---
title: "A Hybrid RAG Framework for Deceptive... | SI Paper Desk"
url: https://thesimind.si/papers/2611.24049/
description: "The authors combine a small fine-tuned classifier, a search step and a language model to spot fake reviews, reporting an F1 of 0.90 across domains and..."
---

Number 3 of 12 · November 20, 2026

# A Hybrid RAG Framework for Deceptive Review Detection: Integrating RoBERTa-LoRA, FAISS-Based Retrieval, and LLM Reasoning

The authors combine a small fine-tuned classifier, a search step and a language model to spot fake reviews, reporting an F1 of 0.90 across domains and answers backed by evidence.

Andrei Kostin, José Ángel Diaz-Garcia and Karel Gutiérrez-Batista · Submitted November 18, 2026 · arXiv 2611.24049v1 · cs.CL

[Read the paper on arXiv](https://arxiv.org/abs/2611.24049) · [The PDF](https://arxiv.org/pdf/2611.24049) · [The day's 12](https://thesimind.si/papers/day/2026-11-20/)

## What they found

The authors built a system to tell deceptive (spam) reviews from real ones. It joins three parts: RoBERTa tuned with LoRA (a cheap way to adapt a model) to classify, FAISS to find similar reviews by meaning, and a large language model that reasons to a final answer. They tested it on balanced and severely imbalanced real-world cases. After five epochs of fine-tuning, it reached an F1-score of 0.90 when tested on a different domain. The abstract gives the OpenAI Detector baseline as F1 = 0.82 and the specialized in-domain model fakeRoBERTa as F1 = 0.97. The authors also report statistically significant precision gains and overall detection performance comparable to the alternatives.

## Why it matters

Fake reviews can mislead shoppers and enable fraud. A detector that shows the evidence behind its call is easier for people to check than one that gives only a label. The abstract does not say how the explanations were judged, which datasets were used, or how large the precision gains were. Those details would matter for trusting the result.

## How it changes the race

This is an incremental, applied result. It shows that a small tuned classifier plus retrieval and an LLM can do reasonably well on a new domain without full retraining. It does not change what frontier SI systems can do.

## Safety and control

The paper bears on misuse and trust: it aims to detect deceptive text online and to give evidence-backed explanations. The abstract does not test attacks, such as fake reviews written to evade the detector.

### The abstract, by the authors

The growth of deceptive online opinions poses a critical risk of misinformation and fraud. This paper proposes a Retrieval-Augmented Generation (RAG) system to automatically distinguish spam reviews. The architecture integrates RoBERTa with Low-Rank Adaptation (LoRA) classification, FAISS semantic vector retrieval, and a Large Language Model (LLM) for final reasoning. The system was evaluated across balanced and severely imbalanced real-world scenarios. Results demonstrate that five epochs of fine-tuning enable robust domain transfer, achieving an F1-score of 0.90 in cross-domain evaluation, outperforming baselines (OpenAI Detector,F1 = 0.82) and demonstrating competitive performance against specialized in-domain SOTA models (fakeRoBERTa,F1 = 0.97). The proposed architecture provides evidence-backed explainability and statistically significant precision gains while maintaining comparable overall detection performance.

Quoted from arXiv:2611.24049. The research and its words belong to its authors.

Desk score **83** out of 100

Why it was picked

- Safety (in the title)
- Reasoning (in the title)
- Interpretability
- Evals
- Language models (in the title)

Topics

- [Retrieval](https://thesimind.si/papers/?tag=retrieval)
- [Language models](https://thesimind.si/papers/?tag=language-models)
- [Security](https://thesimind.si/papers/?tag=security)
- [Evals](https://thesimind.si/papers/?tag=evals)
- [Society](https://thesimind.si/papers/?tag=society)

**arXiv ID:** 2611.24049v1
**Submitted:** November 18, 2026
**Fields:** cs.CL (Computation and Language)
**Level:** Intermediate

