Tanmay Agrawal
2025
Read Tanmay Agrawal (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-00663.txt.
Introduction
Language models hallucinate, and enterprise users need those fabrications caught automatically. Tanmay Agrawal presents structured visualizations for automated hallucination detection in 2025, under the banner of graphing the truth. In enterprise settings, models are augmented with closed-source domain knowledge to give informed answers.
Grounding answers in private documents helps only if unsupported claims are still detected. Turning model statements into structured graphs makes claims checkable against sources, claim by claim. Visual detection tools could let compliance and support teams audit AI answers at scale, rather than sampling them by hand and hoping. Closed-source knowledge makes verification harder, since auditors cannot inspect the whole corpus. Structured graphs also make disagreements between model and source easy to point at.
Catalog record
| YEAR | 2025 |
|---|---|
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 19 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2512-00663
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Tanmay Agrawal Abstract Graphing the Truth: Structured Visualizations for Automated Hallucination Detection in LLMs Abstract Large Language Models have rapidly advanced in their ability to interpret and generate natural language. In enterprise settings, they are frequently augmented with closed-source domain knowledge to deliver more contextually informed responses. However, operational constraints such as limited context windows and inconsistencies between pre-training data and supplied knowledge often lead to hallucinations, some of which appear highly credible and escape routine human review. Current mitigation strategies either depend on costly, large-scale gold-standard Q&A curation or rely on secondary model verification, neither of which offers deterministic assurance. This paper introduces a framework that organizes proprietary knowledge and model-generated content into interactive visual knowledge graphs. The objective is to provide end users with a clear, intuitive view of potential hallucination zones by linking model assertions to underlying sources of truth and indicating confidence levels. Through this visual interface, users can diagnose inconsistencies, identify weak reasoning chains, and supply corrective feedback. The resulting human-in-the-loop workflow creates a structured feedback loop that can enhance model reliability and continuously improve response quality. Key Words: Hallucination Detection, Knowledge Graph Visualization, Retrieval-Augmented Generation, Human-in-the-Loop Auditing, Closed-Source Knowledge Integration 1 Introduction Large Language Models (LLMs) have become central to modern natural language processing due to their strong performance in language understanding and generation [ 13 ] . As organizations increasingly deploy these models in high-stakes, knowledge-centric environments, they often augment LLMs with closed-source datasets to improve factual grounding. While this improves contextual relevance, it also introduces new risks. Conflicts between pre-training distributions and domain-specific knowledge, combined with finite context windows, can lead LLMs to generate hallucinations that appear confident yet factually incorrect [ 7 , 1 , 2 ] . Much of the current research targets hallucination prevention during generation through improved prompting, retrieval-augmented generation, or enhanced training pipelines. However, enterprise use cases frequently require post-hoc auditing of model outputs against authoritative sources, a capability that remains comparatively underdeveloped. Recent work explores several post-hoc detection strategies, including evaluator-model frameworks, explainability-based inspection, and structured graph-based analysis [ 6 , 11 , 9 ] . Systems such as GraphEval [ 9 ] demonstrate the promise of triplet-based comparisons for identifying factual inconsistencies, although limitations in scalability, transparency, and traceability persist. These challenges highlight the need for more interpretable and operationally efficient solutions. Visualization techniques present an opportunity to address these gaps. Prior research shows that visual encodings of semantic relationships can support intuitive inspection, large-scale comparison, and human-in-the-loop validation of textual content [ 10 , 8 , 4 ] . Motivated by these insights, this work proposes a visualization-driven framework that structures closed-source knowledge and LLM-generated statements as entity-relationship graphs. Through interactive visual exploration, the system enables users to identify potential hallucinations, trace their origins, and provide corrective feedback that supports more trustworthy and reliable deployment of LLMs in mission-critical applications. 