Multilingual retrieval-augmented generation (MRAG) requires models to effectively acquire and integrate beneficial external knowledge from multilingual collections. However, most existing studies employ a unitive process where queries of equivalent semantics across different languages are processed through a single-turn retrieval and subsequent optimization. Such a “one-size-fits-all” strategy is often suboptimal in multilingual settings, as the models occur to knowledge bias and conflict during the interaction with the search engine. To alleviate the issues, we propose LcRL, a multilingual search-augmented reinforcement learning framework that integrates a language-coupled Group Relative Policy Optimization into the policy and reward models. We adopt the language-coupled group sampling in the rollout module to reduce knowledge bias, and regularize an auxiliary anti-consistency penalty in the reward models to mitigate the knowledge conflict. Experimental results demonstrate that  not only achieves competitive performance but is also appropriate for various practical scenarios such as constrained training data and retrieval over collections encompassing a large number of languages. Our code is available at https://anonymous.4open.science/r/LcRL-B4EF.

Current large reasoning models (LRMs) prefer to generate verbose chain-of-thought (CoT) traces (a.k.a. the “overthinking” issue), especially in multilingual scenarios due to the imbalanced proficiency across different languages. However, existing studies on CoT compression focus exclusively on English and overlook language characteristics, resulting in performance degradation in multilingual scenarios. To address this issue, we develop a multilingual CoT compression method based on reinforcement learning, named Language-aware Entropy-based (LAE), which optimizes LLMs with multilingual group sampling and entropy gap-based rewards. Specifically, we extend GRPO with multilingual group sampling, where each query is jointly optimized across multiple languages. We further design a language-aware entropy-gap reward that penalizes redundant uncertainty in the thinking phase and aligns reasoning complexity across languages, enabling concise yet accurate multilingual CoT reasoning. Extensive experiments on the multilingual reasoning benchmarks MGSM, MMATH and MMMLU demonstrate that our method achieves competitive performance and generalization, compared with several strong baselines.

Knowledge distillation (KD) is widely used to compress and post-train large language models (LLMs), yet many existing frameworks execute teacher inference with the same training-oriented backend as student optimization, leading to suboptimal efficiency. In this paper, we propose KDFlow, a novel framework for LLM distillation that features a decoupled architecture and employs SGLang for teacher inference. KDFlow combines SGLang for teacher inference with PyTorch FSDP2 for student optimization, allowing each model to run on a backend tailored to its workload. To enable efficient full-vocabulary distillation in this decoupled architecture, KDFlow transfers the teacher’s final hidden states via Ray’s object store and recomputes teacher logits on each student worker using a frozen copy of the teacher’s output head. Furthermore, our framework supports both off-policy and on-policy distillation and incorporates cross-tokenizer algorithms through highly extensible and user-friendly APIs. Experiments show that KDFlow achieves a 1.44× to 6.36× speedup over MS-SWIFT in off-policy distillation and a 1.43× to 1.75× speedup over verl in on-policy distillation. KDFlow further scales to 64 GPUs, achieving 3.68× and 2.52× strong-scaling speedups in two representative model configurations. The code and documentation are publicly available. The code and documentation are publicly available.

Large Reasoning Models (LRMs) have achieved remarkable performance on complex reasoning tasks by adopting the “think-then-answer” paradigm, which enhances both accuracy and interpretability. However, current LRMs exhibit two critical limitations when processing non-English languages: (1) They often struggle to maintain input-output language consistency; (2) They generally perform poorly with wrong reasoning paths and lower answer accuracy compared to English. These limitations significantly compromise the interpretability of reasoning processes and degrade the user experience for non-English speakers, hindering the global deployment of LRMs. To address these limitations, we propose M-Thinker, which is trained by the GRPO algorithm that involves a Language Consistency (LC) reward and a novel Cross-lingual Thinking Alignment (CTA) reward. Specifically, the LC reward defines a strict constraint on the language consistency between the input, thought, and answer. Besides, the CTA reward compares the model’s non-English reasoning paths with its English reasoning path to transfer its own reasoning capability from English to non-English languages. Through an iterative RL procedure, our M-Thinker-1.5B/4B/7B models not only achieve nearly 100% language consistency and superior performance on two multilingual benchmarks (MMATH and PolyMath), but also exhibit excellent generalization on out-of-domain languages.