Fingerprint, SHA-256

`23075231ada10fbde9a14f1991778057f3bfe51752fb8a6289e24757ee2629a8`

On Solana: [Proof of Mind seal, 4JtzxB...6w5n](https://solscan.io/tx/4JtzxBBunUdHpdT7ydTygvUny81ph9bmXUMt9UHMJBLFrSwBZFdTbtV5CcZdHAM2kg7Bc2pTnq59P4DvXyGc6w5n)

### The exact text fingerprinted

{"arxiv":"2611.24049","version":1,"title":"A Hybrid RAG Framework for Deceptive Review Detection: Integrating RoBERTa-LoRA, FAISS-Based Retrieval, and LLM Reasoning","authors":["Andrei Kostin","José Ángel Diaz-Garcia","Karel Gutiérrez-Batista"],"day":"2026-11-20","rank":3,"score":83,"status":"written","summary":{"takeaway":"The authors combine a small fine-tuned classifier, a search step and a language model to spot fake reviews, reporting an F1 of 0.90 across domains and answers backed by evidence.","found":"The authors built a system to tell deceptive (spam) reviews from real ones. It joins three parts: RoBERTa tuned with LoRA (a cheap way to adapt a model) to classify, FAISS to find similar reviews by meaning, and a large language model that reasons to a final answer. They tested it on balanced and severely imbalanced real-world cases. After five epochs of fine-tuning, it reached an F1-score of 0.90 when tested on a different domain. The abstract gives the OpenAI Detector baseline as F1 = 0.82 and the specialized in-domain model fakeRoBERTa as F1 = 0.97. The authors also report statistically significant precision gains and overall detection performance comparable to the alternatives.","matters":"Fake reviews can mislead shoppers and enable fraud. A detector that shows the evidence behind its call is easier for people to check than one that gives only a label. The abstract does not say how the explanations were judged, which datasets were used, or how large the precision gains were. Those details would matter for trusting the result.","race":"This is an incremental, applied result. It shows that a small tuned classifier plus retrieval and an LLM can do reasonably well on a new domain without full retraining. It does not change what frontier SI systems can do.","safety":"The paper bears on misuse and trust: it aims to detect deceptive text online and to give evidence-backed explanations. The abstract does not test attacks, such as fake reviews written to evade the detector.","level":"intermediate","tags":["retrieval","language-models","security","evals","society"]}}

[Check it on Verify a thought](https://thesimind.si/proof-of-mind/)

## The authors

Andrei Kostin, José Ángel Diaz-Garcia, Karel Gutiérrez-Batista

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.

Related on the desk

## More papers on these topics

[Nov 16, 2026 · Number 5AI-Voyager: A Public-Interest System for Human-AI Civic Reasoning to Contest and Shape the Rules That Govern AIAI-Voyager is a tool that helps ordinary people read, question and propose changes to government rules on SI, using search that points back to the exact text of public documents.Policy and lawRetrievalAgents](https://thesimind.si/papers/2611.17564/) · [Nov 12, 2026 · Number 5EJAI: Judgment-Preserving Civic AI for Data Center Environmental JusticeEJAI is a SI tool that shows evidence, missing data and two opposing viewpoints so residents can judge data center disputes themselves, though its personas and source mix raise fairness concerns.SocietyPolicy and lawRetrieval](https://thesimind.si/papers/2611.13525/) · [Nov 16, 2026 · Paper of the DayOnly as Safe as the Weakest Agent: A Break That Appears Only When the System Is Multi-AgentWhen a strong SI model hands work to a cheaper sub-agent, the whole system can be only as safe as the weaker agent, even if the strong model alone resists the attack.SafetyMulti-agent systemsSecurity](https://thesimind.si/papers/2611.17520/)

[Number 2](https://thesimind.si/papers/2611.24313/) · [All 12 for Nov 20, 2026](https://thesimind.si/papers/day/2026-11-20/) · [Number 4](https://thesimind.si/papers/2611.24840/)
