Searcharxiv⌕ Search

arXiv subjects

Xiang Song

Publications and source records attributed to Xiang Song.

At least 19 recordsLinked to original sources
arXiv2026

BPG: Balancing Plasticity and Generalization for Domain Incremental Learning

Deep neural networks excel in various tasks but struggle to generalize across evolving data distributions, leading to significant performance degradation under domain shifts. Domain incremental learning (DIL) addresses this challenge by enabling models to continuously adapt while retaining prior knowledge. Among existing DIL approaches, the parameter-isolation paradigm achieves state-of-the-art performance. However, these methods often adopt a one-size-fits-all approach to adapt to new domains, resulting in either insufficient learning capacity or redundant parameters. In this work, we propose BPG, a unified framework that addresses both challenges through two complementary components: BPG-Adapter, which dynamically determines each domain's adapter hidden dimension based on domain-specific feature separability, and BPG-Inference, a soft domain mixture strategy that integrates multiple domain-specific models at test time, mitigating domain ID misselection. Experimental results on DomainNet, CDDB, and CORe50 demonstrate that BPG consistently outperforms uniform adapter-based approaches and hard domain selection strategies, achieving state-of-the-art average accuracy while reducing forgetting to as low as 0.22% on DomainNet.

cs.CV↗
arXiv2026

Toward an Organizational Science of Multi-Agent LLM Systems: Decoupling Who, How, and Which Algorithm

Multi-agent frameworks built on large language models (LLMs) routinely entangle three logically distinct concerns: who is on the team (organization), how members align (coordination), and which algorithm fuses their work (collaboration protocol). IMACS (Intelligent Multi-Agent Collaboration System) separates the three into orthogonal, independently swappable layers. Classic organizational theory (Belbin roles, Mintzberg coordination, RACI accountability) becomes executable, validated configuration, and the framework places six published collaboration algorithms behind a common interface while exposing roles, coordination, and accountability as independently configurable factors. We use this separation to conduct controlled comparisons in which organizational assignments vary while the collaboration protocol is held fixed. It also turns protocol choice into a variable that can be learned: Adaptive Org Routing, a contextual-bandit meta-protocol, selects a protocol per task under an explicit quality-cost tradeoff, outperforms every fixed protocol in a controlled study, and trains online on real benchmark and LLM-judge rewards. The ablations expose a mechanism. Accountability placement changes outcomes exactly when the protocol routes the deliverable through the accountable agent, and the winning placement flips across model families, so organizational design cannot be hard-coded; it must be revalidated, or learned, for each model binding.

cs.AI↗
arXiv2026

OwlPath: Lossless Knowledge Compression for LLM Bug Repair

LLM-based software engineering agents are constrained by limited context windows: roughly 100K tokens must store structurally relevant code subsets to resolve bugs. Standard retrieval models treat code as plain text, forcing agents to resolve multi-hop dependencies including subclass chains, transitive callers and interface implementations through slow trial and error. We tackle this limitation with lossless knowledge compression, encoding source code into an OWL2 ontology to answer structural queries using minimal relevant code fragments. We present OwlPath, an OWL2 reasoning layer atop CodeGraph, a widely used code intelligence platform with 500K+ GitHub stars, offering a unified CLI for structural code retrieval. Powered by tree-sitter parsing, OwlPath supports multi-language repositories (Python, JavaScript, TypeScript, Go, etc.) and encodes language-specific semantics into a unified OWL2 ontology. It adopts two complementary modules. First, a transitive-closure engine fetches all structurally linked symbols via single SPARQL property-path queries, capturing multi-hop relations missed by string matching. Second, the OWL Software Knowledge Map (OWL-SKM) precomputes a compact 3KB summary with module trees, core APIs and issue-related symbols, directing agents to target modules in the first query. Evaluated on 18 SWE-bench Pro instances, OwlPath obtains a 68.4% strict-apply rate versus 66.7% for the CodeGraph baseline, cutting token usage by 28.8% and runtime by 39.5%. In offline retrieval tests over 67 instances, OwlPath improves recall 2.06 times (0.464 vs 0.226) and reaches 88.1% hit rate compared to CodeGraph's 59.7%. On a 37-question structural retrieval benchmark, recall rises from 4.4% to 28.8%, with 69-80% accuracy on transitive caller and interface tasks.

