arXiv subjects

Junfeng Liu

Publications and source records attributed to Junfeng Liu.

At least 19 recordsLinked to original sources

arXiv2026

What Is a Skill Worth? Structure-Aware Shapley Valuation of Agent Skills

Agent skills are increasingly optimized by automated feedback loops, producing long structured artifacts whose internal value remains unclear. We study skill valuation: assigning credit to the internal units of a fixed skill, such as rules, examples, scripts, and heuristics, under a fixed agent and held-out task distribution. Skill valuation differs from data or prompt-span valuation because skill units are structured: they may depend on other units, belong to a document hierarchy, trigger agent behavior, and consume limited prompt context. We introduce SkillSV, a structure-aware Shapley-style framework for skill valuation. SkillSV compiles a skill into units, dependencies, and hierarchy, so that only valid counterfactual skills are evaluated. It uses paired deletion and length-neutral padding to separate content value from context cost, and estimates the resulting values with a rollout-budgeted estimator for noisy agent evaluations. On four agentic benchmarks, we assess the faithfulness, actionability, and explanation of SkillSV: it recovers unit interactions, preserves aggregate skill lift, and guides safe pruning and compression.

cs.AI↗

arXiv2026

GRAG: Generic Response-Augmented Generation Framework for Personalized Conversational Systems

Deploying highly capable personalized conversational agents in resource-constrained or privacy-sensitive environments remains a significant challenge. We identify a fundamental bottleneck in the existing approaches: current training paradigms treat personalization and grounding as a single monolithic learning problem. Under these paradigms, language models are forced to simultaneously address what to say (content grounding) and how to say it in a user-specific way (personalization), which introduces significant computational and optimization challenges. Consequently, contextual grounding is often sacrificed for persona adherence, or vice versa, resulting in responses that are either weakly grounded in the conversational history or insufficiently personalized. In this work, we propose the Generic Response-Augmented Generation (GRAG) framework that decouples these competing objectives by leveraging offline, generic responses from high-capacity, general-purpose LLMs as a semantic and structural scaffold to guide the fine-tuning of smaller, task-specialized models seamlessly in resource-limited environments. By decoupling the content grounding from personalization, GRAG allows the model to focus exclusively on persona injection while remaining firmly anchored to the conversational context. We instantiate the GRAG in two post- and pre-fusion-based architectural variants and evaluate them on multiple benchmark conversational datasets that cover diverse personalization structures. Our results demonstrate that GRAG significantly outperforms state-of-the-art methods that do not use auxiliary scaffolding, yielding up to 47% improvements in ROUGE-2 and 36% in BLEU scores. Ultimately, GRAG offers a generalizable blueprint for building grounding-aware personalized conversational systems in resource-limited environments.

cs.CL↗

arXiv2026

SenseJudge: Human-Centric Preference-Driven Judgment Framework

Large Language Models (LLMs) as judges across various scenarios such as assessing model responses is becoming an increasingly accepted paradigm. However, existing judgment approaches often rely on trained judgers using fixed preference data, which tend to overlook diverse user preferences and struggle to adapt to real-world human-AI dialogue scenarios. To address these limitations, we propose SenseJudge, a customizable judgment framework driven by human preferences and SenseBench, a diverse and challenging instruction-following benchmark derived from real-world multi-turn interactions. We applied the automatic judgment framework and benchmark to two tasks: (1) LLMs as personalized judges, and (2) model ranking. We conducted extensive experiments, and the results demonstrate that the SenseJudge framework surpasses other judgment methods and models in the LLMs-as-personalized-judges task and achieves model ranking that aligns with real human sense. Additionally, we conducted analyses on position bias and consistency, alongside ablation studies, which affirmed the robustness of SenseJudge.

cs.CL↗

arXiv2026

Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters

We introduce Step 3.5 Flash, a sparse Mixture-of-Experts (MoE) model that bridges frontier-level agentic intelligence and computational efficiency. We focus on what matters most when building agents: sharp reasoning and fast, reliable execution. Step 3.5 Flash pairs a 196B-parameter foundation with 11B active parameters for efficient inference. It is optimized with interleaved 3:1 sliding-window/full attention and Multi-Token Prediction (MTP-3) to reduce the latency and cost of multi-round agentic interactions. To reach frontier-level intelligence, we design a scalable reinforcement learning framework that combines verifiable signals with preference feedback, while remaining stable under large-scale off-policy training, enabling consistent self-improvement across mathematics, code, and tool use. Step 3.5 Flash demonstrates strong performance across agent, coding, and math tasks, achieving 85.4% on IMO-AnswerBench, 86.4% on LiveCodeBench-v6 (2024.08-2025.05), 88.2% on tau2-Bench, 69.0% on BrowseComp (with context management), and 51.0% on Terminal-Bench 2.0, comparable to frontier models such as GPT-5.2 xHigh and Gemini 3.0 Pro. By redefining the efficiency frontier, Step 3.5 Flash provides a high-density foundation for deploying sophisticated agents in real-world industrial environments.