2 Literature Review Hallucination detection research follows three major directions. LLM-as-a-Judge approaches rely on a separate model to evaluate factual consistency, although they frequently inherit correlated biases and knowledge blind spots [ 6 ] . Explainable AI techniques analyze model rationales, attention patterns, or generated explanations, but empirical evaluations indicate that many such explanations do not accurately capture the true generative process [ 11 ] . Graph-based techniques convert source material and model outputs into structured triplets or entity-relationship graphs, enabling deterministic comparisons and improved transparency. These methods have gained traction due to their reproducibility and interpretability [ 9 ] . GraphEval represents a leading example of structured evaluation. It translates claims and source content into (subject, predicate, object) triplets and identifies factual mismatches through direct comparison [ 9 ] . Although effective, the method faces challenges involving computational overhead from Natural Language Inference pipelines, variability across NLI models, and limited traceability that forces users to manually inspect long prompt contexts. These limitations highlight the need for more interpretable and operationally efficient graph-based solutions. Visualization-oriented research offers complementary insights into how complex textual relationships can be made more transparent. Systems such as WizMap position semantically related embeddings in close proximity and provide multi-resolution exploration of large knowledge spaces [ 10 ] . VADIS uses a circular grid layout to simultaneously encode semantic similarity and query relevance, enabling intuitive navigation of document collections [ 8 ] . Belief graph work by Hase et al. shows how node-edge representations can reveal relationships between model-stored factual assertions, supporting clearer inspection of inconsistencies [ 4 ] . Taken together, the literature points to two unresolved gaps: the need for structured and explainable mechanisms for hallucination detection, and the need for visualization frameworks that help users interpret relationships between generated claims and source knowledge. The present study addresses both gaps by integrating graph-based consistency checking with visualization-driven interpretability. 3 Materials and Methods 3.1 Dataset To evaluate the proposed approach, we used the SummEval dataset [ 3 ] , a widely adopted benchmark for assessing factual consistency in abstractive summarization. SummEval contains human evaluations of outputs from 16 summarization models applied to 100 CNN/DailyMail articles. Each generated summary is rated on a 1–5 Likert scale across four categories: consistency, coherence, fluency, and relevance. Building on prior work, including the TRUE benchmark [ 5 ] , our study focuses specifically on the consistency dimension, where a score of 5 denotes complete factual alignment with the source document, and lower scores indicate varying degrees of inconsistency or hallucination. The dataset comprises 1,600 annotated examples with a consistency-label ratio of 33.2%. On average, model outputs contain 63 tokens, while source documents average 359 tokens, making SummEval an effective benchmark for evaluating hallucination detection methods under realistic summarization conditions. 3.2 Structured Knowledge Representation We extract comprehensive entity-relationships from both closed-source knowledge and LLM-generated outputs using triple extraction tech- niques. Currently, we employ subject-verb-object triplets similar to GraphEval [ 9 ] , but have extended this approach to apply extraction methods to both source context and model outputs. While triple extrac- tion offers a balance between simplicity and effectiveness, we continue to explore richer representation formats including atomic sentences with coreference resolution for capturing more complex relationships. GraphEval+ Framework for Hallucination Detection We developed GraphEval+, an enhanced extension of the baseline GraphEval system, by integrating three primary improvements into the evaluation pipeline. First, the method performs bidirectional triple extraction, generating semantic triples from both the source material and the LLM output to ensure structurally comparable representations. Second, it introduces a semantic similarity matching stage that identifies the most relevant context triples for each generated triple, avoiding the inefficiency of comparing against the full prompt and mitigating known limitations of NLI models when handling long premises. Third, GraphEval+ employs targeted NLI evaluation, assessing each output triple only against its closest semantic matches from the source, resulting in a more focused and contextually grounded measure of factual consistency. Alternative Approaches to Hallucination Detection In addition to our primary GraphEval+ framework, we investigated alternative methodologies for detecting hallucinations in LLM-generated content through the Sentence Isolation plus CoReference Isolation (SICI) approach [ 12 ] . Instead of relying on LLM-driven triple extraction to convert text into subject-verb-object structures, SICI adopts a more direct representation by treating individual sentences as the fundamental units of analysis. The method resolves pronouns and