Knowledge-Graph-Gated Defactualization for Style-Controllable and Fact-Preserving Generation in Agentic Conversational AI - Abstract

Sep 15, 2026

Agentic large language models in fact-sensitive customer-support settings must meet two simultaneous goals: maintain factual correctness while producing responses with controllable, task-specific stylistic registers. Activation steering supports finetuning-free style control by perturbing hidden representations, but it lacks an explicit mechanism to separate verifiable facts from stylistic content, causing semantic leakage. The Defactualize-Steer-Rehydrate (DSR) framework combines a typed, salience-weighted knowledge graph with activation steering by replacing salient entities with typed placeholders and deterministically rehydrating verified values after generation. Experiments across six LLaMA-family models on A2A-generated cases, with ablations, show significantly improved verified-entity recovery while preserving effective style control.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

AccessFree to download

Abstract

Index Terms

I. Introduction

Knowledge-Graph-Gated Defactualization for Style-Controllable and Fact-Preserving Generation in Agentic Conversational AI Tanmay Kumar Shrivastava, Darsh Rohit Nandu, and Rajesh Kumar Mundotiya Member, IEEE arXiv : 2608 .20393v 1 [cs .CL] 1 Jul 2026 Abstract—Agentic large language models (LLMs) deployed in fact-sensitive applications such as customer support must simultaneously preserve factual correctness and generate responsesin a controllable stylistic register. Activation steering enables finetuning-free style control by perturbing hidden representations, but it lacks an explicit mechanism for distinguishing verifiable facts from stylistic content, leading to semantic leakage. We address this challenge through Defactualize-Steer-Rehydrate (DSR), a knowledge-engineering framework that integrates a typed, salience-weighted knowledge graph (KG) with activation steering. DSR extracts salient entities using a layered regex or NER or lexical-classifier pipeline, replaces them with typed placeholders prior to steering, and deterministically restores verified values through salience-guided rehydration after generation. DSR is evaluated across six LLaMA-family models (1B–13B parameters) on 600 A2A-generated customer-support cases (1,200 generations), with a dedicated KG ablation study. DSR significantly increases verified-entity recovery relative to a steering-only baseline (Cohen’s d = 0 .225, pBonf = 1 .0 × 10 −4), though the absolute recovery rate remains modest, while preserving effective style control across diverse model families. Layer-wise separability and steering-strength diagnostics further show previously unexplored interactions between representation-level steering and factual grounding. hese results demonstrate that explicit knowledge engineering can systematically enhance trustworthy, controllable, and reproducible generative AI without requiring model finetuning. Code, cached steering vectors, and evaluation scripts are publicly released to support reproducibility.1 Index Terms—Agentic AI, Knowledge Graphs, Knowledge Engineering, Activation Steering, Controllable Text Generation, Hallucination Mitigation, Neuro-Symbolic AI, Trustworthy AI. I. INTRODUCTION AGENTIC large language model (LLM) deployments [14], [15] increasingly generate text under two simultaneous constraints: an exact factual schema supplied by a surrounding system-an order identifier, a product name, a reported defectand a stylistic register requested by a downstream policyan empathetic tone for a distressed customer, a formal tone for an escalation. Fine-tuning [11], [12], retrieval-augmented T. K. Shrivastava and D. R. Nandu are with the Department of Computer Science and Engineering, MATRA Lab (Multimodal and Multilingual AI for Translational Research and Applications Lab), Indian Institute of Technology (IIT) Bhilai, Bhilai, Chhattisgarh, India (e-mail: [tanmayku@iitbhilai.ac.in](tanmayku@iitbhilai.ac.in)). R. K. Mundotiya is an Assistant Professor with the Department of Computer Science and Engineering, MATRA Lab, Indian Institute of Technology (IIT) Bhilai, Bhilai, Chhattisgarh, India. Corresponding author (e-mail: [rmundotiya@iitbhilai.ac.in](rmundotiya@iitbhilai.ac.in)). Manuscript submitted to the IEEE Transactions on Knowledge and Data Engineering Special Issue on Data and Knowledge Empowered Generative Artificial Intelligence. 1[https://github.com/Tanmay-IITDSAI/KG-Gated-Defactualization](https://github.com/Tanmay-IITDSAI/KG-Gated-Defactualization) generation (RAG) [6], and representation-level activation steering [1], [2] each address part of this problem, but none treats the factual half as a first-class constraint. Fine-tuning embeds a register into model weights at the cost of one training run per style and per model [11] . RAG retrieves supporting passages but leaves the generator free to paraphrase, recombine, or drop the entities inside them [13] . Activation steering, which adds a fixed direction to the tr

FAQs

What problem does the paper address in agentic conversational AI?

It addresses the need to preserve factual correctness while generating responses with controllable styles in fact-sensitive deployments like customer support. It highlights that activation steering alone can cause semantic leakage between style signals and fact-bearing content.

How does the Defactualize-Steer-Rehydrate (DSR) framework work?

DSR integrates a typed, salience-weighted knowledge graph with activation steering. It extracts salient entities, replaces them with typed placeholders before steering, then deterministically restores verified values via salience-guided rehydration after generation.

What is the evaluation setup and main outcome?

The framework is evaluated on six LLaMA-family models (1B–13B) using 600 A2A-generated customer-support cases with 1,200 total generations, plus a dedicated KG ablation study. DSR improves verified-entity recovery versus a steering-only baseline while maintaining effective style control.

Sep 15, 2026

Uploaded by Evangeline