.5% compared to the Search-R1 baseline, demonstrating improved answer accuracy and retrieval efficiency.

NeuroLoRA Context-Aware Neuromodulation for Parameter-Efficient Multi-Task Adaptation

针对LoRA微调参数干扰问题，提出神经调制上下文感知方法，提升多任务适应效率。

Authors: Yuxin Yang, Haoran Zhang, Mingxuan Li, Jiachen Xu, Ruoxi Shen, Zhenyu Wang, Tianhao Liu, Siqi Chen, Weilin Huang | Date: 2026-03-12

Link: http://arxiv.org/abs/2603.12378v1

Abstract Parameter-Efficient Fine-Tuning (PEFT) techniques, particularly Low-Rank Adaptation (LoRA), have become essential for adapting Large Language Models (LLMs) to downstream tasks. While the recent FlyLoRA framework successfully leverages bio-inspired sparse random projections to mitigate parameter interference, it relies on a static, magnitude-based routing mechanism that is agnostic to input context. In this paper, we propose NeuroLoRA, a novel Mixture-of-Experts (MoE) based LoRA framework inspired by biological neuromodulation -- the dynamic regulation of neuronal excitability based on context. NeuroLoRA retains the computational efficiency of frozen random projections while introducing a lightweight, learnable neuromodulation gate that contextually rescales the projection space prior to expert selection. We further propose a Contrastive Orthogonality Loss to explicitly enforce separation between expert subspaces, enhancing both task decoupling and continual learning capacity. Extensive experiments on MMLU, GSM8K, and ScienceQA demonstrate that NeuroLoRA consistently outperforms FlyLoRA and other strong baselines across single-task adaptation, multi-task model merging, and sequential continual learning scenarios, while maintaining comparable parameter efficiency.

Alternating Gradient Flow Utility A Unified Metric for Structural Pruning and Dynamic Routing in Deep Networks

针对结构化剪枝指标偏差问题，提出交替梯度流效用方法，优化深度网络剪枝与动态路由。

Authors: Tianhao Qian, Zhuoxuan Li, Jinde Cao, Xinli Shi, Hanjie Liu, Leszek Rutkowski | Date: 2026-03-12

Link: http://arxiv.org/abs/2603.12354v1

Abstract Efficient deep learning traditionally relies on static heuristics like weight magnitude or activation awareness (e.g., Wanda, RIA). While successful in unstructured settings, we observe a critical limitation when applying these metrics to the structural pruning of deep vision networks. These contemporary metrics suffer from a magnitude bias, failing to preserve critical functional pathways. To overcome this, we propose a decoupled kinetic paradigm inspired by Alternating Gradient Flow (AGF), utilizing an absolute feature-space Taylor expansion to accurately capture the network's structural "kinetic utility". First, we uncover a topological phase transition at extreme sparsity, where AGF successfully preserves baseline functionality and exhibits topological implicit regularization, avoiding the collapse seen in models trained from scratch. Second, transitioning to architectures without strict structural priors, we reveal a phenomenon of Sparsity Bottleneck in Vision Transformers (ViTs). Through a gradient-magnitude decoupling analysis, we discover that dynamic signals suffer from signal compression in converged models, rendering them suboptimal for real-time routing. Finally, driven by these empirical constraints, we design a hybrid routing framework that decouples AGF-guided offline structural search from online execution via zero-cost physical priors. We validate our paradigm on large-scale benchmarks: under a 75% compression stress test on ImageNet-1K, AGF effectively avoids the structural collapse where traditional metrics aggressively fall below random sampling. Furthermore, when systematically deployed for dynamic inference on ImageNet-100, our hybrid approach achieves Pareto-optimal efficiency. It reduces the usage of the heavy expert by approximately 50% (achieving an estimated overall cost of 0.92$\times$) without sacrificing the full-model accuracy.