cs.CL↗

arXiv2026

STEP3-VL-10B Technical Report

We present STEP3-VL-10B, a lightweight open-source foundation model designed to redefine the trade-off between compact efficiency and frontier-level multimodal intelligence. STEP3-VL-10B is realized through two strategic shifts: first, a unified, fully unfrozen pre-training strategy on 1.2T multimodal tokens that integrates a language-aligned Perception Encoder with a Qwen3-8B decoder to establish intrinsic vision-language synergy; and second, a scaled post-training pipeline featuring over 1k iterations of reinforcement learning. Crucially, we implement Parallel Coordinated Reasoning (PaCoRe) to scale test-time compute, allocating resources to scalable perceptual reasoning that explores and synthesizes diverse visual hypotheses. Consequently, despite its compact 10B footprint, STEP3-VL-10B rivals or surpasses models 10$\times$-20$\times$ larger (e.g., GLM-4.6V-106B, Qwen3-VL-235B) and top-tier proprietary flagships like Gemini 2.5 Pro and Seed-1.5-VL. Delivering best-in-class performance, it records 92.2% on MMBench and 80.11% on MMMU, while excelling in complex reasoning with 94.43% on AIME2025 and 75.95% on MathVision. We release the full model suite to provide the community with a powerful, efficient, and reproducible baseline.

cs.CV↗

arXiv2026

RadixGraph: A Fast, Space-Optimized Data Structure for Dynamic Graph Storage (Extended Version)

Dynamic graphs model many real-world applications, and as their sizes grow, efficiently storing and updating them becomes critical. We present RadixGraph, a fast and memory-efficient data structure for dynamic graph storage. RadixGraph features a carefully designed radix-tree-based vertex index that strikes an optimal trade-off between query efficiency and space among all pointer-array-based radix trees. For edge storage, it employs a hybrid snapshot-log architecture that enables amortized $O(1)$ update time. RadixGraph supports millions of concurrent updates per second while maintaining competitive performance for graph analytics. Experimental results show that RadixGraph outperforms the most performant baseline by up to $16.27\times$ across various datasets in ingesting graph updates, and reduces memory usage by an average of $40.1\%$. RadixGraph is open-source at https://github.com/ForwardStar/RadixGraph.

cs.DB↗

arXiv2025

Bounded Dynamic Level Maintenance for Efficient Logic Optimization

Logic optimization constitutes a critical phase within the Electronic Design Automation (EDA) flow, essential for achieving desired circuit power, performance, and area (PPA) targets. These logic circuits are typically represented as Directed Acyclic Graphs (DAGs), where the structural depth, quantified by node level, critically correlates with timing performance. Modern optimization strategies frequently employ iterative, local transformation heuristics (\emph{e.g.,} \emph{rewrite}, \emph{refactor}) directly on this DAG structure. As optimization continuously modifies the graph locally, node levels require frequent dynamic updates to guide subsequent decisions. However, a significant gap exists: existing algorithms for incrementally updating node levels are unbounded to small changes. This leads to a total of worst complexity in $O(|V|^2)$ for given local subgraphs $\{\Delta G_i\}_{i=1}^{|V|}$ updates on DAG $G(V,E)$. This unbounded nature poses a severe efficiency bottleneck, hindering the scalability of optimization flows, particularly when applied to large circuit designs prevalent today. In this paper, we analyze the dynamic level maintenance problem endemic to iterative logic optimization, framing it through the lens of partial topological order. Building upon the analysis, we present the first bounded algorithm for maintaining level constraints, with $O(|V| \Delta \log \Delta)$ time for a sequence $|V|$ of updates $\{\Delta G_i\}$, where $\Delta = \max_i \|\Delta G_i\|$ denotes the maximum extended size of $\Delta G_i$. Experiments on comprehensive benchmarks show our algorithm enables an average 6.4$\times$ overall speedup relative to \rw and \rf, driven by a 1074.8$\times$ speedup in the level maintenance, all without any quality sacrifice.

