Mehmet Onur Keskin

Document retrieval increasingly supports high-stakes information access in finance, healthcare, and law. Modern retrieval pipelines vary across two axes: modality (text or multimodal) and architecture (dense or late-interaction). These choices impose a hard compromise: the most effective pipelines are too expensive to deploy at scale, while the fastest fail on complex documents. We show that this compromise is unnecessary, as no static pipeline dominates across 11 financial and scientific benchmarks. We introduce RetrievalRouter, a lightweight query-aware router that predicts, from query text alone, which retrieval pipeline best fits each query. A single tunable parameter exposes the full accuracy-latency frontier, offering operating points that simultaneously surpass static baselines in accuracy and efficiency. Against the best static baseline, RetrievalRouter is 2.5% more accurate and 12.4x faster. Furthermore, compared with prior adaptive strategy selection methods, RetrievalRouter achieves significantly higher nDCG@5 in accuracy-oriented settings while matching or outperforming them in latency-oriented regimes.

Retrieval-augmented generation (RAG) systems traditionally rely on text-based pipelines to extract and retrieve information from documents. While efficient, these systems often struggle with documents where meaning is conveyed through layout, tables, and visual elements. Recent advancements in multimodal pipelines, utilizing advanced vision-language models (VLMs), have significantly improved retrieval quality by jointly encoding visual and textual signals. However, these improvements come with increased memory requirements and higher indexing and retrieval latency. To address these trade-offs, a quantitative, data-driven methodology is proposed for selecting the most suitable RAG pipeline based on empirical performance and resource constraints. This involves evaluating leading contemporary textual and multimodal pipelines, including dense and late-interaction architectures, to analyze their strengths, limitations, and the types of documents they are best equipped to handle. The evaluation focuses on key design dimensions in modern RAG systems: input modality (textual vs. visual) and retrieval architecture (dense vs. late interaction). The study highlights that text-based pipelines typically follow a multi-stage preprocessing procedure, and while efficient, they can lose fine-grained alignment crucial for structured or layout-dependent documents. Multimodal approaches, on the other hand, preserve token-level or patch-level granularity during indexing. The goal is to guide practitioners in identifying the most efficient and accurate pipeline for specific tasks.

The paper introduces a comprehensive human-agent negotiation framework designed to facilitate the development and evaluation of research studies on human-agent negotiation without building each component from scratch. Leveraging the interoperability and reusability of its components, this framework offers various functionalities, including speech-to-text conversion, emotion recognition, a repository of negotiation strategies, and an interaction manager capable of managing gestures designed for Nao, Pepper, and QT, and coordinating message exchanges in a turn-taking fashion. This framework aims to lower the entry barrier for researchers in human-agent negotiation by providing a versatile platform that supports a wide range of research directions, including affective computing, natural language processing, decision-making, and non-verbal communication.

This paper introduces a negotiation framework to solve the Multi-Agent Path Finding (MAPF) Problem for self-interested agents in a decentralized fashion. The framework aims to achieve a good trade-off between the privacy of the agents and the effectiveness of solutions. Accordingly, a token-based bilateral negotiation protocol and two negotiation strategies are presented. The experimental results over four different settings of the MAPF problem show that the proposed approach could find conflict-free path solutions albeit suboptimally, especially when the search space is large and high-density. In contrast, Explicit Estimation Conflict-Based Search (EECBS) struggles to find optimal solutions. Besides, deploying a sophisticated negotiation strategy that utilizes information about local density for generating alternative paths can yield remarkably better solution performance in this negotiation framework.

With the improvement of intelligent systems and robotics, social robots are becoming part of our society. To accomplish complex tasks, robots and humans may need to collaborate, and when necessary, they need to negotiate with each other. While designing such socially interacting robots, it is crucial to consider human factors such as facial expression, emotions, and body language. Since gestures play a crucial role in interaction, this article studies the effect of gestures in human–robot negotiation experiments. Additionally, it compares the performance of variants of the well-known negotiation tactics (i.e., time-based and behavior-based) in automated negotiation literature in the context of human–robot negotiations. Our experimental results support the finding in automated negotiation. That is, the robot gained higher utility when it imitates its opponent’s bidding strategy than employing a time-based negotiation strategy. When adopting a behavior-based technique, there is a statistically significant effect of gestures on the underlying negotiation process, and, therefore, on negotiation outcome.

Negotiation is one of the crucial processes for resolving conflicts between parties. In automated negotiation, agent designers mostly take opponent’s offers and the remaining time into account while designing their strategies. While designing a negotiating agent interacting with a human directly, other information such as opponent’s emotional changes during the negotiation can establish a better interaction and reach an admissible settlement for joint interests. Accordingly, this paper proposes a bidding strategy for humanoid robots, which incorporates their opponents’ emotional states and awareness of the agent’s changing behavior.