Large language models (LLMs) augmented with retrieval have shown impressive performance in open-domain question answering, yet struggle significantly with temporal knowledge graph question answering (TKGQA). The core issue lies in structural misalignment: treating structured, temporally sensitive graph queries as plain text often causes LLMs to retrieve or reason with semantically similar but structurally incorrect facts, resulting in critical inaccuracies. To address this, we introduce SAR (Structure-Aligned Reasoning), a novel TKGQA framework that integrates LLM reasoning tightly with the explicit subject–predicate–object–time schema inherent in knowledge graphs. SAR employs an LLM agent to first decompose natural language questions into structured queries, clearly delineating entities, relationships, and temporal constraints. It then conducts schema-consistent, time-aware retrieval from the knowledge graph to acquire candidate quadruples, which guide a subsequent iterative ReAct-style reasoning process by the LLM. A final verification stage ensures that proposed answers strictly adhere to temporal conditions, reinforcing accuracy and temporal coherence. Experiments on two benchmark datasets, MultiTQ and CronQuestions, demonstrate SAR’s effectiveness, achieving the best results. Specifically, with GPT-4.1, SAR achieves 78.2% Hits@1 on MultiTQ, significantly outperforming existing methods, and similarly establishes a new performance record on CronQuestions. Our results underscore the critical importance of structural alignment in temporal reasoning tasks, particularly in handling complex queries involving multiple temporal constraints and multi-hop reasoning.

Cross-datacenter distributed model training is increasingly important, but WAN packet loss and long propagation delays make retransmission-based recovery expensive. Existing RDMA FEC schemes are application-unaware and uniformly protect all packets, wasting inter-datacenter bandwidth. In this paper, we present SMART, a semantic-aware loss recovery scheme for cross-datacenter model training. SMART distinguishes data-parallel (DP) gradients from pipeline-parallel (PP) activations using lightweight message-level tags; it selectively protects high-priority DP packets, gives earlier PP traffic stronger FEC, zero-fills only small profiled residual PP loss, and compensates tolerated low-priority DP losses. Simulation experiments with Qwen3-3B-derived workloads show that SMART reduces average and P99 flow completion times compared with retransmission-only recovery and uniform FEC, while preserving convergence close to the no-loss baseline.

Simulink is MathWorks’ commercial cyber-physical systems development tool, which enables engineers to do rapid prototyping of their systems through simulation and embedded code generation. When a Simulink model meets requirements, engineers can utilize embedded coder to convert it into embedded code (e.g., C source code) and deploy it in safety-critical applications such as automotive, aerospace, and healthcare. However, bugs or incorrect implementations in code generation may lead to unexpected behaviors in target applications, posing security risks. Therefore, it is crucial to eliminate such bugs in embedded code generation. To address this issue, we propose DESCO, a differential testing approach to test embedded code generation in Simulink. DESCO considers the functional correlation between Simulink blocks for partitioning, aiming to generate diverse and complex bug-triggering Simulink models to thoroughly exercise the embedded code generation. DESCO then detects bugs by analyzing the outputs of these Simulink models by differential testing. The experiments demonstrate that DESCO significantly outperforms existing approaches. In three months, DESCO reported 16 issues, including 12 confirmed as bugs by MathWorks Support.