arXiv subjects

Alexander Gutfraind

Publications and source records attributed to Alexander Gutfraind.

At least 19 recordsLinked to original sources

arXiv2026

Cross-Modal Memory Compression for Efficient Multi-Agent Debate

Multi-agent debate can improve reasoning quality and reduce hallucinations, but it incurs rapidly growing context as debate rounds and agent count increase. Retaining full textual histories leads to token usage that can exceed context limits and often requires repeated summarization, adding overhead and compounding information loss. We introduce DebateOCR, a cross-modal compression framework that replaces long textual debate traces with compact image representations, which are then consumed through a dedicated vision encoder to condition subsequent rounds. This design compresses histories that commonly span tens to hundreds of thousands of tokens, cutting input tokens by more than 92% and yielding substantially lower compute cost and faster inference across multiple benchmarks. We further provide a theoretical perspective showing that diversity across agents supports recovery of omitted information: although any single compressed history may discard details, aggregating multiple agents' compressed views allows the collective representation to approach the information bottleneck with exponentially high probability.

cs.AI↗

arXiv2025

From nuclear safety to LLM security: Applying non-probabilistic risk management strategies to build safe and secure LLM-powered systems

Large language models (LLMs) offer unprecedented and growing capabilities, but also introduce complex safety and security challenges that resist conventional risk management. While conventional probabilistic risk analysis (PRA) requires exhaustive risk enumeration and quantification, the novelty and complexity of these systems make PRA impractical, particularly against adaptive adversaries. Previous research found that risk management in various fields of engineering such as nuclear or civil engineering is often solved by generic (i.e. field-agnostic) strategies such as event tree analysis or robust designs. Here we show how emerging risks in LLM-powered systems could be met with 100+ of these non-probabilistic strategies to risk management, including risks from adaptive adversaries. The strategies are divided into five categories and are mapped to LLM security (and AI safety more broadly). We also present an LLM-powered workflow for applying these strategies and other workflows suitable for solution architects. Overall, these strategies could contribute (despite some limitations) to security, safety and other dimensions of responsible AI.

cs.CR↗

arXiv2023

On strategies for risk management and decision making under uncertainty shared across multiple fields

Decision theory recognizes two principal approaches to solving problems under uncertainty: probabilistic models and cognitive heuristics. However, engineers, public planners and decision-makers in other fields seem to employ solution strategies that do not fall into either field, i.e., strategies such as robust design and contingency planning. In addition, identical strategies appear in several fields and disciplines, pointing to an important shared toolkit. The focus of this paper is to develop a systematic understanding of such strategies and develop a framework to better employ them in decision making and risk management. The paper finds more than 110 examples of such strategies and this approach to risk is termed RDOT: Risk-reducing Design and Operations Toolkit. RDOT strategies fall into six broad categories: structural, reactive, formal, adversarial, multi-stage and positive. RDOT strategies provide an efficient response even to radical uncertainty or unknown unknowns that are challenging to address with probabilistic methods. RDOT could be incorporated into decision theory using workflows, multi-objective optimization and multi-attribute utility theory. Overall, RDOT represents an overlooked class of versatile responses to uncertainty. Because RDOT strategies do not require precise estimation or forecasting, they are particularly helpful in decision problems affected by uncertainty and for resource-constrained decision making.

q-fin.RM↗

arXiv2023

On solving decision and risk management problems subject to uncertainty

Uncertainty is a pervasive challenge in decision and risk management and it is usually studied by quantification and modeling. Interestingly, engineers and other decision makers usually manage uncertainty with strategies such as incorporating robustness, or by employing decision heuristics. The focus of this paper is then to develop a systematic understanding of such strategies, determine their range of application, and develop a framework to better employ them. Based on a review of a dataset of 100 decision problems, this paper found that many decision problems have pivotal properties, i.e. properties that enable solution strategies, and finds 14 such properties. Therefore, an analyst can first find these properties in a given problem, and then utilize the strategies they enable. Multi-objective optimization methods could be used to make investment decisions quantitatively. The analytical complexity of decision problems can also be scored by evaluating how many of the pivotal properties are available. Overall, we find that in the light of pivotal properties, complex problems under uncertainty frequently appear surprisingly tractable.

cs.AI↗

arXiv2019

Using massive health insurance claims data to predict very high-cost claimants: a machine learning approach

