Reinforcement Learning with Verifiable Rewards (RLVR) offers a promising framework for optimizing large language models in reasoning tasks. However, existing RLVR algorithms focus on different granularities, and each has complementary strengths and limitations. Group Relative Policy Optimization (GRPO) updates the policy with token-level importance ratios, which preserves fine-grained credit assignment but often suffers from high variance and instability. In contrast, Group Sequence Policy Optimization (GSPO) applies single sequence-level importance ratios across all tokens in a response that better matches sequence-level rewards, but sacrifices token-wise credit assignment. In this paper, we propose Dynamic Hybrid Policy Optimization (DHPO) to bridge GRPO and GSPO within a single clipped surrogate objective. DHPO combines token-level and sequence-level importance ratios using weighting mechanisms. We explore two variants of the mixing mechanism, including an averaged mixing and an entropy-guided mixing. To further stabilize training, we employ a branch-specific clipping strategy that constrains token-level and sequence-level ratios within separate trust regions before mixing, preventing outliers in either branch from dominating the update. Across seven challenging mathematical reasoning benchmarks, experiments on both dense and MoE models from the Qwen3 series show that DHPO consistently outperforms GRPO and GSPO. Our code is publicly available at https://github.com/XMUDeepLIT/DHPO.

Reinforcement learning with verifiable rewards (RLVR) improves mathematical reasoning, but sparse sequence-level rewards make training heavily dependent on long, repeated rollout generation. On-policy distillation (OPD) makes these rollouts more informative by asking a larger teacher to score every student trajectory token by token, but this inserts large-teacher forward passes into every RL step and requires online resources beyond standard student-only RL. We propose SPEC-Distill, a speculative prefix distillation framework in which the teacher shapes rollout construction rather than defining an auxiliary reward. Instead of querying the teacher during training, SPEC-Distill runs it once offline to build a reusable pool of expert trajectories. During RL, the student treats these trajectories as speculative drafts, accepts a prefix under a lenience-controlled criterion, completes the response under its own policy, and is updated only by verifiable rewards with standard GRPO. Against OPD, this removes online teacher scoring; against standard autoregressive rollouts, accepted prefixes reduce the number of fresh student-generated tokens. In the primary non-think setting, SPEC-Distill improves Qwen3-4B from 66.1 to 73.2 average accuracy over GRPO, and outperforms OPD on the two scales. Compared with full-response expert reuse such as Luffy, SPEC-Distill reaches comparable accuracy on Qwen3-1.7B while reducing online rollout tokens by about 60% and rollout time from 10.90 to 7.06 hours. Analysis further shows that teacher prefixes are best used as an early scaffold, after which reusing the student’s own rollouts provides the strongest balance between expert guidance and on-policy exploration.

Long audio-video reasoning is difficult for omni-modal LLMs because the decisive evidence is often sparse, cross-modal, and too expensive to preserve with uniformly high-fidelity inputs. We introduce OmniReasoner, a tool-use post-training framework for Thinking with Long Audio-Video: omni-modal LLMs learn, via supervised fine-tuning and reinforcement learning, to decide whether and where to call a zoom-in tool before answering. OmniReasoner first builds a low-cost global preview of the full stream and then, when needed, calls the zoom-in tool with a requested temporal interval for higher-fidelity visual and audio inspection before answering. Because the model observes different sampling granularities before and after this call—a sparse global preview and a denser local clip—we introduce TimeAnchor, which keeps the tool’s temporal argument valid and round-trip-consistent across these granularities, rather than tied to frame indices from a particular sampling rate. To make this tool-use behavior trainable without expensive manual interval annotation, we build a Temporal Augmented Data Engine that synthesizes tool-use post-training trajectories by video editing and composition. Experiments across omni-modal and video benchmarks show that OmniReasoner improves both answer accuracy and temporal grounding while concentrating high-fidelity computation on informative regions. Code is available at https://github.com/RockyChen0205/OmniReasoner.

As LLM-based shopping agents enter production, existing benchmarks fail to capture how a shopper’s requirements arrive: stated implicitly in the query, recorded in a profile, or revealed only when the right question is asked. Benchmarks that expose full intent upfront and grade only the final choice can neither pose this long-horizon challenge nor explain which requirement an agent missed. To address this gap, we introduce EComAgentBench, a benchmark of 662 tasks grounded in real Amazon products and reviews. Each task scatters these requirements across a visible query, a tool-gated profile, and scripted clarification; an agent must uncover hidden intent, verify candidates against attributes and review evidence, and commit to a single product within 100 tool calls. Moreover, typed, source-tagged rubrics grade every task, attributing each failure to a requirement and its source. Construction is automated yet reliable, with every answer fixed in code before any text is generated and every sample validated. Our evaluation of seven models reveals that even the strongest attains only 57.1% overall accuracy, and rubric satisfaction degrades from visible to hidden sources. Overall, we believe EComAgentBench will serve as a reproducible foundation for moving shopping agents from single-query search toward dependable assistance over long horizons.

