# Right Answers, Inconsistent Evidence: Source-Level Counterfactual Auditing of Large Audio-Language Models

**Authors:** Yi Zhang, Yi Li, Hongwei Du  
**Published:** 11 Nov 2026  
**Links:** [arXiv](https://arxiv.org/abs/2611.15463) · [PDF](https://arxiv.org/pdf/2611.15463) · [Code](https://github.com/yikise/TRACE) · [This page](https://omen-bit.github.io/ai-paper-daily/paper/2611.15463.html)

## Key numbers

- **6.3%**: 6.3% of correct answers fail this ordering
- **42.5%**: A target-only score check misses 42.5% of them
- **4.9% → 8.4%**: The share rises from 4.9% to 8.4% as target SNR falls from +5 to -5 dB

## Abstract

A large audio-language model (LALM) can answer "is a cat present?" about a mixture correctly, and its confidence can drop when the cat is removed, yet this does not show that it can tell the cat's presence from its absence. We introduce TRACE. It deletes the queried source and each rival in turn and asks whether every view still containing that sound outscores the view with it removed. Across 8 LALMs and 3,456 controlled mixtures, 6.3% of correct answers fail this ordering: the target-absent view scores at or above some target-present view, a conflict no shared threshold repairs. A target-only score check misses 42.5% of them. The share rises from 4.9% to 8.4% as target SNR falls from +5 to -5 dB, even on units correct at every level, and survives loudness and re-pairing controls. Accuracy certifies the answer; TRACE audits whether it orders presence above absence. Code and data: https://github.com/yikise/TRACE.

---

Notes written by AI from the paper, numbers copied from the abstract. Check the paper before you rely on them. From AI Papers Daily.