arXiv subjects

Jetlir Duraj

Publications and source records attributed to Jetlir Duraj.

12 recordsLinked to original sources

arXiv2026

Iterating Toward Better Search: A Two-Agent Simulation Framework for Evaluating Agentic Search Architectures in E-Commerce

We present a modular two-agent simulation framework for evaluating conversational shopping assistant architectures. An independent buyer agent, configured with personas, missions, and patience levels, is paired with an interchangeable responder that integrates with a real e-commerce search API. Holding the buyer constant across experiments enables controlled comparison of responder designs on identical scenarios. Using 2011 conversations across 14 persona buckets, we establish four empirical findings. First, rolling-window memory outperforms intent-extraction memory on all quality metrics while being 35% faster per query. Second, illustrating rapid evidence-driven iteration, a systematic failure analysis of a responder version enables targeted fixes that reduce failure and near-failure rates by 62% across the full dataset. Third, swapping the responder LLM backbone from Gemini~2.5 to Llama~3.3~70B costs 0.16--0.45 points despite identical architecture. Finally, we document systematic philosophical disagreement between frontier LLM judges: Gemini rewards process correctness while Claude demands concrete outcomes, despite using the same evaluation prompt.

cs.AI↗

arXiv2026

A Chain-of-Thought Approach to Semantic Query Categorization in e-Commerce Taxonomies

Search in e-Commerce is powered at the core by a structured representation of the inventory, often formulated as a category taxonomy. An important capability in e-Commerce with hierarchical taxonomies is to select a set of relevant leaf categories that are semantically aligned with a given user query. In this scope, we address a fundamental problem of search query categorization in real-world e-Commerce taxonomies. A correct categorization of a query not only provides a way to zoom into the correct inventory space, but opens the door to multiple intent understanding capabilities for a query. A practical and accurate solution to this problem has many applications in e-commerce, including constraining retrieved items and improving the relevance of the search results. For this task, we explore a novel Chain-of-Thought (CoT) paradigm that combines simple tree-search with LLM semantic scoring. Assessing its classification performance on human-judged query-category pairs, relevance tests, and LLM-based reference methods, we find that the CoT approach performs better than a benchmark that uses embedding-based query category predictions. We show how the CoT approach can detect problems within a hierarchical taxonomy. Finally, we also propose LLM-based approaches for query-categorization of the same spirit, but which scale better at the range of millions of queries.

cs.IR↗

arXiv2023

A multi-agent targeted trading equilibrium with transaction costs

We prove the existence of a continuous-time Radner equilibrium with multiple agents and transaction costs. The agents are incentivized to trade towards a targeted number of shares throughout the trading period and seek to maximize their expected wealth minus a penalty for deviating from their targets. Their wealth is further reduced by transaction costs that are proportional to the number of stock shares traded. The agents' targeted number of shares is publicly known, making the resulting equilibrium fully revealing. In equilibrium, each agent optimally chooses to trade for an initial time interval before stopping trade. Our equilibrium construction and analysis involves identifying the order in which the agents stop trade. The transaction cost level impacts the equilibrium stock price drift. We analyze the equilibrium outcomes and provide numerical examples.

q-fin.MF↗

arXiv2023

Dynamic Information Design with Diminishing Sensitivity Over News

A Bayesian agent experiences gain-loss utility each period over changes in belief about future consumption ("news utility"), with diminishing sensitivity over the magnitude of news. Diminishing sensitivity induces a preference over news skewness: gradual bad news, one-shot good news is worse than one-shot resolution, which is in turn worse than gradual good news, one-shot bad news. So, the agent's preference between gradual information and one-shot resolution can depend on his consumption ranking of different states. In a dynamic cheap-talk framework where a benevolent sender communicates the state over multiple periods, the babbling equilibrium is essentially unique without loss aversion. More loss-averse agents may enjoy higher news utility in equilibrium, contrary to the commitment case. We characterize the family of gradual good news equilibria that exist with high enough loss aversion, and find the sender conveys progressively larger pieces of good news. We discuss applications to media competition and game shows.

econ.TH↗

arXiv2020

Invariance principles for integrated random walks conditioned to stay positive

Let $S(n)$ be a centered random walk with finite second moment. We consider the integrated random walk $T(n) = S(0)+S(1)+\dots+S(n)$. We prove invariance principles for the meander and for the bridge of this process, under the condition that the integrated random walk remains positive. Furthermore, we prove the functional convergence of its Doob's $h$-transform to the $h$-transform of the Kolmogorov diffusion conditioned to stay positive.