cs.CC↗

arXiv2025

SetAD: Semi-Supervised Anomaly Learning in Contextual Sets

Semi-supervised anomaly detection (AD) has shown great promise by effectively leveraging limited labeled data. However, existing methods are typically structured around scoring individual points or simple pairs. Such {point- or pair-centric} view not only overlooks the contextual nature of anomalies, which are defined by their deviation from a collective group, but also fails to exploit the rich supervisory signals that can be generated from the combinatorial composition of sets. Consequently, such models struggle to exploit the high-order interactions within the data, which are critical for learning discriminative representations. To address these limitations, we propose SetAD, a novel framework that reframes semi-supervised AD as a Set-level Anomaly Detection task. SetAD employs an attention-based set encoder trained via a graded learning objective, where the model learns to quantify the degree of anomalousness within an entire set. This approach directly models the complex group-level interactions that define anomalies. Furthermore, to enhance robustness and score calibration, we propose a context-calibrated anomaly scoring mechanism, which assesses a point's anomaly score by aggregating its normalized deviations from peer behavior across multiple, diverse contextual sets. Extensive experiments on 10 real-world datasets demonstrate that SetAD significantly outperforms state-of-the-art models. Notably, we show that our model's performance consistently improves with increasing set size, providing strong empirical support for the set-based formulation of anomaly detection.

cs.LG↗

arXiv2025

On 10x Better Scalability: KV Stores Scale Up KV Cache

Large language models (LLMs) rely on Key-Value (KV) cache to reduce time-to-first-token (TTFT) latency, but existing disk-based KV cache systems using file-per-object layouts suffer from severe scalability bottlenecks due to file system metadata overhead, I/O inefficiency, and poor spatial locality. This paper presents SGLANG-LSM, a database-inspired system that leverages Log-Structured Merge-tree (LSM-tree) architectures for scalable KV cache management. SGLANG-LSM implements a layered system design with three coordinated components: (1) a prefix-preserving storage engine that maintains token sequence locality while efficiently storing large KV cache tensors through key-value separation, (2) an adaptive controller that dynamically optimizes LSM-tree configurations based on shifting workload characteristics, and (3) runtime services including batch operations and automatic resource management for production deployment. Evaluation on large-scale dynamic workloads demonstrates that SGLANG-LSM significantly improves cache hits by up to 143% and reduces TTFT by up to 24% compared to state-of-the-art systems, representing the first systematic application of database storage architectures to large-scale LLM cache management.

cs.DB↗

arXiv2025

BoolSkeleton: Boolean Network Skeletonization via Homogeneous Pattern Reduction

Boolean equivalence allows Boolean networks with identical functionality to exhibit diverse graph structures. This gives more room for exploration in logic optimization, while also posing a challenge for tasks involving consistency between Boolean networks. To tackle this challenge, we introduce BoolSkeleton, a novel Boolean network skeletonization method that improves the consistency and reliability of design-specific evaluations. BoolSkeleton comprises two key steps: preprocessing and reduction. In preprocessing, the Boolean network is transformed into a defined Boolean dependency graph, where nodes are assigned the functionality-related status. Next, the homogeneous and heterogeneous patterns are defined for the node-level pattern reduction step. Heterogeneous patterns are preserved to maintain critical functionality-related dependencies, while homogeneous patterns can be reduced. Parameter K of the pattern further constrains the fanin size of these patterns, enabling fine-tuned control over the granularity of graph reduction. To validate BoolSkeleton's effectiveness, we conducted four analysis/downstream tasks around the Boolean network: compression analysis, classification, critical path analysis, and timing prediction, demonstrating its robustness across diverse scenarios. Furthermore, it improves above 55% in the average accuracy compared to the original Boolean network for the timing prediction task. These experiments underscore the potential of BoolSkeleton to enhance design consistency in logic synthesis.

cs.AR↗

arXiv2025

WebRouter: Query-specific Router via Variational Information Bottleneck for Cost-sensitive Web Agent

