Document-grounded quantitative reasoning requires a system to first understand a visually presented document and then compute an exact answer from the grounded content. We describe LILIPUT, a zero-shot submission to the DocSem 2026 shared task, in which Gemini 3.8 Flash performs visual evidence localization and converts the selected content into a faithful, self-contained math problem, while DeepSeek V4 performs the downstream numerical reasoning. On 1,730 held-out test instances, our accepted submission obtains 49.54% joint exact accuracy, 55.09% answer accuracy, and 66.76% macro evidence F1. In contrast, experiments on 908 labeled training instances show much higher answer accuracy when the relevant evidence is successfully extracted: Gemini 3.7 Flash achieves 95.59%, while text-based solvers achieve 92.51%–97.03%. Although these results are not directly comparable, the large gap suggests that calculation is not the main challenge. Instead, the main difficulty is correctly understanding and extracting information from the document. We found that clean and clear documents were easier to process, while blurry, dense, and visually complex documents made extraction much harder, even after image enhancement. Since DeepSeek V4 uses the extracted content for calculation, errors during visual extraction can also lead to incorrect answers. These findings highlight visual document understanding as a key challenge for document-grounded quantitative reasoning.

When training data is very limited, the way different training signals are weighted over time can strongly affect what a model learns. We study hybrid replaced-token detection (RTD) and masked language modeling (MLM) pretraining under the BabyLM 2026 Strict-Small budget. We train five single-seed DeBERTa-v2 models along three curriculum axes: objective scheduling, masking policy, and data order. The primary system, bgb-decay, retains generator MLM, discriminator RTD, and auxiliary discriminator MLM throughout training while reducing the RTD coefficient in stages. We compare it with a matched fixed-RTD reference, a variant that adds easy-to-hard data ordering and adaptive auxiliary masking, and two stronger RTD-to-MLM handoff schedules with whole-word and adaptive masking. Among the five runs, bgb-decay achieves the best overall balance, while the more aggressive variants gain more on some tasks but drop more on others. Training diagnostics reveal a trade-off between replacement discrimination and auxiliary vocabulary prediction: under fixed high RTD weighting, RTD gradients are larger and nearly orthogonal to auxiliary-MLM gradients, while staged decay preserves RTD but shifts late optimization toward MLM. These results suggest that gradual objective rebalancing is a promising direction for small-data pretraining.