Chain-of-thought prompts break VLM answer decoding in standard benchmarks
Researchers have identified a critical failure mode in vision-language model evaluation: when chain-of-thought prompts precede answer extraction, models collapse to near-random performance despite retaining correct answers in their hidden representations. On ScienceQA, Qwen2.5-VL-7B plummets from 81% to 45% accuracy, with 94% of predictions defaulting to the first option. The underlying cause is a probability redistribution toward continuation tokens rather than answer tokens, though linear probes confirm the information persists in late layers. This finding exposes a fundamental mismatch between how VLMs encode reasoning instructions and how evaluators decode answers, with implications for benchmark reliability and the design of multi-step reasoning workflows in production systems.
Modelwire context
ExplainerThe critical insight here isn't just that chain-of-thought prompts hurt performance on ScienceQA, it's that the model's reasoning capability and answer knowledge remain intact in its internal representations. The failure is purely in how the model's output probabilities get reallocated during decoding, not in what it actually knows.
This connects directly to the evaluation methodology fragility documented in recent work on LLM graders and likelihood ranking. Just as a single instruction phrase ('never give partial credit') flipped LLM evaluation systems from production-ready to unusable, and likelihood-based ranking diverged sharply from prompted answering across model scales, this paper shows that how we structure prompts and extract answers fundamentally reshapes measured performance independent of actual model capability. The pattern is consistent: evaluation protocol design matters more than model scale or training quality. These three findings together suggest the field has been conflating benchmark design artifacts with genuine capability differences.
Test whether this probability redistribution effect replicates on other VLM benchmarks (MMVP, MMBench) and whether it persists across different chain-of-thought formulations. If the collapse is specific to ScienceQA's format or to Qwen2.5-VL-7B, the finding is a quirk; if it generalizes, it signals a systematic vulnerability in how VLMs handle multi-step reasoning workflows that will require architectural or decoding-level fixes before these models can reliably handle production reasoning tasks.
Coverage we drew on
- Likelihood Ranking doesn't Scale Like Prompting in LLMs · arXiv cs.CL
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsQwen2.5-VL-7B · ScienceQA · Vision-language models · Chain-of-thought
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Reasoning Instructions Can Break Answer Decoding in Vision--Language Models”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.