Self-consistency (SC) improves chain-of-thought reasoning by aggregating multiple sampled answers, yet all existing variants operate in the output space—they refine answer selection but leave the noisy input context unchanged. We observe that when multiple reasoning traces independently cite the same support items, those items are likely key evidence for the question. Building on this input-side consensus signal, we propose SVote-2S, a two-stage self-consistency framework for multi-hop question answering. In Stage 1, SVote-2S performs confidence-weighted voting over support items cited across N sampled traces and selects a compact evidence set. In Stage 2, the model re-reasons over a filtered context containing only the selected evidence paragraphs (about 15% of the original length on average on our benchmarks), enabling more focused reasoning. The final answer is determined by confidence-weighted fusion of samples from both stages. The fusion serves as a robustness safeguard, while the primary gain comes from evidence voting and re-reasoning. Across three benchmarks (HotpotQA, MuSiQue, 2WikiMultihopQA) and three LLMs, SVote-2S improves over SC by 5.36 EM on average and consistently outperforms the strongest output-side baseline (CISC) across all nine settings, without additional training or external retrieval.

This paper describes the Kingsoft AI Lab’s submission to the WMT2019 news translation shared task. We participated in two language directions: English-Chinese and Chinese-English. For both language directions, we trained several variants of Transformer models using the provided parallel data enlarged with a large quantity of back-translated monolingual data. The best translation result was obtained with ensemble and reranking techniques. According to automatic metrics (BLEU) our Chinese-English system reached the second highest score, and our English-Chinese system reached the second highest score for this subtask.