The goal of open relation extraction (OpenRE) is to develop an RE model that can generalize to new relations not encountered during training. Existing studies primarily formulate OpenRE as a clustering task. They first cluster all test instances based on the similarity between the instances, and then manually assign a new relation to each cluster. However, their reliance on human annotation limits their practicality. In this paper, we propose an OpenRE framework based on large language models (LLMs), which directly predicts new relations for test instances by leveraging their strong language understanding and generation abilities, without human intervention. Specifically, our framework consists of two core components: (1) a relation discoverer (RD), designed to predict new relations for test instances based on demonstrations formed by training instances with known relations; and (2) a relation predictor (RP), used to select the most likely relation for a test instance from n candidate relations, guided by demonstrations composed of their instances. To enhance the ability of our framework to predict new relations, we design a self-correcting inference strategy composed of three stages: relation discovery, relation denoising, and relation prediction. In the first stage, we use RD to preliminarily predict new relations for all test instances. Next, we apply RP to select some high-reliability test instances for each new relation from the prediction results of RD through a cross-validation method. During the third stage, we employ RP to re-predict the relations of all test instances based on the demonstrations constructed from these reliable test instances. Extensive experiments on three OpenRE datasets demonstrate the effectiveness of our framework. We release our code at https://github.com/XMUDeepLIT/LLM-OREF.git.

Direct Preference Optimization (DPO) has emerged as a promising framework for aligning Large Language Models (LLMs) with human preferences by directly optimizing the log-likelihood difference between chosen and rejected responses. However, existing methods assign equal importance to all tokens in the response, while humans focus on more meaningful parts. This leads to suboptimal preference optimization, as irrelevant or noisy tokens disproportionately influence DPO loss. To address this limitation, we propose Optimal Transport-based token weighting scheme for enhancing direct Preference Optimization (OTPO). By emphasizing semantically meaningful token pairs and de-emphasizing less relevant ones, our method introduces a context-aware token weighting scheme that yields a more contrastive reward difference estimate. This adaptive weighting enhances reward stability, improves interpretability, and ensures that preference optimization focuses on meaningful differences between responses. Extensive experiments have validated OTPO’s effectiveness in improving instruction-following ability across various settings.

Diverse NMT aims at generating multiple diverse yet faithful translations given a source sentence. In this paper, we investigate a common shortcoming in existing diverse NMT studies: the model is usually trained with single reference, while expected to generate multiple candidate translations in inference. The discrepancy between training and inference enlarges the confidence variance and quality gap among candidate translations and thus hinders model performance. To deal with this defect, we propose a multi-candidate optimization framework for diverse NMT. Specifically, we define assessments to score the diversity and the quality of candidate translations during training, and optimize the diverse NMT model with two strategies based on reinforcement learning, namely hard constrained training and soft constrained training. We conduct experiments on NIST Chinese-English and WMT14 English-German translation tasks. The results illustrate that our framework is transparent to basic diverse NMT models, and universally makes better trade-off between diversity and quality. Our source codeis available at https://github.com/DeepLearnXMU/MultiCanOptim.

Controllable paraphrase generation (CPG) incorporates various external conditions to obtain desirable paraphrases. However, existing works only highlight a special condition under two indispensable aspects of CPG (i.e., lexically and syntactically CPG) individually, lacking a unified circumstance to explore and analyze their effectiveness. In this paper, we propose a general controllable paraphrase generation framework (GCPG), which represents both lexical and syntactical conditions as text sequences and uniformly processes them in an encoder-decoder paradigm. Under GCPG, we reconstruct commonly adopted lexical condition (i.e., Keywords) and syntactical conditions (i.e., Part-Of-Speech sequence, Constituent Tree, Masked Template and Sentential Exemplar) and study the combination of the two types. In particular, for Sentential Exemplar condition, we propose a novel exemplar construction method — Syntax-Similarity based Exemplar (SSE). SSE retrieves a syntactically similar but lexically different sentence as the exemplar for each target sentence, avoiding exemplar-side words copying problem. Extensive experiments demonstrate that GCPG with SSE achieves state-of-the-art performance on two popular benchmarks. In addition, the combination of lexical and syntactical conditions shows the significant controllable ability of paraphrase generation, and these empirical results could provide novel insight to user-oriented paraphrasing.

