🤖 AI

SIRIN: A Unified Toolkit for Detecting Contextual Hallucinations in Retrieval-Augmented and Memory-Grounded LLM Systems

SIRIN is a unified toolkit and interactive web interface that integrates multiple detection paradigms to identify contextual hallucinations and assess query answerability in retrieval-augmented, agentic, and memory-grounded LLM systems.

Original authors: Julia Belikova, Rauf Parchiev, Mikhail Filimonov, Konstantin Polev, Andrey Savchenko, Maksim Makarenko

Published 2026-08-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Julia Belikova, Rauf Parchiev, Mikhail Filimonov, Konstantin Polev, Andrey Savchenko, Maksim Makarenko

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are chatting with a super-smart, super-fast robot that has read almost every book in the library. You ask it a question, and it answers instantly with perfect grammar and a confident tone. But here's the catch: sometimes, that robot is just making things up. It might tell you a historical event happened on the wrong date or that a character in a story did something they never did. This is called a "hallucination." It's not the robot lying on purpose; it's just so good at sounding plausible that it forgets to check its facts.

In the world of Artificial Intelligence, there are two main ways we try to stop this. One way is to give the robot a specific set of notes (called "context" or "evidence") and tell it, "Only answer using these notes." This is like giving a student a reference sheet during a test. The other way is to build a "truth detector" that can sniff out when the robot is drifting away from those notes. For a long time, building these truth detectors has been messy. Some detectors only work if you can peek inside the robot's brain (its internal code), while others only work if you can read its final answer. Some check if the robot is confident, while others act like a strict teacher grading the essay. Because these tools are all built differently, it's been hard to compare them or use them together to keep our AI honest.

Enter SIRIN, a new toolkit created by researchers at Sber AI Lab that acts like a universal translator and a Swiss Army knife for catching these AI lies. Think of SIRIN as a "Hallucination Control Tower." Instead of forcing you to choose between different types of truth detectors, SIRIN brings them all under one roof. It unifies three very different ways of checking the AI:

The "Brain Scan" (Probing): This looks inside the robot's brain while it's thinking to see if it's hesitating or confused, even before it speaks.
The "Strict Teacher" (Judge): This uses a second, independent AI to read the answer and the notes side-by-side to see if they match.
The "Confidence Meter" (Uncertainty): This measures how shaky the robot feels about its answer, like a nervous student who isn't sure if they studied enough.

SIRIN doesn't just stop at catching lies after the robot speaks. It also checks before the robot starts writing to see if the notes it has are actually enough to answer the question at all. If the notes are missing, SIRIN says, "Stop! You can't answer this yet," preventing the robot from guessing.

The researchers found that by combining these methods, they could build a much stronger safety net. When they tested SIRIN on various tasks, the "Strict Teacher" method (using a trained AI judge) was the most accurate at spotting lies, especially when they had lots of examples to learn from. However, the "Brain Scan" method was surprisingly good at catching errors even when they didn't have many examples to study, making it a lightweight and efficient option.

Perhaps the most exciting discovery was how SIRIN works as a "gatekeeper" for long-term memory systems. Imagine an AI agent that remembers your conversations over weeks. If that agent misremembers a detail, it might build a whole new false memory on top of it, compounding the error. The researchers showed that by using SIRIN as a checkpoint before the agent answers, they could boost the accuracy of the answers it did give from about 62% to nearly 79%, while cutting the number of completely made-up facts in half.

In short, the paper suggests that catching AI hallucinations isn't about finding one perfect detector. Instead, it's about building a flexible system that can mix and match different types of checks. SIRIN proves that when you unify these tools into a single, easy-to-use interface, you can make AI systems significantly more trustworthy, catching unsupported claims and highlighting exactly where the robot got it wrong, all without needing to rewrite the AI's core code.

Technical Summary: SIRIN – A Unified Toolkit for Detecting Contextual Hallucinations

Problem Statement

Large Language Models (LLMs) frequently generate fluent and plausible responses that are not grounded in the provided evidence, a failure mode known as contextual hallucination. This issue is particularly critical in Retrieval-Augmented Generation (RAG) and agentic systems where models manage long-term memory, as ungrounded claims can compound over time.

Currently, the landscape of hallucination detection is fragmented. Existing open-source toolkits (e.g., LM-Polygraph, RAGAS, DeepEval) typically support only a subset of detection paradigms, operate at specific granularities (response-level vs. span-level), or lack unified interfaces for comparing different methods. Furthermore, most tools focus solely on post-generation faithfulness, neglecting the complementary task of query answerability (determining if context is sufficient before generation). This fragmentation makes it difficult to combine complementary signals, benchmark detectors, or deploy them as a unified trust layer in production systems.

Methodology: The SIRIN Framework

SIRIN (Semantic Inconsistency Recognition and Inspection Nexus) addresses this fragmentation by treating hallucination detection as an infrastructure problem. It provides a unified, plug-in toolkit that integrates three distinct detector paradigms and a pre-generation task under a single interface, configuration system, and evaluation pipeline.

