publications

2026

1. SVR-MAD: A Bayesian-Inspired Framework for Posterior-Guided Multi-Agent Debate

   Weifan Jiang, Rana Shahout, Minghao Li, Zhenting Qi, Yilun Du, Michael Mitzenmacher, and Minlan Yu

   Findings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2026

   Abs PDF

   Multi-Agent Debate (MAD) improves LLM-agent accuracy but suffers from rapid context growth, limiting scalability in larger multi-agent settings. Existing methods prune low-utility communications using prior signals, such as token-level log-likelihoods or LLM self-reported confidence. However, these signals become unreliable under hallucination, degrading the accuracy of MAD methods that rely on them. We propose SVR-MAD, a Bayesian-inspired MAD framework that treats pre-debate signals as priors and debate outcomes as posterior-style evidence for estimating agent correctness. SVR-MAD uses this evidence to incrementally construct the communication graph, prioritizing agents whose answers survive peer challenges. Experiments across multiple LLMs and benchmarks show that SVR-MAD reduces token cost by up to 61% while matching or improving accuracy relative to the most accurate competing MAD baseline.

2025

1. Don’t Stop Me Now: Embedding Based Scheduling for LLMs

   Rana Shahout, Eran Malach, Chunwei Liu, Weifan Jiang, Minlan Yu, and Michael Mitzenmacher

   International Conference on Learning Representations (ICLR), 2025

   Abs PDF Poster

   Efficient scheduling is crucial for interactive Large Language Model (LLM) applications, where low request completion time directly impacts user engagement. Size-based scheduling algorithms like Shortest Remaining Process Time (SRPT) aim to reduce average request completion time by leveraging known or estimated request sizes and allowing preemption by incoming jobs with shorter service times. However, two main challenges arise when applying size-based scheduling to LLM systems. First, accurately predicting output lengths from prompts is challenging and often resource-intensive, making it impractical for many systems. As a result, the state-of-the-art LLM systems default to first-come, first-served scheduling, which can lead to head-of-line blocking and reduced system efficiency. Second, preemption introduces extra memory overhead to LLM systems as they must maintain intermediate states for unfinished (preempted) requests. In this paper, we propose TRAIL, a method to obtain output predictions from the target LLM itself. After generating each output token, we recycle the embedding of its internal structure as input for a lightweight classifier that predicts the remaining length for each running request. Using these predictions, we propose a prediction-based SRPT variant with limited preemption designed to account for memory overhead in LLM systems. This variant allows preemption early in request execution when memory consumption is low but restricts preemption as requests approach completion to optimize resource utilization. On the theoretical side, we derive a closed-form formula for this SRPT variant in an M/G/1 queue model, which demonstrates its potential value. In our system, we implement this preemption policy alongside our embedding-based prediction method. Our refined predictions from layer embeddings achieve 2.66x lower mean absolute error compared to BERT predictions from sequence prompts. TRAIL achieves 1.66x to 2.01x lower mean latency on the Alpaca dataset and 1.76x to 24.07x lower mean time to the first token compared to the state-of-the-art serving system.

2021

1. Towards a Traffic Map of the Internet: Connecting the Dots between Popular Services and Users

   Thomas Koch, Weifan Jiang, Tao Luo, Petros Gigis, Yunfan Zhang, Kevin Vermeulen, Emile Aben, Matt Calder, Ethan Katz-Bassett, Lefteris Manassakis, Georgios Smaragdakis, and Narseo Vallina-Rodriguez

   ACM Workshop on Hot Topics in Networks (HotNets), 2021

   Abs PDF

   The impact of Internet phenomena depends on how they impact users, but researchers lack visibility into how to translate Internet events into their impact. Distressingly, the research community seems to have lost hope of obtaining this information without relying on privileged viewpoints. We argue for optimism thanks to new network measurement methods and changes in Internet structure which make it possible to construct an "Internet traffic map". This map would identify the locations of users and major services, the paths between them, and the relative activity levels routed along these paths. We sketch our vision for the map, detail new measurement ideas for map construction, and identify key challenges that the research community should tackle. The realization of an Internet traffic map will be an Internet-scale research effort with Internet-scale impacts that reach far beyond the research community, and so we hope our fellow researchers are excited to join us in addressing this challenge.

2. Towards Identifying Networks with Internet Clients Using Public Data

   Weifan Jiang, Tao Luo, Thomas Koch, Yunfan Zhang, Ethan Katz-Bassett, and Matt Calder

   ACM Internet Measurement Conference (IMC), 2021

   Abs PDF Slides Talk

   Does an outage impact any users? Can a geolocation database known to be good at locating users and bad at infrastructure be trusted for a particular prefix? Is a content-heavy network likely to peer with a particular network? For these questions and many more, knowing which prefixes contain Internet users aids in interpreting Internet analysis. However, existing datasets of Internet activity are out of date, unvalidated, based on privileged data, or too coarse. As a step towards identifying which IP prefixes contain users, we present multiple novel techniques to identify which IP prefixes host web clients without relying on privileged data. Our techniques identify client activity in ASes responsible for 98.8% of Microsoft CDN traffic and in prefixes responsible for 95.2% of Microsoft CDN traffic. Less than 1% of prefixes identified by our technique as active do not contact Microsoft at all. We present measurements of Internet usage worldwide and sketch future directions for extending the techniques to measure relative activity levels across prefixes.