math.PR↗

arXiv2020

Martin boundary of random walks in convex cones

We determine the asymptotic behavior of the Green function for zero-drift random walks confined to multidimensional convex cones. As a consequence, we prove that there is a unique positive discrete harmonic function for these processes (up to a multiplicative constant); in other words, the Martin boundary reduces to a singleton.

math.PR↗

arXiv2018

Mechanism Design with News Utility

News utility is the idea that the utility of an agent depends on changes in her beliefs over consumption and money. We introduce news utility into otherwise classical static Bayesian mechanism design models. We show that a key role is played by the timeline of the mechanism, i.e. whether there are delays between the announcement stage, the participation stage, the play stage and the realization stage of a mechanism. Depending on the timing, agents with news utility can experience two additional news utility effects: a surprise effect derived from comparing to pre-mechanism beliefs, as well as a realization effect derived from comparing post-play beliefs with the actual outcome of the mechanism. We look at two distinct mechanism design settings reflecting the two main strands of the classical literature. In the first model, a monopolist screens an agent according to the magnitude of her loss aversion. In the second model, we consider a general multi-agent Bayesian mechanism design setting where the uncertainty of each player stems from not knowing the intrinsic types of the other agents. We give applications to auctions and public good provision which illustrate how news utility changes classical results. For both models we characterize the optimal design of the timeline. A timeline featuring no delay between participation and play but a delay in realization is never optimal in either model. In the screening model the optimal timeline is one without delays. In auction settings, under fairly natural assumptions the optimal timeline has delays between all three stages of the mechanism.

econ.TH↗

arXiv2018

Dynamic Random Subjective Expected Utility

Dynamic Random Subjective Expected Utility (DR-SEU) allows to model choice data observed from an agent or a population of agents whose beliefs about objective payoff-relevant states and tastes can both evolve stochastically. Our observable, the augmented Stochastic Choice Function (aSCF) allows, in contrast to previous work in decision theory, for a direct test of whether the agent's beliefs reflect the true data-generating process conditional on their private information as well as identification of the possibly incorrect beliefs. We give an axiomatic characterization of when an agent satisfies the model, both in a static as well as in a dynamic setting. We look at the case when the agent has correct beliefs about the evolution of objective states as well as at the case when her beliefs are incorrect but unforeseen contingencies are impossible. We also distinguish two subvariants of the dynamic model which coincide in the static setting: Evolving SEU, where a sophisticated agent's utility evolves according to a Bellman equation and Gradual Learning, where the agent is learning about her taste. We prove easy and natural comparative statics results on the degree of belief incorrectness as well as on the speed of learning about taste. Auxiliary results contained in the online appendix extend previous decision theory work in the menu choice and stochastic choice literature from a technical as well as a conceptual perspective.

econ.TH↗

arXiv2018

Green function of a random walk in a cone

This paper studies the asymptotic behavior of the Green function of a multidimensional random walk killed when leaving a convex cone with smooth boundary. Our results imply uniqueness, up to a multiplicative factor, of the positive harmonic function for the killed random walk.

math.PR↗

arXiv2015

Invariance principles for random walks in cones

We prove invariance principles for a mulditimensional random walk conditioned to stay in a cone. Our first result concerns convergence towards the Brownian meander in the cone. Furthermore, we prove functional convergence of $h$-transformed random walk to the corresponding $h$-transform of the Brownian motion. Finally, we prove an invariance principle for bridges of a random walk in a cone.

math.PR↗

arXiv2013

On harmonic Functions of killed Random Walks in two dimensional convex Cones

We prove the existence of uncountably many positive harmonic functions for random walks on the euclidean lattice with non-zero drift, killed when leaving two dimensional convex cones with vertex in 0. Our proof is an adaption of the proof for the positive quadrant from [Ignatiouk-Robert, Loree]. We also make the natural conjecture about the Martin boundary for general convex cones in two dimensions. This is still an open problem and here we only indicate where the proof technique for the positive quadrant breaks down.

math.PR↗

arXiv2013

Random Walks in Cones: the Case of Nonzero Drift

We consider multidimensional discrete valued random walks with nonzero drift killed when leaving general cones of the euclidian space. We find the asymptotics for the exit time from the cone and study weak convergence of the process conditioned on not leaving the cone. We get quasistationarity of its limiting distribution. Finally we construct a version of the random walk conditioned to never leave the cone.

math.PR↗