analiza aiqbench
This research investigates whether recurrent computation in Transformers can mitigate the performance drop observed in multimodal BabyLM models compared to text-only baselines. The study finds that while recurrence improves efficiency, the primary cause of accuracy loss stems from changes in language data distribution and presentation during multimodal training rather than a lack of computational depth. These findings offer critical insights for optimizing compute-constrained pretraining strategies in next-generation foundation models.
Loop Before You Leap: What Recurrent Depth Can and Cannot Buy in Data-Limited Pretraining
arXiv:2611.06206v1 Announce Type: new Abstract: Multimodal models have often scored below text-only baselines in the BabyLM Challenge (Klerings et al., 2024; Takmaz et al., 2025). We test whether additional recurrent computation helps close this gap. Recurrence increases computation by reusing layers without adding parameters or training data. Using a single implementation, we compare standard and looped Transformers under a 100M-word budget, with three seeds per main condition. Each image is represented by one feature vector from a frozen DINOv2 encoder, used as a prefix token. Caption words count toward the budget, so multimodal training also changes the language data. In a pre-registered text-only comparison, a 12-layer stack repeated four times exceeds a standard 48-layer Transformer by 5.4 BLiMP percentage points at matched inference compute. However, multimodal training lowers BLiMP accuracy by about 3.8 points in both architectures, with no clear mitigation from recurrence. An exploratory loop-side no-image control, trained with the multimodal language data and batching schedule, scores 5.7 points below the text-only baseline. Adding the visual prefix raises accuracy by 1.85 points on average, but this contrast is uncertain with three seeds. These results suggest that changes in language data and its presentation are a major source of the drop. Diagnostic tests show that the models use visual information; its contribution to grammatical generalization remains uncertain under this single-vector interface.
Abstract:Multimodal models have often scored below text-only baselines in the BabyLM Challenge (Klerings et al., 2024; Takmaz et al., 2025). We test whether additional recurrent computation helps close this gap. Recurrence increases computation by reusing layers without adding parameters or training data. Using a single implementation, we compare standard and looped Transformers under a 100M-word budget, with three seeds per main condition. Each image is represented by one feature vector from a frozen DINOv2 encoder, used as a prefix token. Caption words count toward the budget, so multimodal training also changes the language data. In a pre-registered text-only comparison, a 12-layer stack repeated four times exceeds a standard 48-layer Transformer by 5.4 BLiMP percentage points at matched inference compute. However, multimodal training lowers BLiMP accuracy by about 3.8 points in both architectures, with no clear mitigation from recurrence. An exploratory loop-side no-image control, trained with the multimodal language data and batching schedule, scores 5.7 points below the text-only baseline. Adding the visual prefix raises accuracy by 1.85 points on average, but this contrast is uncertain with three seeds. These results suggest that changes in language data and its presentation are a major source of the drop. Diagnostic tests show that the models use visual information; its contribution to grammatical generalization remains uncertain under this single-vector interface.
Accepted to the BabyLM Challenge 2026 (Strict track). 20 pages, 5 figures, 12 tables. This version no longer reports a training-time shuffled-feature control that appeared in the workshop version, because its execution could not be verified (see Limitations)
Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv:2611.06206 [cs.CL]
(or arXiv:2611.06206v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2611.06206 arXiv-issued DOI via DataCite (pending registration)
Text integral sub licența deschisă a publicației · 276 cuvinte