3. Cost-Aware Robust Tree Ensembles for Security Applications

   Yizheng Chen, Shiqi Wang, Weifan Jiang, Asaf Cidon, and Suman Jana

   USENIX Security Symposium, 2021

   Abs PDF

   There are various costs for attackers to manipulate the features of security classifiers. The costs are asymmetric across features and to the directions of changes, which cannot be precisely captured by existing cost models based on Lp-norm robustness. In this paper, we utilize such domain knowledge to increase the attack cost of evading classifiers, specifically, tree ensemble models that are widely used by security tasks. We propose a new cost modeling method to capture the feature manipulation cost as constraint, and then we integrate the cost-driven constraint into the node construction process to train robust tree ensembles. During the training process, we use the constraint to find data points that are likely to be perturbed given the feature manipulation cost, and we use a new robust training algorithm to optimize the quality of the trees. Our cost-aware training method can be applied to different types of tree ensembles, including gradient boosted decision trees and random forest models. Using Twitter spam detection as the case study, our evaluation results show that we can increase the attack cost by 10.6⇥ compared to the baseline. Moreover, our robust training method using cost-driven constraint can achieve higher accuracy, lower false positive rate, and stronger cost-aware robustness than the state-of-theart training method using L•-norm cost model. Our code is available at https://github.com/surrealyz/growtrees.

2020

1. Cloud Provider Connectivity in the Flat Internet

   Todd Arnold, Jia He, Weifan Jiang, Matt Calder, Italo Cunha, Vasileios Giotsas, and Ethan Katz-Bassett

   ACM Internet Measurement Conference (IMC), 2020

   Abs PDF

   The Tier-1 ISPs have been considered the Internet’s backbone since the dawn of the modern Internet 30 years ago, as they guarantee global reachability. However, their influence and importance are waning as Internet flattening decreases the demand for transit services and increases the importance of private interconnections. Conversely, major cloud providers – Amazon, Google, IBM, and Microsoft– are gaining in importance as more services are hosted on their infrastructures. They ardently support Internet flattening and are rapidly expanding their global footprints, which enables them to bypass the Tier-1 ISPs and other large transit providers to reach many destinations.In this paper we seek to quantify the extent to which the cloud providers’ can bypass the Tier-1 ISPs and other large transit providers. We conduct comprehensive measurements to identify the neighbor networks of the major cloud providers and combine them with AS relationship inferences to model the Internet’s AS-level topology to calculate a new metric, hierarchy-free reachability, which characterizes the reachability a network can achieve without traversing the networks of the Tier-1 and Tier-2 ISPs. We show that the cloud providers are able to reach over 76% of the Internet without traversing the Tier-1 and Tier-2 ISPs, more than virtually every other network.

2. Poster: Footprint and Performance of Large Cloud Networks

   Jia He, Weifan Jiang, Ege Gürmeriçliler, Georgia Essig, Arpit Gupta, Matt Calder, Vasileios Giotsas, Italo Cunha, Ethan Katz-Bassett, and Todd Arnold

   N2Women Workshop at ACM SIGCOMM, 2020

   Abs PDF Poster

   The Tier-1 and Tier-2 transit providers have historically been considered the backbone of the Internet as they guarantee global reachability. In recent years, Internet flattening has reduced the need for transit providers, an effect greatly contributed to by the top cloud providers, such as Google, Amazon, Microsoft, and IBM. Recently, these cloud providers started offering two performance tiers for routing traffic. One tier, referred to as "Premium Tier" (PT), the Google-specific term, uses the cloud provider’s private network as much as possible, while "Standard Tier" (ST), uses the public Internet as much as possible. Through analysis of measurements made to gather performance and connectivity data, we find that the cloud provider networks’ points-of-presence (PoPs) tend to be deployed closer to population centers than the transit providers’ PoPs. We also find that the performance improvement from PT service is dependent on variables such as the ST/PT path length difference. These metrics demonstrate how cloud providers connect within the Internet, and what benefits their private networks provide to users.

3. Irrigation Detection by Car: Computer Vision and Sensing for the Detection and Geolocation of Irrigated and Non-irrigated Farmland

   Weifan Jiang, Vivek Kumar, Nikhil Mehta, Jack Bott, and Vijay Modi

   IEEE Global Humanitarian Technology Conference (GHTC), 2020

   Abs PDF

   Irrigation can greatly increase the income of smallholder farmers in sub-Saharan Africa. By providing information about current irrigation utilization, or lack thereof, we seek to encourage investment in irrigation systems and their supporting infrastructure. In this paper, we describe the design, prototyping, and testing of a novel, cost-effective, and reliable computer vision system that is capable of locating irrigated plots at scale. Our system will be mounted to a vehicle and record the depth of objects in the camera’s view while the vehicle is in motion. The GPS coordinates of objects are computed based on estimated depth, vehicle coordinates, and orientation, available from included sensors. We tested our prototype on objects at various distances from the system and achieved feasible accuracy with acceptable error in the estimated depth. In the future, we hope to deploy the system in parts of sub-Saharan Africa, to detect and geolocate irrigated agricultural plots during the dry season. Then we plan to use that collected data to inform and train machine learning models that use remote sensing and satellite imagery.