cs.SE↗
arXiv2026

KernelSight-LM: A Kernel-Level LLM Inference Simulator

As large language models (LLMs) move into production serving, practitioners must rapidly evaluate inference performance across diverse hardware, models, and serving parameters to meet cost and latency targets. However, the end-to-end behavior of LLMs couples serving-layer policies with low-level GPU kernel execution and rapidly evolving architectures, forcing slow, deployment-specific benchmarking that is hard to generalize. We present KernelSight-LM, a fine-grained inference simulator that models token-level execution and produces kernel-level latency breakdowns. It decomposes each serving step into a roofline kernel model with a learned efficiency term, a communication model, and a host-overhead model, composed through a discrete-event scheduler that also captures mechanisms like prefix caching and continuous batching. KernelSight-LM offers two prediction tiers that trade target-GPU data for accuracy. The cross-generation tier uses no target-GPU measurements, only hardware specifications and kernel microbenchmarks from previously profiled GPUs, and predicts per-kernel latency on an unseen GPU generation to 12.1% error, a 1.8x improvement over the roofline baseline (22.0%). A second target-measured tier adds one model-agnostic kernel-microbenchmark sweep on the target GPU, sharpening per-kernel error to 3.8%, a 7.3x improvement over a comparable baseline (27.7%). Both tiers require far less target-GPU data than the prior systems they extend. In our simulator, these predictions yield end-to-end median (p50) errors across six model families of 15.4%, 12.8%, and 3.0% (TTFT, TPOT, throughput) in the cross-generation tier and 14.3%, 6.2%, and 2.7% in the target-measured tier, matching dedicated profiling tools while collecting far less on-device data. Beyond prediction, its kernel-level bottleneck breakdowns support hardware/software co-design and capacity planning.

cs.PF↗
arXiv2026

NetCause: Counterfactual Learning for Root Cause Analysis in Large-Scale Networks

Can a learned model capture how faults propagate through a large-scale network and use this knowledge to causally attribute customer impact to its underlying root cause? Existing root cause analysis techniques often rely on static rules, correlation heuristics, or topology-local reasoning, which struggle to generalize in dynamic environments where faults propagate across complex physical and logical dependencies. We present NetCause, a self-supervised learning-based framework that models network incidents as graph-temporal processes and uses counterfactual simulation to rank candidate root causes. This approach produces an interpretable ranking of root cause hypotheses and integrates naturally with operator-defined mitigation and remediation actions. We train the model on over 1,500 incidents collected over six months from a leading cloud provider's production network and evaluate it on 31 expert-labeled incidents. NetCause consistently improves root cause ranking quality in the regime most relevant to operational decision-making, achieving a 16.1% accuracy improvement over a rule-based heuristic baseline. While training is computationally intensive, inference is lightweight, requiring only seconds of GPU runtime per incident (well below typical telemetry collection latencies).

