• Computer Science > Machine Learning [Submitted on 28 Jan 2026] Title:Reducing Text Bias in Synthetically Generated MCQAs for VLMs in Autonomous Driving View PDF HTML (experimental)Abstract:Multiple Choice Question Answering (MCQA) benchmarks are an established standard for measuring Vision Language Model (VLM) performance in driving tasks. • However, we observe the known phenomenon that synthetically generated MCQAs are highly susceptible to hidden textual cues that allow models to exploit linguistic patterns rather than visual context. • Our results show that a VLM fine-tuned on such data can achieve accuracy comparable to human-validated benchmarks even without visual input. • Our proposed method reduces blind accuracy from +66.9% above random to +2.9%, eliminating the vast majority of exploitable textual shortcuts. • By decoupling the correct answer from linguistic artifacts and employing a curriculum learning strategy, we force the model to rely on visual grounding, ensuring that performance accurately reflects perceptual understanding. • Submission history From: Christoffer Heckman [view email][v1] Wed, 28 Jan 2026 20:30:26 UTC (1,685 KB) Current browse context: cs.LG References & Citations export BibTeX citation Loading…
Article Summaries:
- Researchers have shown that synthetic multiple‑choice question answering (MCQA) datasets used to evaluate vision‑language models (VLMs) for autonomous driving contain hidden textual cues that let models achieve high accuracy without visual input. In a recent study, a VLM fine‑tuned on such data reached human‑level scores even when the image was omitted, with “blind” accuracy 66.9 % above random. The team introduced a method that decouples correct answers from linguistic artifacts and applies curriculum learning, reducing blind accuracy to just 2.9 % above random. This forces the model to rely on visual grounding, ensuring performance better reflects perceptual understanding.
Sources: