Verify the mechanism behind deterministic Qwen2-VL-2B descriptions

Determine whether 4-bit quantization narrows the Qwen2-VL-2B output probability distribution sufficiently to cause nominally sampled descriptions to collapse to argmax selection under the stated deployment configuration.

Background

The paper reports that Qwen2-VL-2B produces byte-identical descriptions across four independent runs on the same 360 images, despite using sampling with temperature 0.7 and without fixing a random seed. The authors attribute this apparent determinism to 4-bit quantization narrowing the output probability distribution, but they explicitly state that they did not directly verify this causal mechanism. Establishing whether quantization is responsible would clarify the reproducibility behavior of the deployed vision-LLM and distinguish it from other possible causes of deterministic generation.

References

We attribute this to quantization narrowing the output probability distribution sufficiently that sampling collapses to argmax selection in practice, though we did not verify this mechanism directly.

— GoDeep: Annotation-Free Open-Vocabulary 3D Scene Understanding via Language-Space Lifting  (2609.09082 - Betsas et al., 8 Sep 2026) in Section 4, Ablation Studies, paragraph “VLM Description Consistency” (Paragraph \ref{para:vlmconsistency})