cs.NI↗
arXiv2026

Trace2Policy: From Expert Behavior Traces to Self-Evolving Decision Agents

Decision rules that enterprise experts apply tacitly -- in auditing, compliance, and contract review -- can be systematically recovered and improved through iterative error analysis. We present \textbf{Trace2Policy}, whose core mechanism -- \textbf{EISR} (\textbf{E}rror-driven \textbf{I}terative \textbf{S}kill \textbf{R}efinement) -- maintains a human-readable rule document as its optimization target: each round executes the rules on a validation set, clusters errors by root cause into MISSING, WRONG, or CONFLICT types, applies targeted patches, and commits only those that pass a regression gate. \textbf{For this class of compliance-sensitive, skewed-base-rate decision tasks, we identify rule quality -- not model capability -- as the dominant performance lever}: across five LLMs, one-shot distillation plateaus near $\sim$70\% on the deployed pool, while eight EISR rounds lift the same rules to 79.6\% when compiled into deterministic Python -- zero LLM calls at inference. \textbf{Execution form compounds the gain: in production, the same EISR-refined content runs 9.8~pp higher as compiled Python than as an LLM prompt, a form-and-engineering bundle the 22-day deployment matured together.} Deployed for 22 days at a major logistics carrier (3,349 audit cases), the compiled pipeline outperforms the pure-LLM baseline it replaced (72.7\%); on these calibrated, skewed-base-rate workloads, re-enabling LLM fallback monotonically degrades accuracy. An LLM-driven variant, \textbf{Auto-EISR}, reproduces this refinement at \$5--\$10 per cycle versus $\sim$70 expert-hours, and transfers to four public benchmarks spanning legal reasoning (LegalBench) and process-mining decisions (BPIC 2012) without re-engineering.

cs.AI↗
arXiv2026

DualKV: Shared-Prompt Flash Attention for Efficient RL Training with Large Rollouts and Long Contexts

Modern RL post-training methods such as GRPO and DAPO train on N response sequences of R tokens sampled from a shared prompt of P tokens, but standard FlashAttention replicates all P prompt tokens N times across both forward and backward passes -- duplicating compute and memory on identical hidden states. In large-rollout, long-context RL training (N>=16, P>=8K), this redundancy dominates the policy update cost. We observe that in decoder-only models, causal masking makes prompt representations invariant across sequences at every layer, so all per-token operations (norms, projections, MLP) and attention can process the prompt once -- a property not yet exploited at the kernel level for training. We propose DualKV, the first FlashAttention kernel variant that eliminates shared-prompt replication during RL training, via (1) fused CUDA forward and backward kernels that iterate over two disjoint KV regions -- shared context and per-sequence response -- in a single kernel launch, and (2) a data-pipeline redesign in veRL that repacks N(P+R) tokens into P+NR tokens per micro-batch, extending the token reduction from attention to the entire model by a factor rho = N(P+R)/(P+NR). DualKV is mathematically equivalent to standard attention and introduces no approximation. On Qwen3-8B GRPO training with 8xH100 GPUs (N=32, 8K-context), DualKV achieves 1.63--2.09x policy-update speedup, enables 2x larger micro-batches, and raises MFU from 36% to 76%. Similar gains hold for DAPO (2.47x speedup, 77% MFU). At 30B MoE scale on 16xH100, DualKV achieves 3.82x policy-update and 3.38x end-to-end step speedup over FlashAttention (which requires 4-way Ulysses sequence parallelism to avoid OOM). DualKV also extends to hybrid sliding/global attention with head dimension 512 (which FA2 does not support) and integrates with Ulysses sequence parallelism, demonstrated on Gemma-4-31B GRPO at 64K context.

cs.LG↗
arXiv2026

Relatron: Automating Relational Machine Learning over Relational Databases

Predictive modeling over relational databases (RDBs) powers applications, yet remains challenging due to capturing both cross-table dependencies and complex feature interactions. Relational Deep Learning (RDL) methods automate feature engineering via message passing, while classical approaches like Deep Feature Synthesis (DFS) rely on predefined non-parametric aggregators. Despite performance gains, the comparative advantages of RDL over DFS and the design principles for selecting effective architectures remain poorly understood. We present a comprehensive study that unifies RDL and DFS in a shared design space and conducts architecture-centric searches across diverse RDB tasks. Our analysis yields three key findings: (1) RDL does not consistently outperform DFS, with performance being highly task-dependent; (2) no single architecture dominates across tasks, underscoring the need for task-aware model selection; and (3) validation accuracy is an unreliable guide for architecture choice. This search yields a model performance bank that links architecture configurations to their performance; leveraging this bank, we analyze the drivers of the RDL-DFS performance gap and introduce two task signals -- RDB task homophily and an affinity embedding that captures size, path, feature, and temporal structure -- whose correlation with the gap enables principled routing. Guided by these signals, we propose Relatron, a task embedding-based meta-selector that chooses between RDL and DFS and prunes the within-family search. Lightweight loss-landscape metrics further guard against brittle checkpoints by preferring flatter optima. In experiments, Relatron resolves the "more tuning, worse performance" effect and, in joint hyperparameter-architecture optimization, achieves up to 18.5% improvement over strong baselines with 10x lower cost than Fisher information-based alternatives.

