Determine whether the generation limit causes lower sample accuracy for Gemma-4-12B

Determine whether reaching the generation limit causes the lower sample accuracy observed for Gemma-4-12B on the Composite task.

Background

On the Composite task, Gemma-4-12B has lower sample accuracy when using native reasoning than when answering without reasoning. The model also reaches the prescribed generation limit relatively often, with 42.0% of its thinking-mode runs hitting that limit. Because an unfinished reasoning trace may prevent the model from producing a final answer, the reported results do not establish whether the generation limit itself is responsible for the observed reduction in sample accuracy. The paper subsequently studies when an answer becomes usable during reasoning and proposes early stopping, but it does not directly identify whether the generation limit caused Gemma-4-12B’s lower sample accuracy.

References

This model also often reaches the generation limit; these results do not show whether that limit causes its lower sample accuracy.

— From Perception to Integration: Revisiting the Internal Dynamics of Reasoning in Vision-Language Models  (2609.34809 - Xu et al., 28 Sep 2026) in Section 3, subsection “Development of Composite Answer Readiness,” paragraph “Reasoning improves counterfactual consistency”