arXiv subjects

Domagoj Matijević

Publications and source records attributed to Domagoj Matijević.

5 recordsLinked to original sources

arXiv2026

ARdena: Scenario-driven control of real-time LLM agents

Large language models (LLMs) have enabled increasingly capable conversational agents, but reliably controlling their behavior in real-time interactive environments remains a significant challenge. Existing approaches often rely on model fine-tuning or alignment procedures that are difficult to adapt to changing interaction requirements. This paper introduces layered scenario-driven LLM control, a framework that enables runtime behavior control through structured prompting. By combining persistent context with scenario-specific constraints, the approach allows agent behavior to be modified during interaction without changing the underlying model. The framework is implemented in ARDena, a real-time multimodal embodied agent that integrates speech interaction, visual perception, tool use, and avatar-based response generation. The proposed approach is evaluated with respect to control effectiveness, response latency, and operational stability. The results demonstrate that scenario definitions alone can produce substantially different interaction behaviors while maintaining stable real-time operation, highlighting the effectiveness of scenario-driven prompting for controlling LLM agents.

cs.AI↗

arXiv2026

Learning to Place Guards by Reinforcement: A Geo-Free Neural Policy for the Vertex-Guard Art Gallery Problem

Neural combinatorial optimization (NCO) has shown that policies trained by reinforcement can construct strong solutions to NP-hard problems directly from raw instances. What such a policy actually learns, as opposed to what its decoder expresses, remains much less clear. We study this distinction on the vertex-guard Art Gallery Problem, the NP-hard task of choosing polygon vertices from which to observe an entire region. A pointer-network policy is trained from a coverage-aware reward over its own rollouts under the constraint we call geo-free inference: at test time it sees only vertex coordinates, with no visibility computation and no geometric oracle. The policy places guards economically but leaves a tail of under-covered polygons that widens far beyond the training range. To locate the cause, we freeze the trained encoder and read its embeddings with a small single-shot classifier, still geo-free at inference. The classifier closes most of the feasibility gap, in and out of distribution and at up to roughly five times the training range, cutting under-covered polygons by about an order of magnitude at an explicitly reported cost in guard count. We read this as evidence that the reinforcement-trained representation already encodes the geometry required for feasibility, and that residual failures reflect decoder calibration rather than missing knowledge. Probing a frozen encoder thus offers a practical way to ask what a neural combinatorial solver has internalized.

cs.LG↗

arXiv2023

Compressing Sentence Representation with maximum Coding Rate Reduction

In most natural language inference problems, sentence representation is needed for semantic retrieval tasks. In recent years, pre-trained large language models have been quite effective for computing such representations. These models produce high-dimensional sentence embeddings. An evident performance gap between large and small models exists in practice. Hence, due to space and time hardware limitations, there is a need to attain comparable results when using the smaller model, which is usually a distilled version of the large language model. In this paper, we assess the model distillation of the sentence representation model Sentence-BERT by augmenting the pre-trained distilled model with a projection layer additionally learned on the Maximum Coding Rate Reduction (MCR2)objective, a novel approach developed for general-purpose manifold clustering. We demonstrate that the new language model with reduced complexity and sentence embedding size can achieve comparable results on semantic retrieval benchmarks.

cs.CL↗

arXiv2021

Anti Tai Mapping for Unordered Labeled Trees

The well-studied Tai mapping between two rooted labeled trees $T_1(V_1, E_1)$ and $T_2(V_2, E_2)$ defines a one-to-one mapping between nodes in $T_1$ and $T_2$ that preserves ancestor relationship. For unordered trees the problem of finding a maximum-weight Tai mapping is known to be NP-complete. In this work, we define an anti Tai mapping $M\subseteq V_1\times V_2$ as a binary relation between two unordered labeled trees such that any two $(x,y), (x', y')\in M$ violate ancestor relationship and thus cannot be part of the same Tai mapping, i.e. $(x\le x' \iff y\not \le y') \vee (x'\le x \iff y'\not \le y)$, given an ancestor order $x<x'$ meaning that $x$ is an ancestor of $x'$. Finding a maximum-weight anti Tai mapping arises in the cutting plane method for solving the maximum-weight Tai mapping problem via integer programming. We give an efficient polynomial-time algorithm for finding a maximum-weight anti Tai mapping for the case when one of the two trees is a path and further show how to extend this result in order to provide a polynomially computable lower bound on the optimal anti Tai mapping for two unordered labeled trees. The latter result stems from the special class of anti Tai mapping defined by the more restricted condition $x\sim x' \iff y\not\sim y'$, where $\sim$ denotes that two nodes belong to the same root-to-leaf path. For this class, we give an efficient algorithm that solves the problem directly on two unordered trees in $O(|V_1|^2|V_2|^2)$.

cs.DS↗

arXiv2008

Finding the theta-Guarded Region

We are given a finite set of n points (guards) G in the plane R^2 and an angle 0 < theta < 2 pi. A theta-cone is a cone with apex angle theta. We call a theta-cone empty (with respect to G) if it does not contain any point of G. A point p in R^2 is called theta-guarded if every theta-cone with its apex located at p is non-empty. Furthermore, the set of all theta-guarded points is called the theta-guarded region, or the theta-region for short. We present several results on this topic. The main contribution of our work is to describe the theta-region with O(n/theta) circular arcs, and we give an algorithm to compute it. We prove a tight O(n) worst-case bound on the complexity of the theta-region for theta >= pi/2. In case theta is bounded from below by a positive constant, we prove an almost linear bound O(n^(1+epsilon)) for any epsilon > 0 on the complexity. Moreover, we show that there is a sequence of inputs such that the asymptotic bound on the complexity of their theta-region is Omega(n^2). In addition we point out gaps in the proofs of a recent publication that claims an O(n) bound on the complexity for any constant angle theta.

cs.CG↗