LLM-brained web agents offer powerful capabilities for web automation but face a critical cost-performance trade-off. The challenge is amplified by web agents' inherently complex prompts that include goals, action histories, and environmental states, leading to degraded LLM ensemble performance. To address this, we introduce WebRouter, a novel query-specific router trained from an information-theoretic perspective. Our core contribution is a cost-aware Variational Information Bottleneck (ca-VIB) objective, which learns a compressed representation of the input prompt while explicitly penalizing the expected operational cost. Experiments on five real-world websites from the WebVoyager benchmark show that WebRouter reduces operational costs by a striking 87.8\% compared to a GPT-4o baseline, while incurring only a 3.8\% accuracy drop.

cs.CL↗

arXiv2025

[Extended Version] ArceKV: Towards Workload-driven LSM-compactions for Key-Value Store Under Dynamic Workloads

Key-value stores underpin a wide range of applications due to their simplicity and efficiency. Log-Structured Merge Trees (LSM-trees) dominate as their underlying structure, excelling at handling rapidly growing data. Recent research has focused on optimizing LSM-tree performance under static workloads with fixed read-write ratios. However, real-world workloads are highly dynamic, and existing workload-aware approaches often struggle to sustain optimal performance or incur substantial transition overhead when workload patterns shift. To address this, we propose ElasticLSM, which removes traditional LSM-tree structural constraints to allow more flexible management actions (i.e., compactions and write stalls) creating greater opportunities for continuous performance optimization. We further design Arce, a lightweight compaction decision engine that guides ElasticLSM in selecting the optimal action from its expanded action space. Building on these components, we implement ArceKV, a full-fledged key-value store atop RocksDB. Extensive evaluations demonstrate that ArceKV outperforms state-of-the-art compaction strategies across diverse workloads, delivering around 3x faster performance in dynamic scenarios.

cs.DB↗

arXiv2025

Step-3 is Large yet Affordable: Model-system Co-design for Cost-effective Decoding

Large language models (LLMs) face low hardware efficiency during decoding, especially for long-context reasoning tasks. This paper introduces Step-3, a 321B-parameter VLM with hardware-aware model-system co-design optimized for minimizing decoding costs. Step-3 innovates in two key dimensions: (1) A novel Multi-Matrix Factorization Attention (MFA) mechanism that significantly reduces both KV cache size and computation while maintaining high attention expressiveness, and (2) Attention-FFN Disaggregation (AFD), a distributed inference system that decouples attention and Feed-Forward Network (FFN) layers into specialized subsystems. This co-design achieves unprecedented cost efficiency: Step-3 significantly reduces theoretical decoding costs compared with models like DeepSeek-V3 and Qwen3 MoE 235B, with the gains widening at longer context. Step-3 achieves low cost while activating 38B parameters per token (more than DeepSeek-V3 and Qwen3 MoE 235B), demonstrating that hardware-aligned attention arithmetic intensity, MoE sparsity, and AFD are critical to cost-effectiveness. We perform a head-to-head comparison with DeepSeek-V3 in its favorable scenarios. Our implementation on Hopper GPUs achieves a decoding throughput of up to 4,039 tokens per second per GPU under 50ms TPOT SLA (4K context, FP8, no MTP). It is higher than DeepSeek-V3's 2,324 in the same setup and sets a new Pareto frontier for LLM decoding.

cs.LG↗

arXiv2025

Evaluating Learned Indexes in LSM-tree Systems: Benchmarks,Insights and Design Choices

LSM-tree-based data stores are widely used in industry due to their exceptional performance. However, as data volumes grow, efficiently querying large-scale databases becomes increasingly challenging. To address this, recent studies attempted to integrate learned indexes into LSM-trees to enhance lookup performance, which has demonstrated promising improvements. Despite this, only a limited range of learned index types has been considered, and the strengths and weaknesses of different learned indexes remain unclear, making them difficult for practical use. To fill this gap, we provide a comprehensive and systematic benchmark to pursue an in-depth understanding of learned indexes in LSM-tree systems. In this work, we summarize the workflow of 8 existing learned indexes and analyze the associated theoretical cost. We also identify several key factors that significantly influence the performance of learned indexes and conclude them with a novel configuration space, including various index types, boundary positions, and granularity. Moreover, we implement different learned index designs on a unified platform to evaluate across various configurations. Surprisingly, our experiments reveal several unexpected insights, such as the marginal lookup enhancement when allocating a large memory budget to learned indexes and modest retraining overhead of learned indexes. Besides, we also offer practical guidelines to help developers intelligently select and tune learned indexes for custom use cases.

