Self-Ensemble: Mitigating Confidence Distortion for Large Language Models

Zicheng Xu, Guanchu Wang, Guangyao Zheng, Yu-Neng Chuang, Alex Szalay, Xia Hu, Vladimir BravermanEMNLP 2025

TL;DR

Large Language Models (LLMs) struggle with confidence distortion in multi-choice question answering (MCQA), especially when there are many options. This paper introduces "Self-Ensemble," a novel, plug-and-play method that divides MCQA problems into smaller sub-problems and ensembles their predictions. Self-Ensemble significantly improves LLM accuracy and reliability on MCQA tasks without requiring additional training data.

What Problem Does This Paper Address?

Large Language Models (LLMs) have shown remarkable capabilities across various Natural Language Processing (NLP) tasks. However, when faced with multi-choice question answering (MCQA), particularly those with a large number of answer choices (many-choice QA), LLMs exhibit a phenomenon called "confidence distortion." This means they tend to be under-confident in their correct predictions and over-confident in their incorrect ones. This distortion leads to a significant degradation in performance, making LLMs unreliable for such tasks. Existing methods to address this often require labeled validation data for fine-tuning, which can be costly and labor-intensive to obtain. The paper aims to solve this confidence distortion problem in a way that is efficient, effective, and doesn't require extra labeled data.

Methods

The core innovation is Self-Ensemble. The method works by taking a K-choice MCQA problem and breaking it down into multiple smaller, m-choice (where m << K) sub-problems. This is achieved by randomly grouping the answer choices into several subsets. The LLM then makes predictions on each of these smaller sub-problems. Finally, the predictions from these sub-problems are ensembled to arrive at a final decision.

A key aspect of Self-Ensemble is its plug-and-play nature. To integrate this into existing LLM architectures without requiring separate inference passes for each sub-problem, the authors designed specialized attention masks and positional re-encoding mechanisms. These allow the LLM to process the question and all the choice groups in a single forward pass, making the process efficient. The attention mask ensures that the model focuses only on the relevant group of choices during each part of the ensemble process, while positional re-encoding maintains the correct contextual relationships.

Datasets

The paper evaluates Self-Ensemble on the following datasets:

  • QASC (Question Answering via Sentence Composition): A multi-hop reasoning dataset, specifically used in its 8-choice QA format.
  • TruthfulQA: A benchmark designed to measure how well models mimic human falsehoods, filtered to include questions with at least one correct choice and five incorrect choices.
  • MMLU-Pro (Massive Multitask Language Understanding - Professional): An enhanced version of MMLU, focusing on questions where state-of-the-art LLMs often fail. The Biology subset was used.

Key Insights and Ideas

  • Confidence Distortion is Pervasive: The paper clearly demonstrates that confidence distortion is a significant issue for LLMs, especially as the number of answer choices increases. This isn't just a minor glitch; it directly impacts performance.
  • Divide and Conquer for Confidence: The core idea of breaking down a complex problem (many choices) into simpler ones (few choices) and then combining the results is a powerful strategy for mitigating confidence issues.
  • Plug-and-Play Efficiency: The clever use of attention masks and positional re-encoding to enable a single forward pass for the ensemble process is crucial. This makes Self-Ensemble practical for real-world deployment without significant computational overhead compared to standard inference.
  • Model Agnosticism: Self-Ensemble works across different LLM families (LLaMA, Mistral, Qwen2) and scales, suggesting it's a generalizable solution rather than being tied to a specific model architecture.

Key Results

  • Accuracy Improvement: Self-Ensemble consistently outperforms standard inference and baseline methods. On average, it improves accuracy by 8% across the tested LLMs and datasets. For example, on the QASC dataset, Self-Ensemble achieved up to a 7.45% accuracy improvement for LLaMA-3-8B.
  • Mitigation of Distortion: Figure 6 visually shows that Self-Ensemble leads to a higher proportion of confidence exceeding a threshold for correct answers and a lower proportion for incorrect answers, effectively addressing the confidence distortion.
  • Efficiency: The single-pass implementation of Self-Ensemble significantly reduces inference time (e.g., from 0.78s to 0.16s per question for LLaMA-3-8B on QASC) with minimal impact on memory usage.
  • Quantized LLM Performance: Self-Ensemble can offset the accuracy loss incurred by aggressive quantization (e.g., 4-bit) in LLMs, allowing quantized models to outperform their full-precision counterparts.

Limitations and Future Work

  • Limited to MCQA: The current application of Self-Ensemble is strictly for multi-choice question-answering tasks. The authors suggest exploring its application to open-ended QA tasks as a future research direction.
  • Orthogonal to Existing Methods: While Self-Ensemble is designed to be orthogonal to existing calibration techniques (like Vector Scaling or Dirichlet Calibration), enabling synergistic integration, its full potential in combination with these methods is an area for further exploration.
  • Hyperparameter Sensitivity: While the paper shows robustness to random grouping seeds, the optimal group size ($m$) and number of trials ($N$) might still require some consideration, though accuracy is relatively insensitive to group size. The authors note that increasing trials stabilizes performance, indicating robustness.

Summary updated

BibTeX
@inproceedings{xu2025self,
  author = {Zicheng Xu and Guanchu Wang and Guangyao Zheng and Yu-Neng Chuang and Alex Szalay and Xia Hu and Vladimir Braverman},
  title = {{Self-Ensemble: Mitigating Confidence Distortion for Large Language Models}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure