Constructing scientific multimodal document reasoning datasets for foundation model training involves an inherent trade-off among scale, faithfulness, and realism. To address this challenge, we introduce the synthesize-and-reground framework, a two-stage pipeline comprising: (1) Claim-Centric QA Synthesis, which generates faithful, isolated QA pairs and reasoning on focused segments, and (2) Document-Scale Regrounding, which programmatically re-embeds these pairs into full-document tasks to ensure realistic complexity. We present SciMDR, a large-scale training dataset for cross-modal comprehension, comprising 300K QA pairs with explicit reasoning chains across 20K scientific papers. We further construct SciMDR-Eval, an expert-annotated benchmark to evaluate multimodal comprehension within full-length scientific workflows. Experiments demonstrate that models fine-tuned on SciMDR achieve significant improvements across multiple scientific QA benchmarks, particularly in tasks requiring complex document-level reasoning.

Static bias benchmarks score a model on isolated questions, yet LLMs that answer safely at first can still reveal bias as a conversation unfolds. We frame bias evaluation as multi-turn conversational elicitation and learn an interviewer that probes a target LLM, conditioned on a demographic group, a bias category, and a conversational strategy. The interviewer is trained with supervised fine-tuning on synthetic dialogues and refined with preference optimization on conversation-level feedback. Across all models we test, bias grows as the conversation continues, exposing failures that single-turn evaluation misses. Training the interviewer matters: it surfaces substantially more bias than the same model prompted with identical instructions, and preference optimization pushes it further. Multi-turn elicitation and single-turn search turn out to be complementary: at matched query budgets they surface similar amounts of bias, but conversation uncovers more severe cases. Human annotations corroborate these judgments. Our results argue that bias evaluation of conversational AI should move beyond single-turn probes.