1. Unified Detector Families

SIRIN groups detectors into three families, supporting both white-box (access to internal model states) and black-box (text-only) settings:

Probing (White-box): Trains lightweight classifiers (e.g., linear heads, CatBoost, TabPFN) over frozen internal features such as hidden states, attention maps, and token probabilities. These probes separate faithful from hallucinated outputs based on representation signals.
Judge-style (Black-box/Text-only): Relies on observable inputs and outputs. This includes encoder-based classifiers (e.g., ModernBERT, DeBERTa) and LLM-as-a-judge verification (e.g., Qwen, GPT-5.4-mini) to check for semantic consistency or contradictions between evidence and the answer.
Uncertainty Estimation (UE): Wraps the LM-Polygraph library to expose various uncertainty estimators (likelihood-based, sampling-based, attention-aware) that score generations based on model confidence patterns (e.g., entropy, perplexity).

2. Pipeline Architecture

SIRIN implements a configurable pipeline that processes input triples $(c, q, a)$ for hallucination detection or pairs $(c, q)$ for answerability.

Shared Report Schema: All detectors output a standardized report containing a calibrated probability score ( $s \in [0, 1]$ ), a predicted label, and optional evidence (span scores, rationales).
Efficient Processing: For probing, the pipeline extracts features from a frozen LLM in a single forward pass, allowing multiple feature families to be processed simultaneously without redundant model calls.
Granularity: The system supports both response-level scores (for runtime gating) and span-level scores (for inspecting specific unsupported text segments), utilizing a reusable splitter/aggregator.
Extensibility: New detectors are added via three plug points: a feature processor, a detector, or an aggregator. Configuration is handled via typed, Hydra-compatible files, allowing task switching without pipeline rewrites.

3. Interactive Web UI

SIRIN includes a Streamlit-based web interface that allows users to input context-query-answer triples, select detectors from any family, and visualize results. Features include:

Live hallucination scoring.
Highlighting of unsupported spans with per-span risk scores.
Side-by-side comparison of different detector outputs.
Support for curated examples from datasets like RAGTruth, PsiloQA, and Mu-SHROOM.

Key Contributions

Unified Toolkit: SIRIN is the first open-source toolkit to unify probing, judge-style, and uncertainty estimation methods, alongside query answerability, under a shared interface supporting both white-box and black-box regimes at response and span granularity.
Interactive Inspection: A live web UI enabling the analysis of context-query-answer triples, visualization of unsupported spans, and direct comparison of detector performance.
Empirical Demonstration: The authors demonstrate SIRIN's utility in:
- Contextual hallucination detection and query answerability across multiple benchmarks.
- Integration as a faithfulness gate within the SimpleMem long-term memory system, acting as a selective prediction layer.

Experimental Results

Response-Level Detection

Evaluated on RAGTruth and SQuAD 2.0:

Probing: Lightweight probes achieved strong results (e.g., 88.0 AUROC on RAGTruth QA), outperforming fine-tuned encoder judges in data-scarce scenarios.
Judges: LoRA-fine-tuned decoder judges (e.g., Qwen3-4B) achieved the highest performance when sufficient annotations were available (93.5 AUROC on RAGTruth QA), outperforming probes by 5.5–9.0 AUROC points. Zero-shot GPT-5.4-mini performed competitively.
Uncertainty: Unsupervised uncertainty estimators (e.g., MeanTokenEntropy, RAUQ) provided weaker baselines (58.8–69.0 AUROC) but required no training data.
Answerability: The same pipeline successfully handled pre-generation answerability checks, with the attention-pooling probe reaching 87.9 AUROC on SQuAD 2.0.

Span-Level Detection

Evaluated on PsiloQA (English subset):

Probes provided high-quality ranking (up to 76.0 AUROC) and remained effective when span annotations were scarce.
Trained judges (LoRA-Qwen3-4B) achieved the best localization metrics (74.8 F1, 69.3 IoU) in annotation-rich settings, while zero-shot span tagging failed significantly.

Memory System Integration (SimpleMem)

SIRIN was integrated as a trust layer in the SimpleMem system on LongMemEval:

Mechanism: Two gates were applied: an answerability gate (pre-generation) to expand retrieval if the query is uncovered, and a faithfulness gate (post-generation) to validate spans against evidence.
Outcome: Combining both gates increased the accuracy of served answers from a baseline of 62.4% to 79.3%, while reducing strict unfaithfulness from 20.2% to 9.8%. This was achieved without extra generation costs, using selective prediction (serve, retry, or abstain).

Significance and Claims

The paper posits that contextual hallucination detection is an infrastructure problem rather than a single-detector challenge. SIRIN's significance lies in:

Standardization: Providing a common environment for developing, benchmarking, and deploying detectors across different paradigms.
Practicality: Demonstrating that lightweight probes are effective when labels are scarce, while trained judges dominate with sufficient data.
Systemic Improvement: Showing that a unified detection layer can effectively act as a "trust layer" for agentic memory systems, significantly improving reliability and reducing hallucinations without requiring changes to the underlying memory architecture.

The authors conclude that SIRIN represents a step toward making verification a standard layer in the LLM stack, where every grounded response is checked, unsupported spans are visible, and new detectors are easily integrated via plug-ins.

Limitations

The authors note that current detectors are trained on English data and have not been validated in multilingual settings. Performance depends on the quality and domain match of annotated data, potentially requiring re-calibration for out-of-domain deployment. Additionally, the memory gating results were evaluated with a single generator backbone, and the Web UI is currently limited by GPU memory constraints on the number of simultaneously loaded models.