Chain-of-Thought (CoT) reasoning has driven recent gains of large language models (LLMs) on reasoning-intensive tasks by externalizing intermediate steps.However, excessive or redundant reasoning -so-called overthinking -can increase inference costs and lead LLMs toward incorrect conclusions.In this paper, we present REFRAIN (REFlective-Redundancy for Adaptive INference), a training-free framework that adaptively determines when to stop reasoning to mitigate overthinking.REFRAIN integrates a two-stage stop discriminator to identify reflective yet redundant reasoning and a sliding-window Upper Confidence Bound (SW-UCB) multi-armed bandit controller to dynamically adjust stopping thresholds according to problem difficulty without supervision or fine-tuning.Across four representative benchmarks and two model families, REFRAIN reduces token usage by 20-55% while maintaining or improving accuracy compared to standard CoT prompting.Extensive ablation and robustness analyses demonstrate its stability across models, scorers, and prompt variations.In summary, our findings highlight when-tostop as a new and practical axis of test-time scaling -enabling models to reason not just more, but just enough.

Hengyuan Zhang, Chenming Shang, Sizhe Wang, Dongdong Zhang, Yiyao Yu, Feng Yao, Renliang Sun, Yujiu Yang, Furu Wei. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

We introduce MMMU: a new benchmark designed to evaluate multimodal models on massive multi-discipline tasks demanding college-level subject knowledge and deliberate reasoning. MMMU includes 11.5K meticulously collected multimodal questions from college exams, quizzes, and text-books, covering six core disciplines: Art & Design, Busi-ness, Science, Health & Medicine, Humanities & Social Science, and Tech & Engineering. These questions span 30 subjects and 183 subfields, comprising 30 highly het-erogeneous image types, such as charts, diagrams, maps, tables, music sheets, and chemical structures. Unlike existing benchmarks, MMMU focuses on advanced perception and reasoning with domain-specific knowledge, challenging models to perform tasks akin to those faced by experts. The evaluation of 28 open-source LMMs as well as the propri-etary GPT-4V(ision) and Gemini highlights the substantial challenges posed by MMMU. Even the advanced GPT-4V and Gemini Ultra only achieve accuracies of 56% and 59% respectively, indicating significant room for improvement. We believe MMMU will stimulate the community to build next-generation multimodal foundation models towards expert artificial general intelligence.

Sentence Simplification is a valuable technique that can benefit language learners and children a lot.However, current research focuses more on English sentence simplification.The development of Chinese sentence simplification is relatively slow due to the lack of data.To alleviate this limitation, this paper introduces CSS, a new dataset for assessing sentence simplification in Chinese.We collect manual simplifications from human annotators and perform data analysis to show the difference between English and Chinese sentence simplifications.Furthermore, we test several unsupervised and zero/few-shot learning methods on CSS and analyze the automatic evaluation and human evaluation results.In the end, we explore whether Large Language Models can serve as high-quality Chinese sentence simplification systems by evaluating them on CSS.

One of the major problems with text simplification is the lack of high-quality data.The sources of simplification datasets are limited to Wikipedia and Newsela, restricting further development of this field.In this paper, we analyzed the similarity between text summarization and text simplification and exploited summarization data to help simplify.First, we proposed an alignment algorithm to extract sentence pairs from summarization datasets.Then, we designed four attributes to characterize the degree of simplification and proposed a method to filter suitable pairs.We named these pairs Sum4Simp (S4S).Next, we conducted human evaluations to show that S4S is high-quality and compared it with a real simplification dataset.Finally, we conducted experiments to illustrate that the S4S can improve the performance of several mainstream simplification models, especially in low-resource scenarios.

Predicting therapeutic responses in cancer patients is a major challenge in the field of precision medicine due to high inter- and intra-tumor heterogeneity. Most drug response models need to be improved in terms of accuracy, and there is limited research to assess therapeutic responses of particular tumor types. Here, we developed a novel method DROEG (Drug Response based on Omics and Essential Genes) for prediction of drug response in tumor cell lines by integrating genomic, transcriptomic and methylomic data along with CRISPR essential genes, and revealed that the incorporation of tumor proliferation essential genes can improve drug sensitivity prediction. Concisely, DROEG integrates literature-based and statistics-based methods to select features and uses Support Vector Regression for model construction. We demonstrate that DROEG outperforms most state-of-the-art algorithms by both qualitative (prediction accuracy for drug-sensitive/resistant) and quantitative (Pearson correlation coefficient between the predicted and actual IC50) evaluation in Genomics of Drug Sensitivity in Cancer and Cancer Cell Line Encyclopedia datasets. In addition, DROEG is further applied to the pan-gastrointestinal tumor with high prevalence and mortality as a case study at both cell line and clinical levels to evaluate the model efficacy and discover potential prognostic biomarkers in Cisplatin and Epirubicin treatment. Interestingly, the CRISPR essential gene information is found to be the most important contributor to enhance the accuracy of the DROEG model. To our knowledge, this is the first study to integrate essential genes with multi-omics data to improve cancer drug response prediction and provide insights into personalized precision treatment.

