arXiv subjects

Oded Margalit

Publications and source records attributed to Oded Margalit.

6 recordsLinked to original sources

arXiv2026

Governed Shared Memory for Multi-Agent LLM Systems

Multi-agent LLM environments require robust mechanisms for shared knowledge management. This paper formalizes the fleet-memory problem and identifies four foundational failure modes: unauthorized leakage, stale propagation, contradiction persistence, and provenance collapse. To address these, we define explicit systems-level primitives: scoped retrieval, temporal supersession, provenance tracking, and policy-governed memory propagation. These primitives are implemented in MemClaw, a production multi-tenant memory service, and evaluated via ArgusFleet, a reproducible harness testing four governance dimensions. Rather than a baseline comparison, this study measures a live production service, emphasizing real-world architectural insights and negative results. Key Evaluation Results Provenance: Successfully reconstructed 100% of depth-four derivation chains with correct writer identity at sub-second per-hop latency. Propagation: Demonstrated high intra-fleet visibility with zero cross-fleet leakage. Under strong write mode, write-to-visible latency was optimized to a single search round-trip. Production Architectural Issues Discovered Asymmetric Scope Enforcement: Tenant isolation held, but sub-tenant scope was initially bypassed on direct GET-by-id requests for agent-scoped credentials (disclosed and remediated during the study). Pipeline Ordering Conflict: While contradiction supersession works for admitted writes, a synchronous near-duplicate gate can prematurely reject contradictory writes before the asynchronous contradiction detector can evaluate them. Conclusion: Long-context retrieval alone is insufficient for production multi-agent memory. Governed shared memory demands explicit systems-level abstractions, and live evaluation is vital to expose enforcement and pipeline-ordering failures missed by design-only treatments.

cs.AI↗

arXiv2026

PeerRank: Autonomous LLM Evaluation Through Web-Grounded, Bias-Controlled Peer Review

Evaluating large language models typically relies on human-authored benchmarks, reference answers, and human or single-model judgments, approaches that scale poorly, become quickly outdated, and mismatch open-world deployments that depend on web retrieval and synthesis. We introduce PeerRank, a fully autonomous end-to-end evaluation framework in which models generate evaluation tasks, answer them with category-scoped live web grounding, judge peer responses and aggregate dense peer assessments into relative performance estimates, without human supervision or gold references. PeerRank treats evaluation as a multi-agent process where each model participates symmetrically as task designer, respondent, and evaluator, while removing biased judgments. In a large-scale study over 12 commercially available models and 420 autonomously generated questions, PeerRank produces stable, discriminative rankings and reveals measurable identity and presentation biases. Rankings are robust, and mean peer scores agree with Elo. We further validate PeerRank on TruthfulQA and GSM8K, where peer scores correlate with objective accuracy. Together, these results suggest that bias-aware peer evaluation with selective web-grounded answering can scale open-world LLM assessment beyond static and human curated benchmarks.

cs.AI↗

arXiv2025

Black-Box Bug-Amplification for Multithreaded Software

Bugs, especially those in concurrent systems, are often hard to reproduce because they manifest only under rare conditions. Testers frequently encounter failures that occur only under specific inputs, even when occurring with low probability. We propose an approach to systematically amplify the occurrence of such elusive bugs. We treat the system under test as a black-box and use repeated trial executions to train a predictive model that estimates the probability of a given input configuration triggering a bug. We evaluate this approach on a dataset of 17 representative concurrency bugs spanning diverse categories. Several model-based search techniques are compared against a brute-force random sampling baseline. Our results show that an ensemble of regression models can significantly increase bug occurrence rates across nearly all scenarios, often achieving an order-of-magnitude improvement over random sampling. The contributions of this work include: (i) a novel formulation of bug-amplification as a rare-event regression problem; (ii) an empirical evaluation of multiple techniques for amplifying bug occurrence, demonstrating the effectiveness of model-guided search; and (iii) a practical, non-invasive testing framework that helps practitioners expose hidden concurrency faults without altering the internal system architecture.

cs.SE↗

arXiv2024

Evolving Assembly Code in an Adversarial Environment

In this work, we evolve Assembly code for the CodeGuru competition. The goal is to create a survivor -- an Assembly program that runs the longest in shared memory, by resisting attacks from adversary survivors and finding their weaknesses. For evolving top-notch solvers, we specify a Backus Normal Form (BNF) for the Assembly language and synthesize the code from scratch using Genetic Programming (GP). We evaluate the survivors by running CodeGuru games against human-written winning survivors. Our evolved programs found weaknesses in the programs they were trained against and utilized them. To push evolution further, we implemented memetic operators that utilize machine learning to explore the solution space effectively. This work has important applications for cyber-security as we utilize evolution to detect weaknesses in survivors. The Assembly BNF is domain-independent; thus, by modifying the fitness function, it can detect code weaknesses and help fix them. Finally, the CodeGuru competition offers a novel platform for analyzing GP and code evolution in adversarial environments. To support further research in this direction, we provide a thorough qualitative analysis of the evolved survivors and the weaknesses found.

cs.NE↗

arXiv2023

Generalized Coverage Criteria for Combinatorial Sequence Testing

We present a new model-based approach for testing systems that use sequences of actions and assertions as test vectors. Our solution includes a method for quantifying testing quality, a tool for generating high-quality test suites based on the coverage criteria we propose, and a framework for assessing risks. For testing quality, we propose a method that specifies generalized coverage criteria over sequences of actions, which extends previous approaches. Our publicly available tool demonstrates how to extract effective test suites from test plans based on these criteria. We also present a Bayesian approach for measuring the probabilities of bugs or risks, and show how this quantification can help achieve an informed balance between exploitation and exploration in testing. Finally, we provide an empirical evaluation demonstrating the effectiveness of our tool in finding bugs, assessing risks, and achieving coverage.

cs.SE↗

arXiv2022

The Normalized Edit Distance with Uniform Operation Costs is a Metric

We prove that the normalized edit distance proposed in [Marzal and Vidal 1993] is a metric when the cost of all the edit operations are the same. This closes a long standing gap in the literature where several authors noted that this distance does not satisfy the triangle inequality in the general case, and that it was not known whether it is satisfied in the uniform case where all the edit costs are equal. We compare this metric to two normalized metrics proposed as alternatives in the literature, when people thought that Marzal's and Vidal's distance is not a metric, and identify key properties that explain why the original distance, now known to also be a metric, is better for some applications. Our examination is from a point of view of formal verification, but the properties and their significance are stated in an application agnostic way.

cs.FL↗