# Beyond Semantic Accuracy: Consequence-Aware Evaluation for Safety-Critical Language Understanding

arXiv: https://arxiv.org/abs/2608.24621 · HTML: https://mlsanity.com/paper/2608.24621/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 19 of its 64 references that are in the map.
- First posted on arXiv 25 Aug 2026.

## Builds on (most cited first, 19 of 19)

- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [Beyond Accuracy: Behavioral Testing of NLP models with CheckList](https://mlsanity.com/paper/2005.04118/) (2020)
- [R-Judge: Benchmarking Safety Risk Awareness for LLM Agents](https://mlsanity.com/paper/2401.10019/) (2024)
- [SafetyPrompts: a Systematic Review of Open Datasets for Evaluating and Improving Large Language Model Safety](https://mlsanity.com/paper/2404.05399/) (2024)
- [SALM: Speech-augmented Language Model with In-context Learning for Speech Recognition and Translation](https://mlsanity.com/paper/2310.09424/) (2023)
- [ASR Error Correction and Domain Adaptation Using Machine Translation](https://mlsanity.com/paper/2003.07692/) (2020)
- [Look Before You Leap: An Exploratory Study of Uncertainty Measurement for Large Language Models](https://mlsanity.com/paper/2307.10236/) (2023)
- [A Novel Evaluation Benchmark for Medical LLMs: Illuminating Safety and Effectiveness in Clinical Domains](https://mlsanity.com/paper/2507.23486/) (2025)
- [MediTOD: An English Dialogue Dataset for Medical History Taking with Comprehensive Annotations](https://mlsanity.com/paper/2410.14204/) (2024)
- [ATCO2 corpus: A Large-Scale Dataset for Research on Automatic Speech Recognition and Natural Language Understanding of Air Traffic Control Communications](https://mlsanity.com/paper/2211.04054/) (2022)
- [Not All Errors are Equal: Learning Text Generation Metrics using Stratified Error Synthesis](https://mlsanity.com/paper/2210.05035/) (2022)
- [Automatic Speech Recognition Benchmark for Air-Traffic Communications](https://mlsanity.com/paper/2006.10304/) (2020)
- [ATCSpeech: a multilingual pilot-controller speech corpus from real Air Traffic Control environment](https://mlsanity.com/paper/1911.11365/) (2019)
- [ECLM: Entity Level Language Model for Spoken Language Understanding with Chain of Intent](https://mlsanity.com/paper/2403.04481/) (2024)
- [Read the Scene, Not the Script: Outcome-Aware Safety for LLMs](https://mlsanity.com/paper/2510.04320/) (2025)
- [CIF-PT: Bridging Speech and Text Representations for Spoken Language Understanding via Continuous Integrate-and-Fire Pre-Training](https://mlsanity.com/paper/2305.17499/) (2023)
- [Adapting Automatic Speech Recognition for Accented Air Traffic Control Communications](https://mlsanity.com/paper/2502.20311/) (2025)
- [PilotBench: A Benchmark for General Aviation Agents with Safety Constraints](https://mlsanity.com/paper/2604.08987/) (2026)
- [World2Rules: A Neuro-Symbolic Framework for Learning World-Governing Safety Rules for Aviation](https://mlsanity.com/paper/2603.28952/) (2026)

## Abstract (arXiv)

Can language models be trusted in safety- critical operations? In such settings, strong per- formance on semantic metrics does not guaran- tee operational reliability: a misread altitude, a dropped execution condition, or a confused call- sign may score well under standard F1 yet carry sharply asymmetric operational consequences. We study this problem in air traffic control (ATC), where controller-pilot communication demands near-zero error tolerance, and use consequence-aware evaluation to test whether semantic scores misstate operational reliabil- ity. The framework is instantiated in a con- trolled diagnostic ATC benchmark grounded in aviation standards and feedback from 40 air traffic controllers across three countries. Evaluating 8 models, we uncover a system- atic semantic-safety gap: conventional scores give substantially higher performance estimates than consequence-aware evaluation, even for models that appear reliable under standard met- rics. Risk-aware fine-tuning narrows but does not close this gap, showing that consequence- aware evaluation is a necessary complement to standard NLP metrics before any real safety- critical deployment claim