ELITR-Bench: A Meeting Assistant Benchmark for Long-Context Language Models

TL;DR

ELITR-Bench introduces a realistic long-context benchmark for meeting assistants by augmenting the ELITR corpus with 271 ground-truth QA pairs and ASR-noisy transcript variants. The study benchmarks 12 long-context LLMs, revealing clear gaps between state-of-the-art GPT-4 family models and open models, especially under multi-turn and noisy conditions. A GPT-4-based evaluator is validated against human judgments and Prometheus, showing strong correlations with humans but limited fine-grained discrimination on a 10-point scale. The work highlights robustness gaps to transcript noise and suggests future directions such as retrieval-augmented evaluation and de-identification impacts to better reflect real-world deployment.

Abstract

Research on Large Language Models (LLMs) has recently witnessed an increasing interest in extending the models' context size to better capture dependencies within long documents. While benchmarks have been proposed to assess long-range abilities, existing efforts primarily considered generic tasks that are not necessarily aligned with real-world applications. In contrast, we propose a new benchmark for long-context LLMs focused on a practical meeting assistant scenario in which the long contexts consist of transcripts obtained by automatic speech recognition, presenting unique challenges for LLMs due to the inherent noisiness and oral nature of such data. Our benchmark, ELITR-Bench, augments the existing ELITR corpus by adding 271 manually crafted questions with their ground-truth answers, as well as noisy versions of meeting transcripts altered to target different Word Error Rate levels. Our experiments with 12 long-context LLMs on ELITR-Bench confirm the progress made across successive generations of both proprietary and open models, and point out their discrepancies in terms of robustness to transcript noise. We also provide a thorough analysis of our GPT-4-based evaluation, including insights from a crowdsourcing study. Our findings indicate that while GPT-4's scores align with human judges, its ability to distinguish beyond three score levels may be limited.

Figures (9)

• Figure 2: Comparison of the scores obtained for GPT-4o, LLaMA-3.1-8B, and Phi-3-small using transcripts with varied levels of noise on the test set of ELITR-Bench-QA in single-turn mode. Indicated levels of noise correspond to the target Word Error Rates set in our noise injection procedure.
• Figure 3: (a) Distribution of GPT-4 and Silver Human scores with respect to each Gold Human score bin (1-10); the N below a score bin indicates the bin size. (b) Pearson correlation between evaluators.
• Figure 4: Results restricted to QA/Conv differentiating questions. The score reported for each model and evaluation setting corresponds to the average of the scores obtained on the dev subset (16 questions) and the test subset (17 questions). Best viewed in color.
• Figure 5: Interface for our Prolific crowdsourcing study to collect Silver Human score annotations.
• Figure 6: Answer prompt used to obtain LLMs' responses. Questions are appended to this prompt as described in Section \ref{['sec:exp-setup']}. The element in blue and enclosed in curly brackets corresponds to a meeting-specific text span that is dynamically adapted.