Due to escalating healthcare costs, accurately predicting which patients will incur high costs is an important task for payers and providers of healthcare. High-cost claimants (HiCCs) are patients who have annual costs above $\$250,000$ and who represent just 0.16% of the insured population but currently account for 9% of all healthcare costs. In this study, we aimed to develop a high-performance algorithm to predict HiCCs to inform a novel care management system. Using health insurance claims from 48 million people and augmented with census data, we applied machine learning to train binary classification models to calculate the personal risk of HiCC. To train the models, we developed a platform starting with 6,006 variables across all clinical and demographic dimensions and constructed over one hundred candidate models. The best model achieved an area under the receiver operating characteristic curve of 91.2%. The model exceeds the highest published performance (84%) and remains high for patients with no prior history of high-cost status (89%), who have less than a full year of enrollment (87%), or lack pharmacy claims data (88%). It attains an area under the precision-recall curve of 23.1%, and precision of 74% at a threshold of 0.99. A care management program enrolling 500 people with the highest HiCC risk is expected to treat 199 true HiCCs and generate a net savings of $\$7.3$ million per year. Our results demonstrate that high-performing predictive models can be constructed using claims data and publicly available data alone, even for rare high-cost claimants exceeding $\$250,000$. Our model demonstrates the transformational power of machine learning and artificial intelligence in care management, which would allow healthcare payers and providers to introduce the next generation of care management programs.

cs.LG↗

arXiv2019

Risk analysis beyond vulnerability and resilience - characterizing the defensibility of critical systems

A common problem in risk analysis is to characterize the overall security of a system of valuable assets (e.g., government buildings or communication hubs), and to suggest measures to mitigate any hazards or security threats. Currently, analysts typically rely on a combination of indices, such as resilience, robustness, redundancy, security, and vulnerability. However, these indices are not by themselves sufficient as a guide to action; for example, while it is possible to develop policies to decrease vulnerability, such policies may not always be cost-effective. Motivated by this gap, we propose a new index, defensibility. A system is considered defensible to the extent that a modest investment can significantly reduce the damage from an attack or disruption. To compare systems whose performance is not readily commensurable (e.g., the electrical grid vs. the water-distribution network, both of which are critical, but which provide distinct types of services), we defined defensibility as a dimensionless index. After defining defensibility quantitatively, we illustrate how the defensibility of a system depends on factors such as the defender and attacker asset valuations, the nature of the threat (whether intelligent and adaptive, or random), and the levels of attack and defense strengths and provide analytical results that support the observations arising from the above illustrations. Overall, we argue that the defensibility of a system is an important dimension to consider when evaluating potential defensive investments, and that it can be applied in a variety of different contexts.

cs.CR↗

arXiv2018

Multiscale Planar Graph Generation

The study of network representations of physical, biological, and social phenomena can help us better understand the structural and functional dynamics of their networks and formulate predictive models of these phenomena. However, due to the scarcity of real-world network data owing to factors such as cost and effort required in collection of network data and the sensitivity of this data towards theft and misuse, engineers and researchers often rely on synthetic data for simulations, hypothesis testing, decision making, and algorithm engineering. An important characteristic of infrastructure networks such as roads, water distribution and other utility systems is that they can be embedded in a plane, therefore to simulate these system we need realistic networks which are also planar. While the currently-available synthetic network generators can model networks that exhibit realism, they do not guarantee or achieve planarity. Therefore, in this paper we present a flexible algorithm that can synthesize realistic networks that are planar. The method follows a multi-scale randomized editing approach generating a hierarchy of coarsened networks of a given planar graph and introducing edits at various levels in the hierarchy. The method preserves the structural properties with minimal bias including the planarity of the network, while introducing realistic variability at multiple scales.

cs.SI↗

arXiv2016

Generating realistic scaled complex networks

Research on generative models is a central project in the emerging field of network science, and it studies how statistical patterns found in real networks could be generated by formal rules. Output from these generative models is then the basis for designing and evaluating computational methods on networks, and for verification and simulation studies. During the last two decades, a variety of models has been proposed with an ultimate goal of achieving comprehensive realism for the generated networks. In this study, we (a) introduce a new generator, termed ReCoN; (b) explore how ReCoN and some existing models can be fitted to an original network to produce a structurally similar replica, (c) use ReCoN to produce networks much larger than the original exemplar, and finally (d) discuss open problems and promising research directions. In a comparative experimental study, we find that ReCoN is often superior to many other state-of-the-art network generation methods. We argue that ReCoN is a scalable and effective tool for modeling a given network while preserving important properties at both micro- and macroscopic scales, and for scaling the exemplar data by orders of magnitude in size.

cs.SI↗

arXiv2015

Operational resilience: concepts, design and analysis

Building resilience into today's complex infrastructures is critical to the daily functioning of society and its ability to withstand and recover from natural disasters, epidemics, and cyber-threats. This study proposes quantitative measures that implement the definition of engineering resilience advanced by the National Academy of Sciences. The approach is applicable across physical, information, and social domains. It evaluates the critical functionality, defined as a performance function of time set by the stakeholders. Critical functionality is a source of valuable information, such as the integrated system resilience over a time interval, and its robustness. The paper demonstrates the formulation on two classes of models: 1) multi-level directed acyclic graphs, and 2) interdependent coupled networks. For both models synthetic case studies are used to explore trends. For the first class, the approach is also applied to the Linux operating system. Results indicate that desired resilience and robustness levels are achievable by trading off different design parameters, such as redundancy, node recovery time, and backup supply available. The nonlinear relationship between network parameters and resilience levels confirms the utility of the proposed approach, which is of benefit to analysts and designers of complex systems and networks.

physics.soc-ph↗

arXiv2014

Network installation and recovery: approximation lower bounds and faster exact formulations

We study the Neighbor Aided Network Installation Problem (NANIP) introduced previously which asks for a minimal cost ordering of the vertices of a graph, where the cost of visiting a node is a function of the number of neighbors that have already been visited. This problem has applications in resource management and disaster recovery. In this paper we analyze the computational hardness of NANIP. In particular we show that this problem is NP-hard even when restricted to convex decreasing cost functions, give a linear approximation lower bound for the greedy algorithm, and prove a general sub-constant approximation lower bound. Then we give a new integer programming formulation of NANIP and empirically observe its speedup over the original integer program.

cs.DM↗

arXiv2012

Multiscale Network Generation

Networks are widely used in science and technology to represent relationships between entities, such as social or ecological links between organisms, enzymatic interactions in metabolic systems, or computer infrastructure. Statistical analyses of networks can provide critical insights into the structure, function, dynamics, and evolution of those systems. However, the structures of real-world networks are often not known completely, and they may exhibit considerable variation so that no single network is sufficiently representative of a system. In such situations, researchers may turn to proxy data from related systems, sophisticated methods for network inference, or synthetic networks. Here, we introduce a flexible method for synthesizing realistic ensembles of networks starting from a known network, through a series of mappings that coarsen and later refine the network structure by randomized editing. The method, MUSKETEER, preserves structural properties with minimal bias, including unknown or unspecified features, while introducing realistic variability at multiple scales. Using examples from several domains, we show that MUSKETEER produces the intended stochasticity while achieving greater fidelity across a suite of network properties than do other commonly used network generation algorithms.

cs.DM↗

arXiv2012

Optimal recovery of damaged infrastructure network

Natural disasters or attacks may disrupt infrastructure networks on a vast scale. Parts of the damaged network are interdependent, making it difficult to plan and optimally execute the recovery operations. To study how interdependencies affect the recovery schedule, we introduce a new discrete optimization problem where the goal is to minimize the total cost of installing (or recovering) a given network. This cost is determined by the structure of the network and the sequence in which the nodes are installed. Namely, the cost of installing a node is a function of the number of its neighbors that have been installed before it. We analyze the natural case where the cost function is decreasing and convex, and provide bounds on the cost of the optimal solution. We also show that all sequences have the same cost when the cost function is linear and provide an upper bound on the cost of a random solution for an Erdős-Rényi random graph. Examining the computational complexity, we show that the problem is NP-hard when the cost function is arbitrary. Finally, we provide a formulation as an integer program, an exact dynamic programming algorithm, and a greedy heuristic which gives high quality solutions.

math.OC↗

arXiv2011

Evader Interdiction and Collateral Damage

In network interdiction problems, evaders (e.g., hostile agents or data packets) may be moving through a network towards targets and we wish to choose locations for sensors in order to intercept the evaders before they reach their destinations. The evaders might follow deterministic routes or Markov chains, or they may be reactive}, i.e., able to change their routes in order to avoid sensors placed to detect them. The challenge in such problems is to choose sensor locations economically, balancing security gains with costs, including the inconvenience sensors inflict upon innocent travelers. We study the objectives of 1) maximizing the number of evaders captured when limited by a budget on sensing cost and 2) capturing all evaders as cheaply as possible. We give optimal sensor placement algorithms for several classes of special graphs and hardness and approximation results for general graphs, including for deterministic or Markov chain-based and reactive or oblivious evaders. In a similar-sounding but fundamentally different problem setting posed by Rubinstein and Glazer where both evaders and innocent travelers are reactive, we again give optimal algorithms for special cases and hardness and approximation results on general graphs.