Generalist Large Language Models for Molecular Property Prediction Distilling Knowledge from Specialist Models

针对LLM分子预测性能不足问题，提出树知识蒸馏方法，从专家模型转移知识提升准确性。

Authors: Khiem Le, Sreejata Dey, Marcos Martínez Galindo, Vanessa Lopez, Ting Hua, Nitesh V. Chawla, Hoang Thanh Lam | Date: 2026-03-12

Link: http://arxiv.org/abs/2603.12344v1

Abstract Molecular Property Prediction (MPP) is a central task in drug discovery. While Large Language Models (LLMs) show promise as generalist models for MPP, their current performance remains below the threshold for practical adoption. We propose TreeKD, a novel knowledge distillation method that transfers complementary knowledge from tree-based specialist models into LLMs. Our approach trains specialist decision trees on functional group features, then verbalizes their learned predictive rules as natural language to enable rule-augmented context learning. This enables LLMs to leverage structural insights that are difficult to extract from SMILES strings alone. We further introduce rule-consistency, a test-time scaling technique inspired by bagging that ensembles predictions across diverse rules from a Random Forest. Experiments on 22 ADMET properties from the TDC benchmark demonstrate that TreeKD substantially improves LLM performance, narrowing the gap with SOTA specialist models and advancing toward practical generalist models for molecular property prediction.

One Model, Many Budgets Elastic Latent Interfaces for Diffusion Transformers

针对扩散Transformer计算固定问题，提出弹性潜在接口方法，实现多预算质量-延迟权衡。

Authors: Moayed Haji-Ali, Willi Menapace, Ivan Skorokhodov, Dogyun Park, Anil Kag, Michael Vasilkovsky, Sergey Tulyakov, Vicente Ordonez, Aliaksandr Siarohin | Date: 2026-03-12

Link: http://arxiv.org/abs/2603.12245v1

Abstract Diffusion transformers (DiTs) achieve high generative quality but lock FLOPs to image resolution, limiting principled latency-quality trade-offs, and allocate computation uniformly across input spatial tokens, wasting resource allocation to unimportant regions. We introduce Elastic Latent Interface Transformer (ELIT), a drop-in, DiT-compatible mechanism that decouples input image size from compute. Our approach inserts a latent interface, a learnable variable-length token sequence on which standard transformer blocks can operate. Lightweight Read and Write cross-attention layers move information between spatial tokens and latents and prioritize important input regions. By training with random dropping of tail latents, ELIT learns to produce importance-ordered representations with earlier latents capturing global structure while later ones contain information to refine details. At inference, the number of latents can be dynamically adjusted to match compute constraints. ELIT is deliberately minimal, adding two cross-attention layers while leaving the rectified flow objective and the DiT stack unchanged. Across datasets and architectures (DiT, U-ViT, HDiT, MM-DiT), ELIT delivers consistent gains. On ImageNet-1K 512px, ELIT delivers an average gain of $35.3\%$ and $39.6\%$ in FID and FDD scores. Project page: https://snap-research.github.io/elit/

BiGain Unified Token Compression for Joint Generation and Classification

针对扩散模型加速忽略分类问题，提出双增益令牌压缩方法，联合优化生成与分类质量。

Authors: Jiacheng Liu, Shengkun Tang, Jiacheng Cui, Dongkuan Xu, Zhiqiang Shen | Date: 2026-03-12

Link: http://arxiv.org/abs/2603.12240v1

Abstract Acceleration methods for diffusion models (e.g., token merging or downsampling) typically optimize synthesis quality under reduced compute, yet often ignore discriminative capacity. We revisit token compression with a joint objective and present BiGain, a training-free, plug-and-play framework that preserves generation quality while improving classification in accelerated diffusion models. Our key insight is frequency separation: mapping feature-space signals into a frequency-aware representation disentangles fine detail from global semantics, enabling compression that respects both generative fidelity and discriminative utility. BiGain reflects this principle with two frequency-aware operators: (1) Laplacian-gated token merging, which encourages merges among spectrally smooth tokens while discouraging merges of high-contrast tokens, thereby retaining edges and textures; and (2) Interpolate-Extrapolate KV Downsampling, which downsamples keys/values via a controllable interextrapolation between nearest and average pooling while keeping queries intact, thereby conserving attention precision. Across DiT- and U-Net-based backbones and ImageNet-1K, ImageNet-100, Oxford-IIIT Pets, and COCO-2017, our operators consistently improve the speed-accuracy trade-off for diffusion-based classification, while maintaining or enhancing generation quality under comparable acceleration. For instance, on ImageNet-1K, with 70% token merging on Stable Diffusion 2.0, BiGain increases classification accuracy by 7.15% while improving FID by 0.34 (1.85%). Our analyses indicate that balanced spectral retention, preserving high-frequency detail and low/mid-frequency semantics, is a reliable design rule for token compression in diffusion models. To our knowledge, BiGain is the first framework to jointly study and advance both generation and classification under accelerated diffusion, supporting lower-cost deployment.