Neural machine translation (NMT) has become the predominant approach for automated translation, yet conventional models trained on extensive bilingual datasets exhibit critical limitations, including quadratic scaling of training data, sensitivity to out-of-distribution inputs, and a lack of interpretability. Inspired by the classical "translation pyramid" concept, which advocates for translation via a semantic pivot (interlingua), this work explores the integration of Abstract Meaning Representation (AMR) as a structured semantic intermediary to decouple translation into comprehension (source-to-AMR) and generation (AMR-to-target) phases. We conduct a pilot study using a strong AMR parser to create a multilingual silver-standard AMR corpus from the United Nations Parallel Corpus, training modular semantic understanding and generation components for each language. Experimental results demonstrate that our approach achieves an average improvement of 3% in robustness and over 15% in generalization compared to traditional Seq2Seq baselines. Analysis suggests that enhancing semantic parsing and generation accuracy could bridge the gap to conventional NMT systems. To our knowledge, this is the first work to integrate AMR as a semantic pivot in NMT, offering enhanced transparency, scalability, and robustness. This study underscores the potential of semantic-driven translation frameworks and provides a foundation for future research in interpretable, resource-efficient multilingual systems.

Parallel scaling has emerged as a powerful paradigm to enhance reasoning capabilities in large language models (LLMs) by generating multiple Chain-of-Thought (CoT) traces simultaneously. However, this approach introduces significant computational inefficiency due to *inter-trace redundancy*—our analysis reveals that over 80% of parallel reasoning traces yield identical final answers, representing substantial wasted computation. To address this critical efficiency bottleneck, we propose **DeepPrune**, a novel framework that enables efficient parallel scaling through dynamic pruning. Our method features a specialized judge model trained with oversampling techniques to accurately predict answer equivalence from partial reasoning traces, achieving 0.7072 AUROC on equivalence prediction across unseen reasoning models. This is combined with an online greedy clustering algorithm that dynamically prunes redundant paths while preserving answer diversity. Comprehensive evaluations across three challenging benchmarks (AIME 2024, AIME 2025, and GPQA) and multiple reasoning models demonstrate that DeepPrune achieves remarkable token reduction ranging from 65.73% to 88.50% compared to conventional consensus sampling, while maintaining competitive accuracy within 3.4 percentage points. Our work establishes a new standard for efficient parallel reasoning, making high-performance reasoning more efficient. Our code and data are here: https://github.com/THU-KEG/DeepPrune/