cs.SI↗

arXiv2011

Lanchester Theory and the Fate of Armed Revolts

Major revolts have recently erupted in parts of the Middle East with substantial international repercussions. Predicting, coping with and winning those revolts have become a grave problem for many regimes and for world powers. We propose a new model of such revolts that describes their evolution by building on the classic Lanchester theory of combat. The model accounts for the split in the population between those loyal to the regime and those favoring the rebels. We show that, contrary to classical Lanchesterian insights regarding traditional force-on-force engagements, the outcome of a revolt is independent of the initial force sizes; it only depends on the fraction of the population supporting each side and their combat effectiveness. We also consider the effects of foreign intervention and of shifting loyalties of the two populations during the conflict. The model's predictions are consistent with the situations currently observed in Afghanistan, Libya and Syria (Spring 2011) and it offers tentative guidance on policy.

math.DS↗

arXiv2010

Interdiction of a Markovian Evader

Shortest path network interdiction is a combinatorial optimization problem on an activity network arising in a number of important security-related applications. It is classically formulated as a bilevel maximin problem representing an "interdictor" and an "evader". The evader tries to move from a source node to the target node along a path of the least cost while the interdictor attempts to frustrate this motion by cutting edges or nodes. The interdiction objective is to find the optimal set of edges to cut given that there is a finite interdiction budget and the interdictor must move first. We reformulate the interdiction problem for stochastic evaders by introducing a model in which the evader follows a Markovian random walk guided by the least-cost path to the target. This model can represent incomplete knowledge about the evader, and the resulting model is a nonlinear 0-1 optimization problem. We then introduce an optimization heuristic based on betweenness centrality that can rapidly find high-quality interdiction solutions by providing a global view of the network.

math.OC↗

arXiv2010

Targeting by Transnational Terrorist Groups

Many successful terrorist groups operate across international borders where different countries host different stages of terrorist operations. Often the recruits for the group come from one country or countries, while the targets of the operations are in another. Stopping such attacks is difficult because intervention in any region or route might merely shift the terrorists elsewhere. Here we propose a model of transnational terrorism based on the theory of activity networks. The model represents attacks on different countries as paths in a network. The group is assumed to prefer paths of lowest cost (or risk) and maximal yield from attacks. The parameters of the model are computed for the Islamist-Salafi terrorist movement based on open source data and then used for estimation of risks of future attacks. The central finding is that the USA has an enduring appeal as a target, due to lack of other nations of matching geopolitical weight or openness. It is also shown that countries in Africa and Asia that have been overlooked as terrorist bases may become highly significant threats in the future. The model quantifies the dilemmas facing countries in the effort to cut such networks, and points to a limitation of deterrence against transnational terrorists.

cs.GT↗

arXiv2010

Optimizing topological cascade resilience based on the structure of terrorist networks

Complex socioeconomic networks such as information, finance and even terrorist networks need resilience to cascades - to prevent the failure of a single node from causing a far-reaching domino effect. We show that terrorist and guerrilla networks are uniquely cascade-resilient while maintaining high efficiency, but they become more vulnerable beyond a certain threshold. We also introduce an optimization method for constructing networks with high passive cascade resilience. The optimal networks are found to be based on cells, where each cell has a star topology. Counterintuitively, we find that there are conditions where networks should not be modified to stop cascades because doing so would come at a disproportionate loss of efficiency. Implementation of these findings can lead to more cascade-resilient networks in many diverse areas.

math.CO↗

arXiv2010

Monotonic and Non-Monotonic Epidemiological Models on Networks

Contact networks can significantly change the course of epidemics, affecting the rate of new infections and the mean size of an outbreak. Despite this dependence, some characteristics of epidemics are not contingent on the contact network and are probably predictable based only on the pathogen. Here we consider SIR-like pathogens and give an elementary proof that for any network increasing the probability of transmission increases the mean outbreak size. We also introduce a simple model, termed 2FleeSIR, in which susceptibles protect themselves by avoiding contacts with infectees. The 2FleeSIR model is non-monotonic: for some networks, increasing transmissibility actually decreases the final extent. The dynamics of 2FleeSIR are fundamentally different from SIR because 2FleeSIR exhibits no outbreak transition in densely-connected networks. We show that in non-monotonic epidemics, public health officials might be able to intervene in a fundamentally new way to change the network so as to control the effect of unexpectedly-high virulence. However, interventions that decrease transmissibility might actually cause more people to become infected.

math.PR↗