💬 NLP

Reverse Probing: Supervised Token-level Uncertainty Quantification for Large Language Models in Clinical Text

This paper introduces Reverse Probing, a novel framework that leverages internal model activations to estimate token-level uncertainty in clinical text summarization without requiring new sampling, thereby outperforming existing baselines in accuracy and efficiency while providing interpretable insights into model confidence.

Original authors: Bushi Xiao, Sarvesh Soni, Daisy Zhe Wang

Published 2026-05-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bushi Xiao, Sarvesh Soni, Daisy Zhe Wang

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 a doctor reading a patient's summary written by a very smart, but sometimes overconfident, AI assistant. The AI writes a long report about the patient's hospital stay. Most of the time, the report is perfect. But occasionally, the AI might invent a symptom or get a number wrong—a "hallucination." In the real world, a single wrong word could be dangerous.

The problem is: How do we know which specific word the AI is unsure about?

Most current methods try to solve this by asking the AI to write the same report ten times and seeing if the stories change. If the stories are different, the AI is unsure. But this is like asking a tired doctor to write the same discharge summary ten times just to check their confidence. It takes too long, costs too much money, and isn't practical for long medical documents.

This paper introduces a new method called Reverse Probing. Here is how it works, using simple analogies:

1. The "Reverse" Idea: Reading the Mind, Not the Mouth

Instead of asking the AI to speak (generate new text) to see if it's confident, the researchers decided to feed the AI its own existing text and watch how its brain reacts.

Think of the AI's internal brain as a giant library of connections.

Normal Method: You ask the librarian, "Tell me a story about this patient." If the librarian stammers or tells different stories, you know they are unsure.
Reverse Probing: You hand the librarian a finished story and say, "Read this sentence back to me, but pretend you don't know the last word." Then, you watch the librarian's eyes (the internal signals).
- If the sentence is true (supported by the patient's real records), the librarian's eyes light up with a clear, confident path to the answer.
- If the sentence is fake (unsupported), the librarian's eyes look confused, or they look in the wrong direction. They can't find the "grounding" in the real records.

The researchers call this "Reverse" because they aren't looking at what the AI produces; they are looking at how the AI processes what already exists.

2. The "Probe": Checking the Pulse

The researchers use a technique called probing. Imagine the AI is a black box. You can't see inside, but you can poke it with a stick (a "probe") to see how it vibrates.

In this study, they take a clinical summary and a "Brief Hospital Course" (the raw, factual notes from the hospital). They feed both into the AI.

The Test: They hide one word at a time (like a "fill-in-the-blank") and ask the AI to guess it based on the context.
The Signal: They measure four types of "vibrations" inside the AI's brain:
1. Internal Signals: How much the AI's "thoughts" change from one layer of its brain to the next.
2. Uncertainty Signals: How "scattered" the AI's confidence is. Is it sure it's 99% right, or is it guessing between 50/50 options?
3. Medical Knowledge: Does the AI recognize this word as a medical term?
4. Distributional Signals: This is the big one. They compare how the AI reacts when it sees the real hospital notes versus when it doesn't. If the AI's confidence drops significantly when the real notes are removed, that word is likely unsupported.

3. The "Detective": Finding the Lies

Once they have these "vibrations" (data points) for every single word, they train a smart computer program (a classifier) to act as a detective.

The detective learns: "When the AI's brain shows this specific pattern of confusion, that word is probably a lie."
The result is a map of the document where every word is scored from 0 to 1. A high score means "This word is shaky; check it."

4. The Results: Fast and Accurate

The researchers tested this on real medical data where experts had already marked the fake parts.

Speed: Because they didn't have to ask the AI to write new stories, the process was incredibly fast. They could check 100 documents in 10 minutes. The old methods (asking the AI to write 10 times) took 7 hours for the same amount.
Accuracy: Their "Reverse Probing" method was 4 times better at finding the fake words than the next best method.
The Surprise: They found that smaller AI models (7 billion parameters) were actually better at showing their uncertainty than the giant 70-billion-parameter models. The giant models were so overconfident that they hid their confusion, making it harder to spot the lies.

Summary

Reverse Probing is like a lie detector test for AI text. Instead of making the AI talk more to see if it's nervous, you feed it a sentence and watch its internal "pulse." If its pulse skips a beat when it tries to connect a word to the real medical facts, you know that word is likely a hallucination. It's faster, cheaper, and much more precise at finding specific errors in long medical reports.

Technical Summary: Reverse Probing for Token-Level Uncertainty Quantification in Clinical Text

Problem Statement
The deployment of Large Language Models (LLMs) for clinical text summarization introduces critical safety concerns regarding reliability. While Uncertainty Quantification (UQ) is well-studied in general domains, its application to clinical text faces distinct challenges: clinical errors are often sparse (occurring at single words or numbers within otherwise correct sentences), and the cost of error is high. Existing UQ methods primarily operate at the sequence level or rely on generating multiple output candidates (sampling) to estimate uncertainty. These approaches are ill-suited for clinical settings because:

Granularity: They fail to localize uncertainty to specific tokens or spans within long documents.
Efficiency: Repeated sampling over long clinical documents is computationally expensive.
Data Scarcity: Expert-annotated clinical text is scarce, making it difficult to evaluate new generations reliably.
Architecture: Traditional token-level probing (e.g., masked language modeling) is native to bidirectional models like BERT but not to autoregressive decoder-only LLMs.