cs.LG↗
arXiv2026

XShare: Collaborative in-Batch Expert Sharing for Faster MoE Inference

Mixture-of-Experts (MoE) architectures are increasingly used to efficiently scale large language models. However, in production inference, request batching and speculative decoding significantly amplify expert activation, eroding these efficiency benefits. We address this issue by modeling batch-aware expert selection as a modular optimization problem and designing efficient greedy algorithms for different deployment settings. The proposed method, namely XShare, requires no retraining and dynamically adapts to each batch by maximizing the total gating score of selected experts. It reduces expert activation by up to 30% under standard batching, cuts peak GPU load by up to 3x in expert-parallel deployments, and achieves up to 14% throughput gains in speculative decoding via hierarchical, correlation-aware expert selection even if requests in a batch drawn from heterogeneous datasets.

cs.LG↗
arXiv2026

Feedback Control for Multi-Objective Graph Self-Supervision

Can multi-task self-supervised learning on graphs be coordinated without the usual tug-of-war between objectives? Graph self-supervised learning (SSL) offers a growing toolbox of pretext objectives like mutual information, reconstruction, and contrastive learning, yet combining them reliably remains challenging due to objective interference and training instability. Most multi-pretext pipelines use per-update mixing, forcing every parameter update to be a compromise and leading to three failure modes: Disagreement (conflict-induced negative transfer), Drift (nonstationary objective utility), and Drought (hidden starvation of underserved objectives). We argue that coordination is fundamentally a temporal allocation problem: deciding when each objective receives optimization budget, not merely how to weigh them. We introduce ControlG, a control-theoretic framework that recasts multi-objective graph SSL as feedback-controlled temporal allocation by estimating per-objective difficulty and pairwise antagonism, planning target budgets via a Pareto-aware log-hypervolume planner, and scheduling with a Proportional-Integral-Derivative (PID) controller. Across 9 datasets, ControlG consistently outperforms state-of-the-art baselines while producing an auditable schedule that reveals which objectives drove learning.

cs.LG↗
arXiv2026

P-EAGLE: Parallel-Drafting EAGLE with Scalable Training

Reasoning LLMs produce longer outputs, requiring speculative decoding drafters trained on extended sequences. Parallel drafting - predicting multiple tokens per forward pass - offers latency benefits over sequential generation, but training complexity scales quadratically with the product of sequence length and parallel positions, rendering long-context training impractical. We present P(arallel)-EAGLE, which transforms EAGLE from autoregressive to parallel multi-token prediction via a learnable shared hidden state. To scale training to long contexts, we develop a framework featuring attention mask pre-computation and sequence partitioning techniques, enabling gradient accumulation within individual sequences for parallel-prediction training. We implement P-EAGLE in vLLM and demonstrate speedups of 1.10-1.36x over autoregressive EAGLE-3 across GPT-OSS 120B, 20B, and Qwen3-Coder 30B.

