Hallucination as a Feature, not a Defect - Evaluating a multi-agent architecture to transform speculative language-model outputs into testable scientific hypotheses

Sep 15, 2026

Uploaded by Levi

Contemporary Large Language Models (LLMs) are increasingly aligned to suppress hallucinations, emphasizing factual retrieval over combinatorial creativity. While this reduces misinformation, it can hinder speculative R&D by encouraging semantic overfitting and diversity collapse. This paper proposes a Rust-based multi-agent orchestration that contrasts narrative daydreaming with executive control via an Epistemological Friction loop between a high-entropy generator and a web-grounded evaluator. Experiments generate diverse, viability-rated hypotheses and compare ablations against direct prompting and self-reflection, showing advantages when hypotheses face strong physical, empirical, or institutional constraints.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages25

Reading Time1 h 3 min

AccessFree to download

Introduction

Background

Working hypothesis and research objectives

Working hypothesis

Objectives

arXiv : 2608 . 19206v 1 [cs .CL] 1 1 Jun 2026 Hallucination as a Feature, not a Defect Evaluating a multi-agent architecture to transform speculative language-model outputs into testable scientific hypotheses Nicolás Rodríguez Álvarez IES Parquesol, Valladolid, Castilla y León, Spain Technology and Engineering 2026 Abstract Contemporary Large Language Models (LLMs) are increasingly aligned to suppress hallucinations, prioritizing factual retrieval over combinatorial creativity. While crucial for mitigating misinformation, this alignment may also restrict speculative Research and Development (R&D) by encouraging what this work operationally treats as semantic overfitting and diversity collapse. In this paper, we propose a Rust-based multi-agent orchestration that uses the contrast between narrative daydreaming and executive control as a functional analogy, not as aneurocognitive claim. The system instigates an Epistemological Friction loop between a high-entropy generating agent and a web-grounded evaluating agent, mediated by a low-entropy semantic bottleneck intended to reduce noise and repetition. Initial experiments generated diverse, viability-rated hypotheses across physical and socialscience domains. We additionally report an exploratory paired baseline and ablation study comparing the full system against direct prompting, self-reflection, removal of the semantic filter, removal of search grounding, and removal of lateral lenses. The results place direct prompting among the weakest conditions across most observed metrics, but they do not show a general superiority of the full system over simple self-reflection. Instead, they suggest that each architecture shifts the balance between originality, feasibility, diversity, and empirical grounding in different ways, and that the full system provides its main advantages when hypotheses must survive strong physical, empirical, or institutional constraints. These findings do not show that hallucination is useful in isolation; they suggest that speculative generation gains value only when constrained by architecture, empirical grounding, and explicit evaluation. Keywords: large language models; hallucination; computational creativity; multi-agent systems; scientific hypothesis generation. Note on language versions. Following arXiv’s multilingual submission guidelines, this document contains the full English version first, followed by the complete Spanish version (versión en español) . 1 Introduction As a conceptual starting point, this work draws a functional analogy from the literature on intelligence and cognition: the tension between narrative, associative, or contemplative processes and goal-directed control processes (Nilsson, 2009) . The reference to the Default Mode Network (DMN) is not used here as a strong neuroscientific hypothesis, but as an operational metaphor to separate two computational functions: broad semantic exploration and evaluative constraint. To this end, we propose a multi-agent architecture that translates this analogy into a technical design. A generating agent produces high-entropy associations; other modules introduce empirical critique, semantic filtering, and final evaluation. The aim is not to claim that LLMs reproduce human cognition, but to test whether an explicit separation between speculation and control improves the production of structured hypotheses. 2 Background Paradoxically, the recent trajectory of artificial intelligence research attempts, with all its resources, to suppress an inherent capability of current Large Language Models: the propensity to “hallucinate”(Chen and Shu, 2024; Ji et al., 2023) . The major laboratories perceive this phenomenon fundamentally as a vector that aggravates AI-generated misinformation, producing responses that diverge from truthful information. As a result, substantial efforts have been deployed in alignment, reinforcement learning from human feedback (RLHF), and rigorou

FAQs

What problem does the paper address about hallucinations in LLMs?

The paper addresses how alignment efforts suppress hallucinations by treating them as defects, which can reduce the model’s ability to make divergent, out-of-distribution combinatorial ideas. This may limit speculative R&D and diversity.

What architecture does the paper propose?

It proposes a Rust-based multi-agent orchestration that separates a high-entropy generating agent from a web-grounded evaluating agent. An “Epistemological Friction” loop uses a semantic bottleneck to reduce noise and repetition.

What do the experiments suggest about when speculative generation is valuable?

The results do not show the full system is generally superior to simple self-reflection, and direct prompting performs weakest on many metrics. The main advantages appear when hypotheses must survive strong physical, empirical, or institutional constraints, meaning gains require explicit evaluation and grounding.

Sep 15, 2026

Uploaded by Levi