Attention mechanism has become the dominant module in natural language processing models. It is computationally intensive and depends on massive power-hungry multiplications. In this paper, we rethink variants of attention mechanism from the energy consumption aspects. After reaching the conclusion that the energy costs of several energy-friendly operations are far less than their multiplication counterparts, we build a novel attention model by replacing multiplications with either selective operations or additions. Empirical results on three machine translation tasks demonstrate that the proposed model, against the vanilla one, achieves competitable accuracy while saving 99% and 66% energy during alignment calculation and the whole attention procedure. Our code will be released upon the acceptance.

Recognizing the language of ambiguous texts has become a main challenge in language identification (LID). When using multilingual applications, users have their own language preferences, which can be regarded as external knowledge for LID. Nevertheless, current studies do not consider the inter-personal variations due to the lack of user annotated training data. To fill this gap, we introduce preference-aware LID and propose a novel unsupervised learning strategy. Concretely, we construct pseudo training set for each user by extracting training samples from a standard LID corpus according to his/her historical language distribution. Besides, we contribute the first user labeled LID test set called “U-LID”. Experimental results reveal that our model can incarnate user traits and significantly outperforms existing LID systems on handling ambiguous texts. Our code and benchmark have been released.

Query language identification (Q-LID) plays a crucial role in a cross-lingual search engine. There exist two main challenges in Q-LID: (1) insufficient contextual information in queries for disambiguation; and (2) the lack of query-style training examples for low-resource languages. In this article, we propose a neural Q-LID model by alleviating the above problems from both model architecture and data augmentation perspectives. Concretely, we build our model upon the advanced Transformer model. In order to enhance the discrimination of queries, a variety of external features (e.g., character, word, as well as script) are fed into the model and fused by a multi-scale attention mechanism. Moreover, to remedy the low resource challenge in this task, a novel machine translation–based strategy is proposed to automatically generate synthetic query-style data for low-resource languages. We contribute the first Q-LID test set called QID-21, which consists of search queries in 21 languages. Experimental results reveal that our model yields better classification accuracy than strong baselines and existing LID systems on both query and traditional LID tasks.1

Short texts (STs) present in a variety of scenarios, including query, dialog, and entity names. Most of the exciting studies in neural machine translation (NMT) are focused on tackling open problems concerning long sentences rather than short ones. The intuition behind is that, with respect to human learning and processing, short sequences are generally regarded as easy examples. In this article, we first dispel this speculation via conducting preliminary experiments, showing that the conventional state-of-the-art NMT approach, namely, Transformer (Vaswani et al. 2017), still suffers from over-translation and mistranslation errors over STs. After empirically investigating the rationale behind this, we summarize two challenges in NMT for STs associated with translation error types above, respectively: (1) the imbalanced length distribution in training set intensifies model inference calibration over STs, leading to more over-translation cases on STs; and (2) the lack of contextual information forces NMT to have higher data uncertainty on short sentences, and thus NMT model is troubled by considerable mistranslation errors. Some existing approaches, like balancing data distribution for training (e.g., data upsampling) and complementing contextual information (e.g., introducing translation memory) can alleviate the translation issues in NMT for STs. We encourage researchers to investigate other challenges in NMT for STs, thus reducing ST translation errors and enhancing translation quality.

Translation quality evaluation plays a crucial role in machine translation. According to the input format, it is mainly separated into three tasks, i.e., reference-only, source-only and source-reference-combined. Recent methods, despite their promising results, are specifically designed and optimized on one of them. This limits the convenience of these methods, and overlooks the commonalities among tasks. In this paper, we propose , which is the first unified framework engaged with abilities to handle all three evaluation tasks. Concretely, we propose monotonic regional attention to control the interaction among input segments, and unified pretraining to better adapt multi-task training. We testify our framework on WMT 2019 Metrics and WMT 2020 Quality Estimation benchmarks. Extensive analyses show that our single model can universally surpass various state-of-the-art or winner methods across tasks.Both source code and associated models are available at https://github.com/NLP2CT/UniTE.

In this paper, we present our submission to Shared Metrics Task: RoBLEURT (Robustly Optimizing the training of BLEURT). After investigating the recent advances of trainable metrics, we conclude several aspects of vital importance to obtain a well-performed metric model by: 1) jointly leveraging the advantages of source-included model and reference-only model, 2) continuously pre-training the model with massive synthetic data pairs, and 3) fine-tuning the model with data denoising strategy. Experimental results show that our model reaching state-of-the-art correlations with the WMT2020 human annotations upon 8 out of 10 to-English language pairs.