cs.LG↗
arXiv2025

GRIL: Knowledge Graph Retrieval-Integrated Learning with Large Language Models

Retrieval-Augmented Generation (RAG) has significantly mitigated the hallucinations of Large Language Models (LLMs) by grounding the generation with external knowledge. Recent extensions of RAG to graph-based retrieval offer a promising direction, leveraging the structural knowledge for multi-hop reasoning. However, existing graph RAG typically decouples retrieval and reasoning processes, which prevents the retriever from adapting to the reasoning needs of the LLM. They also struggle with scalability when performing multi-hop expansion over large-scale graphs, or depend heavily on annotated ground-truth entities, which are often unavailable in open-domain settings. To address these challenges, we propose a novel graph retriever trained end-to-end with LLM, which features an attention-based growing and pruning mechanism, adaptively navigating multi-hop relevant entities while filtering out noise. Within the extracted subgraph, structural knowledge and semantic features are encoded via soft tokens and the verbalized graph, respectively, which are infused into the LLM together, thereby enhancing its reasoning capability and facilitating interactive joint training of the graph retriever and the LLM reasoner. Experimental results across three QA benchmarks show that our approach consistently achieves state-of-the-art performance, validating the strength of joint graph-LLM optimization for complex reasoning tasks. Notably, our framework eliminates the need for predefined ground-truth entities by directly optimizing the retriever using LLM logits as implicit feedback, making it especially effective in open-domain settings.

cs.LG↗
arXiv2025

Dynamic Mixture-of-Experts for Incremental Graph Learning

Graph incremental learning is a learning paradigm that aims to adapt trained models to continuously incremented graphs and data over time without the need for retraining on the full dataset. However, regular graph machine learning methods suffer from catastrophic forgetting when applied to incremental learning settings, where previously learned knowledge is overridden by new knowledge. Previous approaches have tried to address this by treating the previously trained model as an inseparable unit and using techniques to maintain old behaviors while learning new knowledge. These approaches, however, do not account for the fact that previously acquired knowledge at different timestamps contributes differently to learning new tasks. Some prior patterns can be transferred to help learn new data, while others may deviate from the new data distribution and be detrimental. To address this, we propose a dynamic mixture-of-experts (DyMoE) approach for incremental learning. Specifically, a DyMoE GNN layer adds new expert networks specialized in modeling the incoming data blocks. We design a customized regularization loss that utilizes data sequence information so existing experts can maintain their ability to solve old tasks while helping the new expert learn the new data effectively. As the number of data blocks grows over time, the computational cost of the full mixture-of-experts (MoE) model increases. To address this, we introduce a sparse MoE approach, where only the top-$k$ most relevant experts make predictions, significantly reducing the computation time. Our model achieved 4.92\% relative accuracy increase compared to the best baselines on class incremental learning, showing the model's exceptional power.

cs.LG↗
arXiv2025

Boosting Domain Incremental Learning: Selecting the Optimal Parameters is All You Need

Deep neural networks (DNNs) often underperform in real-world, dynamic settings where data distributions change over time. Domain Incremental Learning (DIL) offers a solution by enabling continual model adaptation, with Parameter-Isolation DIL (PIDIL) emerging as a promising paradigm to reduce knowledge conflicts. However, existing PIDIL methods struggle with parameter selection accuracy, especially as the number of domains and corresponding classes grows. To address this, we propose SOYO, a lightweight framework that improves domain selection in PIDIL. SOYO introduces a Gaussian Mixture Compressor (GMC) and Domain Feature Resampler (DFR) to store and balance prior domain data efficiently, while a Multi-level Domain Feature Fusion Network (MDFN) enhances domain feature extraction. Our framework supports multiple Parameter-Efficient Fine-Tuning (PEFT) methods and is validated across tasks such as image classification, object detection, and speech enhancement. Experimental results on six benchmarks demonstrate SOYO's consistent superiority over existing baselines, showcasing its robustness and adaptability in complex, evolving environments. The codes will be released in https://github.com/qwangcv/SOYO.

cs.CV↗
arXiv2025

DualCP: Rehearsal-Free Domain-Incremental Learning via Dual-Level Concept Prototype

Domain-Incremental Learning (DIL) enables vision models to adapt to changing conditions in real-world environments while maintaining the knowledge acquired from previous domains. Given privacy concerns and training time, Rehearsal-Free DIL (RFDIL) is more practical. Inspired by the incremental cognitive process of the human brain, we design Dual-level Concept Prototypes (DualCP) for each class to address the conflict between learning new knowledge and retaining old knowledge in RFDIL. To construct DualCP, we propose a Concept Prototype Generator (CPG) that generates both coarse-grained and fine-grained prototypes for each class. Additionally, we introduce a Coarse-to-Fine calibrator (C2F) to align image features with DualCP. Finally, we propose a Dual Dot-Regression (DDR) loss function to optimize our C2F module. Extensive experiments on the DomainNet, CDDB, and CORe50 datasets demonstrate the effectiveness of our method.