Furthermore, the paper distinguishes UQ from hallucination detection. UQ is a form of self-assessment measuring a model's confidence in its own predictions, independent of ground truth, whereas hallucination detection relies on external fact-checking.

Methodology: Reverse Probing
The authors propose Reverse Probing, a framework that estimates token-level uncertainty directly from pre-existing labeled summaries without generating new text. The core innovation is an "inversion" of the standard UQ paradigm: instead of asking the model to generate text and evaluating the output, the method feeds existing clinical text back into the frozen LLM to probe its internal state.

The methodology proceeds as follows:

Prompt-Based Masking: To bridge the architectural gap with decoder-only models, the system sequentially replaces each target token in a clinical summary with a mask placeholder. The model is then prompted to predict the masked token based on the surrounding context.
Contextual Perturbation (The "Reverse" Mechanism): The framework leverages the Brief Hospital Course (BHC) as the source of patient-specific evidence. It performs two forward passes for each token:
- With Context: The summary is fed with the BHC.
- Without Context: The summary is fed without the BHC.
  The core hypothesis is that supported tokens will align with the BHC in the model's representation space, while unsupported (hallucinated) tokens will lack this grounding. The absence of alignment serves as a latent uncertainty signal.
Feature Extraction: The system extracts four categories of features from the model's internal activations and output distributions:
- Internal Signals: Hidden state norms, layer-to-layer changes (L2 distance, cosine similarity), and attention metrics (entropy, weight directed to BHC, inter-layer drift).
- Uncertainty Signals: Output distribution shape metrics (Shannon entropy, margin, Gini coefficient, free energy).
- Medical Knowledge Signals: Domain-specific context via Named Entity Recognition (NER) using scispaCy/UMLS, medical entity density, and token frequency/IDF.
- Distributional Signals: The difference in probabilities and entropy between the "with BHC" and "without BHC" passes (e.g., $\Delta E$ , $\Delta H$ , KL divergence).
Supervised Classification: A supervised classifier (XGBoost or CatBoost) is trained on these features to predict a continuous uncertainty score for each token. The model is trained to identify internal activation patterns associated with unsupported spans.

Key Contributions

First Token-Level UQ Framework for Clinical Summaries: Reverse Probing is the first framework designed specifically to estimate token-level uncertainty in clinical summarization tasks.
Computational Efficiency: By eliminating multi-sampling and operating on pre-existing text, the method reduces inference time from hours (required by sampling-based baselines) to minutes.
New Benchmark: The authors establish the first benchmark for UQ in clinical summaries, evaluating against eight adapted baselines across six LLMs (including general and biomedical models) on two expert-annotated datasets (Hallucinations-MIMIC-DI and Hallucinations-Generated-DI).
Interpretable Insights: The study identifies specific internal signals that correlate with uncertainty, offering a path toward explainable AI in clinical settings.

Results
The evaluation demonstrates that Reverse Probing significantly outperforms all adapted baselines:

Performance: On the MIMIC-DI dataset (human-written summaries), Reverse Probing achieved an AUPRC of 0.5478 (using BioMistral-7B), compared to 0.1313 for the strongest baseline (SE Probes). This represents a roughly 4× improvement.
Robustness: The method maintains consistent superiority across models of varying sizes (7B to 70B) and pretraining conditions (general vs. biomedical), although smaller models (7B/8B) generally outperformed larger 70B models. The authors hypothesize that larger models may be overconfident, compressing the dynamic range of uncertainty signals.
Efficiency: Reverse Probing processes 100 samples in under 10 minutes on a single GPU, whereas sampling-based methods (e.g., Semantic Entropy) require approximately 7 hours for the same task.
Feature Analysis: Feature importance analysis reveals that delta energy (the change in free energy when BHC is present) and neighborhood context (sliding window statistics) are the most consistent predictors of uncertainty across all models. This suggests that uncertainty in clinical text is often a property of local spans rather than isolated tokens.

Significance and Claims
The paper claims that Reverse Probing offers a practical and scalable solution for uncertainty estimation in high-stakes clinical environments where generating new text for evaluation is infeasible. By treating existing text as a probe, the method transforms uncertainty estimation into an interpretable process that traces predictions back to specific internal signals (specifically, the lack of alignment with clinical evidence).

The authors modestly note limitations, including:

Model Coverage: The framework was primarily validated on Mistral and Llama architectures.
Large Model Adaptability: Performance degrades on 70B models, suggesting a need for adaptive feature extraction strategies for deeper networks.
Data Scarcity: The reliance on expert-annotated datasets limits the scale of evaluation.
Task Generalization: The study is limited to discharge summaries, and generalization to other clinical text types (e.g., radiology reports) remains unverified.

Ultimately, the work argues that models internally encode uncertainty in response to unsupported content, and Reverse Probing provides a mechanism to access these signals efficiently without the computational overhead of generative sampling.