TLDR: A new research paper evaluates different methods for Large Language Models (LLMs) to quantify their uncertainty, specifically within an argumentative framework (ArgLLMs) used for claim verification. The study found that a straightforward approach called “direct prompting,” where the LLM directly states its confidence, significantly outperformed more complex methods like Semantic Entropy, Eccentricity, and LUQ across various datasets and LLMs. This suggests that for complex, long-form argumentative tasks, direct prompting is a highly effective and resource-efficient way for LLMs to convey their certainty, making AI systems more reliable.
Large Language Models (LLMs) have become incredibly powerful, capable of tasks from coding to reasoning. However, they sometimes produce incorrect or ‘hallucinated’ answers without indicating their level of certainty. This lack of clarity can be risky, especially in critical fields like healthcare, where users might blindly trust AI responses. This is where Uncertainty Quantification (UQ) comes in, aiming to make AI systems more trustworthy by allowing them to express how confident they are in their outputs.
A recent study delves into how different UQ methods perform within a specific type of LLM framework called Argumentative LLMs (ArgLLMs). ArgLLMs are designed to be more explainable, making decisions based on computational argumentation. In this framework, UQ plays a crucial role because the confidence in the arguments directly influences the final decision, such as verifying a claim.
What are Argumentative LLMs (ArgLLMs)?
ArgLLMs work by generating a network of supporting and attacking arguments for a given statement. Each argument is assigned a ‘base score’ representing its strength, which is essentially the confidence score provided by a UQ method. These arguments and their relationships are then processed using a formal reasoning system called gradual semantics. This system calculates a final strength for the original claim. If the claim’s final score is above a certain threshold (e.g., 0.5), it’s predicted as true; otherwise, it’s false. The confidence scores from the UQ method are central to this entire process, making ArgLLMs an ideal setting to evaluate how effective different UQ methods are.
Exploring Different Uncertainty Quantification Methods
The researchers integrated several LLM UQ methods into ArgLLMs and compared their performance on claim verification tasks. The methods evaluated were:
-
Direct Prompting: This is the simplest method, where the LLM is directly asked to provide a confidence score for the text it generates.
-
Semantic Entropy: This method involves the LLM generating multiple different samples for an input. Samples with similar meanings are then clustered, and an entropy score is calculated based on the distribution of these meanings to determine uncertainty.
-
Eccentricity: This approach also generates multiple samples. It then uses a natural language inference (NLI) model to determine how much one generation entails another, creating similarity scores. These scores are used to build a graph, and uncertainty is computed as the average distance from the center of the graph’s eigenvectors.
-
LUQ (Long-text Uncertainty Quantification): Similar to Eccentricity, LUQ generates multiple responses and uses an NLI model to get ‘entailment’ and ‘contradiction’ values between them. Uncertainty is then calculated as a function of these logits, with higher similarity indicating lower uncertainty.
The Experiments and Key Findings
The study conducted extensive experiments across three claim verification datasets (TruthfulClaim, StrategyClaim, and MedClaim), three different LLMs (Google’s Gemma-2-9b-it, Meta’s Llama-3.1-8B, and OpenAI’s GPT-4o-mini), and various ArgLLM settings (different depths of argumentation and ways to determine the claim’s base score). This resulted in 36 unique configurations.
The results were striking: despite its simplicity, direct prompting consistently emerged as the most effective UQ strategy in ArgLLMs. It outperformed the considerably more complex approaches in terms of accuracy in claim verification. Direct prompting was either the best method or not statistically significantly worse than the best in all 36 configurations, being the outright best in 25 instances.
LUQ also showed relatively strong performance, sometimes being the best in specific setups, which aligns with its known strengths in factuality tasks. However, Semantic Entropy and Eccentricity generally lagged behind.
The researchers suggest that direct prompting’s superior performance, especially with long-form and potentially contentious arguments, might be due to its ability to leverage the LLM’s internal knowledge directly. Sampling-based methods like Semantic Entropy can struggle to capture semantic consistency over long and complex texts. Additionally, direct prompting avoids the need for extra normalization steps, which can introduce noise and further estimation challenges.
Another significant advantage of direct prompting is its lower resource requirement. The other high-performing UQ methods often demand a separate NLI model and multiple generations per instance, increasing both memory and computational time.
Also Read:
- Efficiently Verifying AI’s Step-by-Step Thinking with NCV
- How Language Models Learn to Balance Internal Knowledge with New Information
Conclusion
This research highlights the value of evaluating LLM UQ methods in argumentative and explainable frameworks like ArgLLMs. The findings strongly suggest that for tasks involving complex, long-form, and potentially contentious statements, simply asking the LLM for its confidence through direct prompting is a highly effective and efficient method for uncertainty quantification. This reinforces the idea that verbalized confidence scores can be well-calibrated and accurately reflect an LLM’s internal knowledge when prompted effectively.
For more details, you can read the full research paper here: Evaluating Uncertainty Quantification Methods in Argumentative Large Language Models.
