Vision language models are predominantly evaluated on English-centric multimodal benchmarks, leaving their behavior in other languages and cultures insufficiently understood. We introduce VMMU, a Vietnamese Multitask Multimodal Understanding and Reasoning benchmark of 2,548 multiple-choice questions across seven domains: Mathematics, Physics, Chemistry, Biology, Geography, Driving Test, and IQ Test. Every question requires jointly reasoning over Vietnamese text and non-text visual evidence (charts, diagrams, tables, traffic scenes), and is distributed as a single rendered page image. We evaluate 10 open-source and 5 closed-source VLMs on VMMU. Test-time compute improves performance from 57% to 78% accuracy, but it remains inferior to an expert baseline (99% accuracy). Extensive error analyses show that closed-source VLMs already achieve strong Vietnamese OCR performance, yet still struggle on VMMU. This suggests that the primary bottleneck is multimodal grounding and reasoning rather than text recognition or Vietnamese language understanding. Code and data are available at: https://vmmu-bench.github.io.

Large language models often repeatedly select the same answer even when many alternatives are plausible. Prior work treats this concentration as bias, but it does not distinguish stable model preferences from responses that depend on a particular prompt wording. We introduce a bias depth score that measures both how strongly a model prefers its top answer under direct prompting and whether that answer survives scenario reframing. Across 4,442 opinion prompts and four large language models, only about a quarter of the concentrated preferences survive reframing. We call these persistent cases Deep biases, and the remaining prompt-dependent cases Shallow biases. Our results show that Deep biases are more often inherited from pretraining and preserved through SFT. Under both continued fine-tuning and prompt-based debiasing for diversity, Deep biases are consistently harder to remove than Shallow biases. Bias depth therefore separates stable learned biases from prompt-wording artifacts that single-prompt metrics conflate. Code, models, and data are available at https://deepbias.github.io.

Efficient context compression is critical for retrieval-augmented question answering in resource-constrained settings, where long retrieved contexts increase latency, memory use, and LLM reader cost. We propose a lightweight encoder-only framework for query-driven sentence pruning that preserves answer-critical evidence while aggressively reducing irrelevant context. Our method learns marginal contribution scores for sentences using counterfactual training signals and optimizes a contrastive ranking objective that separates critical evidence from noncritical context. Our approach scores all sentences from a single full-context encoding, enabling fast inference with low computational overhead. Experiments show that it maintains accuracy comparable to the strongest baseline while using 3.7× less peak memory and achieving nearly 3× lower compression latency, demonstrating an