cs.CV↗
arXiv2025

Deal: Distributed End-to-End GNN Inference for All Nodes

Graph Neural Networks (GNNs) are a new research frontier with various applications and successes. The end-to-end inference for all nodes, is common for GNN embedding models, which are widely adopted in applications like recommendation and advertising. While sharing opportunities arise in GNN tasks (i.e., inference for a few nodes and training), the potential for sharing in full graph end-to-end inference is largely underutilized because traditional efforts fail to fully extract sharing benefits due to overwhelming overheads or excessive memory usage. This paper introduces Deal, a distributed GNN inference system that is dedicated to end-to-end inference for all nodes for graphs with multi-billion edges. First, we unveil and exploit an untapped sharing opportunity during sampling, and maximize the benefits from sharing during subsequent GNN computation. Second, we introduce memory-saving and communication-efficient distributed primitives for lightweight 1-D graph and feature tensor collaborative partitioning-based distributed inference. Third, we introduce partitioned, pipelined communication and fusing feature preparation with the first GNN primitive for end-to-end inference. With Deal, the end-to-end inference time on real-world benchmark datasets is reduced up to 7.70 x and the graph construction time is reduced up to 21.05 x, compared to the state-of-the-art.

cs.DC↗
arXiv2025

Space Rotation with Basis Transformation for Training-free Test-Time Adaptation

With the development of visual-language models (VLM) in downstream task applications, test-time adaptation methods based on VLM have attracted increasing attention for their ability to address changes distribution in test-time. Although prior approaches have achieved some progress, they typically either demand substantial computational resources or are constrained by the limitations of the original feature space, rendering them less effective for test-time adaptation tasks. To address these challenges, we propose a training-free feature space rotation with basis transformation for test-time adaptation. By leveraging the inherent distinctions among classes, we reconstruct the original feature space and map it to a new representation, thereby enhancing the clarity of class differences and providing more effective guidance for the model during testing. Additionally, to better capture relevant information from various classes, we maintain a dynamic queue to store representative samples. Experimental results across multiple benchmarks demonstrate that our method outperforms state-of-the-art techniques in terms of both performance and efficiency.

cs.CV↗
arXiv2025

Spectro-Riemannian Graph Neural Networks

Can integrating spectral and curvature signals unlock new potential in graph representation learning? Non-Euclidean geometries, particularly Riemannian manifolds such as hyperbolic (negative curvature) and spherical (positive curvature), offer powerful inductive biases for embedding complex graph structures like scale-free, hierarchical, and cyclic patterns. Meanwhile, spectral filtering excels at processing signal variations across graphs, making it effective in homophilic and heterophilic settings. Leveraging both can significantly enhance the learned representations. To this end, we propose Spectro-Riemannian Graph Neural Networks (CUSP) - the first graph representation learning paradigm that unifies both CUrvature (geometric) and SPectral insights. CUSP is a mixed-curvature spectral GNN that learns spectral filters to optimize node embeddings in products of constant-curvature manifolds (hyperbolic, spherical, and Euclidean). Specifically, CUSP introduces three novel components: (a) Cusp Laplacian, an extension of the traditional graph Laplacian based on Ollivier-Ricci curvature, designed to capture the curvature signals better; (b) Cusp Filtering, which employs multiple Riemannian graph filters to obtain cues from various bands in the eigenspectrum; and (c) Cusp Pooling, a hierarchical attention mechanism combined with a curvature-based positional encoding to assess the relative importance of differently curved substructures in our graph. Empirical evaluation across eight homophilic and heterophilic datasets demonstrates the superiority of CUSP in node classification and link prediction tasks, with a gain of up to 5.3% over state-of-the-art models. The code is available at: https://github.com/amazon-science/cusp.

cs.LG↗