trace-driven emulation

Existing discrete-event simulators (e.g., Hypatia) struggle to accurately evaluate real-world protocols and applications in LEO mega-constellation networks, exhibiting a substantial fidelity gap between simulation and empirical measurements. To bridge this gap, we propose a trajectory-driven satellite network path simulation–replay architecture: leveraging Hypatia for offline, high-fidelity path modeling and feature extraction to generate reusable end-to-end trajectory files, which are then replayed in real time on actual hardware and software platforms. This tightly couples simulation with network emulation. The architecture supports multi-constellation scenarios and precisely reproduces dynamic network characteristics—including latency, bandwidth, and topology evolution. Experimental evaluation demonstrates a correlation of 0.96 between simulated and replayed results, significantly enhancing the realism and reproducibility of protocol and application assessments in LEO satellite networks.

Existing testbeds for autonomous driving video uplink struggle to simultaneously achieve realism, repeatability, and practicality: real-vehicle platforms incur high costs and face challenges in reproducing consistent network conditions, while simulation-based approaches lack authentic end-to-end dynamics. To address this gap, this work proposes a trajectory-driven, cloud-native cellular network emulation testbed that, for the first time, integrates time-synchronized replay of vehicular and network traces with a cloud-native architecture. Implemented on commodity Linux virtual nodes, the system enables route-aware, high-fidelity video uplink experiments. It supports production-grade video protocol stacks and offers low cost, scalability, repeatability, and controlled comparative experimentation—significantly enhancing the practicality and reproducibility of vehicular communication testing.

This work addresses the limitations of existing cybersecurity datasets, which are predominantly static and ill-suited for enabling controllable replay and traceability in heterogeneous, multi-protocol environments. To overcome this, the authors propose a scenario-oriented, container-native testing platform that leverages declarative configuration to parameterize the generation of both adversarial and benign network traffic, log collection, and dataset integration. The platform encapsulates 60 attack scenarios, nine target services, and benign traffic generators within single-purpose containers and integrates them into an automated pipeline for feature extraction and experimental execution. Designed with reproducibility, auditability, and extensibility in mind, the framework significantly reduces operational bias and supports fully traceable, reproducible experiments in complex settings such as IoT and IIoT networks.

This study addresses a critical flaw in existing replay-based log stitching methods for evaluating multi-step LLM agents: the erroneous assumption that switching models does not alter subsequent trajectories, leading to distorted evaluations. The authors conduct controlled branch-rollback experiments in the real-world SWE-bench environment to reconstruct agent trajectories and systematically assess the actual impact of model substitutions on downstream behavior. Their findings reveal that model switching rewrites 61–94% of subsequent actions, causing replay-based evaluation to mispredict all key outcomes—trajectory similarity drops to merely 0.00–0.11. Furthermore, quantization schemes such as FP8 and AWQ significantly affect apparent “determinism.” These results expose severe bias in static replay evaluation and advocate for a shift toward dynamic evaluation paradigms. The authors publicly release all trajectories and their evaluation framework.

Existing network traffic generation methods struggle to accurately model multi-flow interactions and TCP state machines because they directly decode raw packet fields, conflating behavioral semantics with protocol constraints and relying on heuristic post-hoc repairs. This work proposes TraceCodec, the first framework to integrate a neural codec with a deterministic protocol compiler in a collaborative architecture. By shifting the generation space from raw packet headers to a structured latent space of packet actions—each comprising a timestamp, an explicit flow slot, and transmission cues—and modeling sequences of continuous latent variables, TraceCodec decouples generative logic from protocol implementation. This enables synthesis of high-fidelity PCAP traces without requiring post-generation correction. Evaluated on the CICIDS2017 Monday dataset, TraceCodec achieves packet count, protocol composition, and flow size errors below 0.03%, significantly outperforming baselines in flow count accuracy, TCP state fidelity, and preservation of multi-flow interleaving structures.

This work addresses the challenge of catastrophic forgetting and misalignment in continual instruction tuning, where fixed replay ratios fail to adapt to dynamic task distributions. The authors propose PROXYMIX, a novel framework that leverages the “forgetting mirror” hypothesis—empirically validated for the first time—which posits that the relative forgetting sensitivity across tasks remains consistent across model scales. By training a dynamic replay controller on a small proxy model, PROXYMIX transfers this policy to large models without requiring knowledge of future tasks. The controller constructs its state from normalized validation loss and its temporal dynamics, then adaptively blends old and new data via a mask-based mixing mechanism. Evaluated on five sequential instruction-tuning benchmarks with LLaMA-3-8B, PROXYMIX improves average accuracy by 3.4 points, reduces final forgetting by 3.5 points, enhances safety by 5.8 points, and achieves these gains at only 1/50th the policy learning cost of Oracle Target RL.

This work addresses the challenge of debugging non-deterministic programs on microcontrollers, where sensor-driven inputs lead to irreproducible execution paths and existing techniques suffer from snapshot overhead, model dependency, and state explosion. The authors propose a trajectory-based “multi-verse” debugging approach that integrates concolic execution into this paradigm for the first time. By recording lightweight execution trajectories instead of full system snapshots, the method dynamically identifies critical inputs and prunes redundant paths, substantially mitigating state explosion while reducing memory and communication costs—making it suitable for resource-constrained environments. Implemented as a prototype atop the WARDuino WebAssembly virtual machine with a remote debugging architecture, the approach demonstrates significant reductions in state space and enhanced debugging efficiency and scalability in real-world scenarios compared to conventional solutions.