k-nearest-neighbor machine translation (kNN-MT), proposed by Khandelwal et al. (2021), has achieved many state-of-the-art results in machine translation tasks.Although effective, kNN-MT requires conducting kNN searches through the large datastore for each decoding step during inference, prohibitively increasing the decoding cost and thus leading to the difficulty for the deployment in real-world applications.In this paper, we propose to move the time-consuming kNN search forward to the preprocessing phase, and then introduce k Nearest Neighbor Knowledge Distillation (kNN-KD) that trains the base NMT model to directly learn the knowledge of kNN.Distilling knowledge retrieved by kNN can encourage the NMT model to take more reasonable target tokens into consideration, thus addressing the overcorrection problem.Extensive experimental results show that, the proposed method achieves consistent improvement over the stateof-the-art baselines including kNN-MT, while maintaining the same training and decoding speed as the standard NMT model. 1

Text simplification is a valuable technique.However, current research is limited to sentence simplification.In this paper, we define and investigate a new task of document-level text simplification, which aims to simplify a document consisting of multiple sentences.Based on Wikipedia dumps, we first construct a large-scale dataset named D-Wikipedia and perform analysis and human evaluation on it to show that the dataset is reliable.Then, we propose a new automatic evaluation metric called D-SARI that is more suitable for the document-level simplification task.Finally, we select several representative models as baseline models for this task and perform automatic evaluation and human evaluation.We analyze the results and point out the shortcomings of the baseline models.

Most of the research on text simplification is limited to sentence level nowadays.In this paper, we are the first to investigate the helpfulness of document context on sentence simplification and apply it to the sequence-to-sequence model.We firstly construct a sentence simplification dataset in which the contexts for the original sentence are provided by Wikipedia corpus.The new dataset contains approximately 116K sentence pairs with context.We then propose a new model that makes full use of the context information.Our model uses neural networks to learn the different effects of the preceding sentences and the following sentences on the current sentence and applies them to the improved transformer model.Evaluated on the newly constructed dataset, our model achieves 36.52 on SARI value, which outperforms the best performing model in the baselines by 2.46 (7.22%), indicating that context indeed helps improve sentence simplification.In the ablation experiment, we show that using either the preceding sentences or the following sentences as context can significantly improve simplification.

BACKGROUND: Variant calling and refinement from whole genome/exome sequencing data is a fundamental task for genomics studies. Due to the limited accuracy of NGS sequencing and variant callers, IGV-based manual review is required for further false positive variant filtering, which costs massive labor and time, and results in high inter- and intra-lab variability. RESULTS: To overcome the limitation of manual review, we developed a novel approach for Variant Filter by Automated Scoring based on Tagged-signature (VariFAST), and also provided a pipeline integrating GATK Best Practices with VariFAST, which can be easily used for high quality variants detection from raw data. Using the bam and vcf files, VariFAST calculates a v-score by sum of weighted metrics causing false positive variations, and marks tags in the manner of keeping high consistency with manual review, for each variant. We validated the performance of VariFAST for germline variant filtering using the benchmark sequencing data from GIAB, and also for somatic variant filtering using sequencing data of both malignant carcinoma and benign adenomas as well. VariFAST also includes a predictive model trained by XGBOOST algorithm for germline variants refinement, which reveals better MCC and AUC than the state-of-the-art VQSR, especially outcompete in INDEL variant filtering. CONCLUSION: VariFAST can assist researchers efficiently and conveniently to filter the false positive variants, including both germline and somatic ones, in NGS data analysis. The VariFAST source code and the pipeline integrating with GATK Best Practices are available at https://github.com/bioxsjtu/VariFAST.

The ultimate goal of metabolic engineering is to produce desired compounds on an industrial scale in a cost effective manner. To address challenges in metabolic engineering, computational strain optimization algorithms based on genome-scale metabolic models have increasingly been used to aid in overproducing products of interest. However, most of these strain optimization algorithms utilize a metabolic network alone, with few approaches providing strategies that also include transcriptional regulation. Moreover previous integrated approaches generally require a pre-existing regulatory network. In this study, we developed a novel strain design algorithm, named OptRAM (Optimization of Regulatory And Metabolic Networks), which can identify combinatorial optimization strategies including overexpression, knockdown or knockout of both metabolic genes and transcription factors. OptRAM is based on our previous IDREAM integrated network framework, which makes it able to deduce a regulatory network from data. OptRAM uses simulated annealing with a novel objective function, which can ensure a favorable coupling between desired chemical and cell growth. The other advance we propose is a systematic evaluation metric of multiple solutions, by considering the essential genes, flux variation, and engineering manipulation cost. We applied OptRAM to generate strain designs for succinate, 2,3-butanediol, and ethanol overproduction in yeast, which predicted high minimum predicted target production rate compared with other methods and previous literature values. Moreover, most of the genes and TFs proposed to be altered by OptRAM in these scenarios have been validated by modification of the exact genes or the target genes regulated by the TFs, for overproduction of these desired compounds by in vivo experiments cataloged in the LASER database. Particularly, we successfully validated the predicted strain optimization strategy for ethanol production by fermentation experiment. In conclusion, OptRAM can provide a useful approach that leverages an integrated transcriptional regulatory network and metabolic network to guide metabolic engineering applications.