Document-level neural machine translation (NMT) has proven to be of profound value for its effectiveness on capturing contextual information. Nevertheless, existing approaches 1) simply introduce the representations of context sentences without explicitly characterizing the inter-sentence reasoning process; and 2) feed ground-truth target contexts as extra inputs at the training time, thus facing the problem of exposure bias. We approach these problems with an inspiration from human behavior – human translators ordinarily emerge a translation draft in their mind and progressively revise it according to the reasoning in discourse. To this end, we propose a novel Multi-Hop Transformer (MHT) which offers NMT abilities to explicitly model the human-like draft-editing and reasoning process. Specifically, our model serves the sentence-level translation as a draft and properly refines its representations by attending to multiple antecedent sentences iteratively. Experiments on four widely used document translation tasks demonstrate that our method can significantly improve document-level translation performance and can tackle discourse phenomena, such as coreference error and the problem of polysemy.

A well-known limitation in pretrain-finetune paradigm lies in its inflexibility caused by the one-size-fits-all vocabulary. This potentially weakens the effect when applying pretrained models into natural language generation (NLG) tasks, especially for the subword distributions between upstream and downstream tasks with significant discrepancy. Towards approaching this problem, we extend the vanilla pretrain-finetune pipeline with an extra embedding transfer step. Specifically, a plug-and-play embedding generator is introduced to produce the representation of any input token, according to pre-trained embeddings of its morphologically similar ones. Thus, embeddings of mismatch tokens in downstream tasks can also be efficiently initialized. We conduct experiments on a variety of NLG tasks under the pretrain-finetune fashion. Experimental results and extensive analyses show that the proposed strategy offers us opportunities to feel free to transfer the vocabulary, leading to more efficient and better performed downstream NLG models.

A good translation should not only translate the original content semantically, but also incarnate personal traits of the original text. For a real-world neural machine translation (NMT) system, these user traits (e.g., topic preference, stylistic characteristics and expression habits) can be preserved in user behavior (e.g., historical inputs). However, current NMT systems marginally consider the user behavior due to: 1) the difficulty of modeling user portraits in zero-shot scenarios, and 2) the lack of user-behavior annotated parallel dataset. To fill this gap, we introduce a novel framework called user-driven NMT. Specifically, a cache-based module and a user-driven contrastive learning method are proposed to offer NMT the ability to capture potential user traits from their historical inputs under a zero-shot learning fashion. Furthermore, we contribute the first Chinese-English parallel corpus annotated with user behavior called UDT-Corpus. Experimental results confirm that the proposed user-driven NMT can generate user-specific translations.

Recent studies have proven that the training of neural machine translation (NMT) can be facilitated by mimicking the learning process of humans. Nevertheless, achievements of such kind of curriculum learning rely on the quality of artificial schedule drawn up with the handcrafted features, e.g. sentence length or word rarity. We ameliorate this procedure with a more flexible manner by proposing self-paced learning, where NMT model is allowed to 1) automatically quantify the learning confidence over training examples; and 2) flexibly govern its learning via regulating the loss in each iteration step. Experimental results over multiple translation tasks demonstrate that the proposed model yields better performance than strong baselines and those models trained with human-designed curricula on both translation quality and convergence speed.

Query translation (QT) serves as a critical factor in successful cross-lingual information retrieval (CLIR). Due to the lack of parallel query samples, neural-based QT models are usually optimized with synthetic data which are derived from large-scale monolingual queries. Nevertheless, such kind of pseudo corpus is mostly produced by a general-domain translation model, making it be insufficient to guide the learning of QT model. In this paper, we extend the data augmentation with a domain transfer procedure, thus to revise synthetic candidates to search-aware examples. Specifically, the domain transfer model is built upon advanced Transformer, in which layer coordination and mixed attention are exploited to speed up the refining process and leverage parameters from a pre-trained cross-lingual language model. In order to examine the effectiveness of the proposed method, we collected French-to-English and Spanish-to-English QT test sets, each of which consists of 10,000 parallel query pairs with careful manual-checking. Qualitative and quantitative analyses reveal that our model significantly outperforms strong baselines and the related domain transfer methods on both translation quality and retrieval accuracy.

This paper describes the submission systems of Alibaba for WMT18 shared news translation task. We participated in 5 translation directions including English ↔ Russian, English ↔ Turkish in both directions and English → Chinese. Our systems are based on Google’s Transformer model architecture, into which we integrated the most recent features from the academic research. We also employed most techniques that have been proven effective during the past WMT years, such as BPE, back translation, data selection, model ensembling and reranking, at industrial scale. For some morphologically-rich languages, we also incorporated linguistic knowledge into our neural network. For the translation tasks in which we have participated, our resulting systems achieved the best case sensitive BLEU score in all 5 directions. Notably, our English → Russian system outperformed the second reranked system by 5 BLEU score.