cs.DB↗

arXiv2025

HauntAttack: When Attack Follows Reasoning as a Shadow

Emerging Large Reasoning Models (LRMs) consistently excel in mathematical and reasoning tasks, showcasing remarkable capabilities. However, the enhancement of reasoning abilities and the exposure of internal reasoning processes introduce new safety vulnerabilities. A critical question arises: when reasoning becomes intertwined with harmfulness, will LRMs become more vulnerable to jailbreaks in reasoning mode? To investigate this, we introduce HauntAttack, a novel and general-purpose black-box adversarial attack framework that systematically embeds harmful instructions into reasoning questions. Specifically, we modify key reasoning conditions in existing questions with harmful instructions, thereby constructing a reasoning pathway that guides the model step by step toward unsafe outputs. We evaluate HauntAttack on 11 LRMs and observe an average attack success rate of over 70\%, achieving up to 13 percentage points of absolute improvement over the strongest prior baseline. Our further analysis reveals that even advanced safety-aligned models remain highly susceptible to reasoning-based attacks, offering insights into the urgent challenge of balancing reasoning capability and safety in future model development.

cs.CR↗

arXiv2025

RAGDoll: Efficient Offloading-based Online RAG System on a Single GPU

Retrieval-Augmented Generation (RAG) enhances large language model (LLM) generation quality by incorporating relevant external knowledge. However, deploying RAG on consumer-grade platforms is challenging due to limited memory and the increasing scale of both models and knowledge bases. In this work, we introduce RAGDoll, a resource-efficient, self-adaptive RAG serving system integrated with LLMs, specifically designed for resource-constrained platforms. RAGDoll exploits the insight that RAG retrieval and LLM generation impose different computational and memory demands, which in a traditional serial workflow result in substantial idle times and poor resource utilization. Based on this insight, RAGDoll decouples retrieval and generation into parallel pipelines, incorporating joint memory placement and dynamic batch scheduling strategies to optimize resource usage across diverse hardware devices and workloads. Extensive experiments demonstrate that RAGDoll adapts effectively to various hardware configurations and LLM scales, achieving up to 3.6 times speedup in average latency compared to serial RAG systems based on vLLM.

cs.DC↗

arXiv2025

CoMAC: Conversational Agent for Multi-Source Auxiliary Context with Sparse and Symmetric Latent Interactions

Recent advancements in AI-driven conversational agents have exhibited immense potential of AI applications. Effective response generation is crucial to the success of these agents. While extensive research has focused on leveraging multiple auxiliary data sources (e.g., knowledge bases and personas) to enhance response generation, existing methods often struggle to efficiently extract relevant information from these sources. There are still clear limitations in the ability to combine versatile conversational capabilities with adherence to known facts and adaptation to large variations in user preferences and belief systems, which continues to hinder the wide adoption of conversational AI tools. This paper introduces a novel method, Conversational Agent for Multi-Source Auxiliary Context with Sparse and Symmetric Latent Interactions (CoMAC), for conversation generation, which employs specialized encoding streams and post-fusion grounding networks for multiple data sources to identify relevant persona and knowledge information for the conversation. CoMAC also leverages a novel text similarity metric that allows bi-directional information sharing among multiple sources and focuses on a selective subset of meaningful words. Our experiments show that CoMAC improves the relevant persona and knowledge prediction accuracies and response generation quality significantly over two state-of-the-art methods.

cs.CL↗

arXiv2025

Aster: Enhancing LSM-structures for Scalable Graph Database

There is a proliferation of applications requiring the management of large-scale, evolving graphs under workloads with intensive graph updates and lookups. Driven by this challenge, we introduce Poly-LSM, a high-performance key-value storage engine for graphs with the following novel techniques: (1) Poly-LSM is embedded with a new design of graph-oriented LSM-tree structure that features a hybrid storage model for concisely and effectively storing graph data. (2) Poly-LSM utilizes an adaptive mechanism to handle edge insertions and deletions on graphs with optimized I/O efficiency. (3) Poly-LSM exploits the skewness of graph data to encode the key-value entries. Building upon this foundation, we further implement Aster, a robust and versatile graph database that supports Gremlin query language facilitating various graph applications. In our experiments, we compared Aster against several mainstream real-world graph databases. The results demonstrate that Aster outperforms all baseline graph databases, especially on large-scale graphs. Notably, on the billion-scale Twitter graph dataset, Aster achieves up to 17x throughput improvement compared to the best-performing baseline graph system.

cs.DB↗