This paper introduces Light-R1, an opensource suite for training long reasoning modelsusing reproducible and cost-effective methodology. Given the proprietary nature of data usedin the DeepSeek-R1 series, we develop an alternative approach leveraging exclusively publicdata and models. Our curriculum training progressively increases data difficulty, combinedwith multi-staged post-training. Our LightR1-32B model, trained from Qwen2.5-32BInstruct, outperforms DeepSeek-R1-DistillQwen-32B in math reasoning. Experimental results show that this curriculum approachbecomes more effective when distinct, diverse datasets are available for different training stages: fine-tuning DeepSeek-R1-Distilledmodels (pre-tuned by DeepSeek team on proprietary data) with 3,000 challenging examplesfrom our curriculum dataset yielded state-ofthe-art 7B and 14B models, while the 32Bmodel, Light-R1-32B-DS performed comparably to QwQ-32B and DeepSeek-R1. Furthermore, we extend our work by applying GRPOon long reasoning models. Our final Light-R1-14B-DS achieves SOTA performance among14B models in math, with AIME24 & 25 scoresof 74.0 and 60.2 respectively, surpassing many32B models and DeepSeek-R1-Distill-Llama70B. Despite math-focused training, Light-R1-14B-DS demonstrates strong cross-domain generalization. Light-R1 represents a significantadvancement in making sophisticated reasoning models more accessible and implementablein real-world applications. Our models, training data and code have been made available.

Answer selection task requires finding appropriate answers to questions from informative but crowdsourced candidates. A key factor impeding its solution by current answer selection approaches is the redundancy and lengthiness issues of crowdsourced answers. Recently, Deng et al. (2020) constructed a new dataset, WikiHowQA, which contains a corresponding reference summary for each original lengthy answer. And their experiments show that leveraging the answer summaries helps to attend the essential information in original lengthy answers and improve the answer selection performance under certain circumstances. However, when given a question and a set of long candidate answers, human beings could effortlessly identify the correct answer without the aid of additional answer summaries since the original answers contain all the information volume that answer summaries contain. In addition, pretrained language models have been shown superior or comparable to human beings on many natural language processing tasks. Motivated by those, we design a series of neural models, either pretraining-based or non-pretraining-based, to check wether the additional answer summaries are helpful for ranking the relevancy degrees of question-answer pairs on WikiHowQA dataset. Extensive automated experiments and hand analysis show that the additional answer summaries are not useful for achieving the best performance.

Multi-document reading comprehension task requires collecting evidences from different documents for answering questions. Previous research works either use the extractive modeling method to naively integrate the scores from different documents on the encoder side or use the generative modeling method to collect the clues from different documents on the decoder side individually. However, any single modeling method cannot make full of the advantages of both. In this work, we propose a novel method that tries to employ a multi-view fusion and multi-decoding mechanism to achieve it. For one thing, our approach leverages question-centered fusion mechanism and cross-attention mechanism to gather fine-grained fusion of evidence clues from different documents in the encoder and decoder concurrently. For another, our method simultaneously employs both the extractive decoding approach and the generative decoding method to effectively guide the training process. Compared with existing methods, our method can perform both extractive decoding and generative decoding independently and optionally. Our experiments on two mainstream multi-document reading comprehension datasets (Natural Questions and TriviaQA) demonstrate that our method can provide consistent improvements over previous state-of-the-art methods.

The prosperity of Pretrained Language Models(PLM) has greatly promoted the development of Machine Reading Comprehension (MRC). However, these models are vulnerable and not robust to adversarial examples. In this paper, we propose Stable and Contrastive Question Answering (SCQA) to improve invariance of representation to alleviate these robustness issues. Specifically, we first construct positive example pairs which have same answer through data augmentation. Then SCQA learns enhanced representations with better alignment between positive pairs by introducing stability and contrastive loss. Experimental results show that our approach can boost the robustness of QA models cross different MRC tasks and attack sets significantly and consistently.