Current large language models (LLMs) generally show a significant performance gap in alignment between English and other languages.To bridge this gap, existing research typically leverages the model’s responses in English as a reference to select the best/worst responses in other languages, which are then used for Direct Preference Optimization (DPO) training.However, we argue that there are two limitations in the current methods that result in noisy multilingual preference data and further limited alignment performance: 1) Not all English responses are of high quality, and using a response with low quality may mislead the alignment for other languages. 2) Current methods usually use biased or heuristic approaches to construct multilingual preference pairs.To address these limitations, we design a consistency-based data selection method to construct high-quality multilingual preference data for improving multilingual alignment (CM-Align).Specifically, our method includes two parts: consistency-guided English reference selection and cross-lingual consistency-based multilingual preference data construction.Experimental results on three LLMs and three common tasks demonstrate the effectiveness and superiority of our method, which further indicates the necessity of constructing high-quality preference data.

In modern large language models (LLMs), LLM alignment is of crucial importance and is typically achieved through methods such as reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO). However, in most existing methods for LLM alignment, all tokens in the response are optimized using a sparse, response-level reward or preference annotation. The ignorance of token-level rewards may erroneously punish high-quality tokens or encourage low-quality tokens, resulting in suboptimal performance and slow convergence speed. To address this issue, we propose AlignDistil, a RLHF-equivalent distillation method for token-level reward optimization. Specifically, we introduce the reward learned by DPO into the RLHF objective and theoretically prove the equivalence between this objective and a token-level distillation process, where the teacher distribution linearly combines the logits from the DPO model and a reference model. On this basis, we further bridge the accuracy gap between the reward from the DPO model and the pure reward model, by building a contrastive DPO reward with a normal and a reverse DPO model. Moreover, to avoid under- and over-optimization on different tokens, we design a token adaptive logit extrapolation mechanism to construct an appropriate teacher distribution for each token. Experimental results demonstrate the superiority of our AlignDistil over existing methods and showcase fast convergence due to its token-level distributional reward optimization.

Continually expanding new languages for existing large language models (LLMs) is a promising yet challenging approach to building powerful multilingual LLMs.The biggest challenge is to make the model continuously learn new languages while preserving the proficient ability of old languages.To achieve this, recent work utilizes the Mixture-of-Experts (MoE) architecture to expand new languages by adding new experts and avoid catastrophic forgetting of old languages by routing corresponding tokens to the original model backbone (old experts).Although intuitive, this kind of method is parameter-costly when expanding new languages and still inevitably impacts the performance of old languages.To address these limitations, we analyze the language characteristics of different layers in LLMs and propose a layer-wise expert allocation algorithm (LayerMoE) to determine the appropriate number of new experts for each layer.Specifically, we find different layers in LLMs exhibit different representation similarities between languages and then utilize the similarity as the indicator to allocate experts for each layer, i.e., the higher similarity, the fewer experts.Additionally, to further mitigate the forgetting of old languages, we add a classifier in front of the router network on the layers with higher similarity to guide the routing of old language tokens.Experimental results show that our method outperforms the previous state-of-the-art baseline with 60% fewer experts in the single-expansion setting and with 33.3% fewer experts in the lifelong-expansion setting, demonstrating the effectiveness of our method.

Knowledge distillation (KD) is known as a promising solution to compress large language models (LLMs) via transferring their knowledge to smaller models. During this process, white-box KD methods usually minimize the distance between the output distributions of the two models so that more knowledge can be transferred. However, in the current white-box KD framework, the output distributions are from the respective output spaces of the two models, using their own prediction heads. We argue that the space discrepancy will lead to low similarity between the teacher model and the student model on both representation and distribution levels. Furthermore, this discrepancy also hinders the KD process between models with different vocabularies, which is common for current LLMs. To address these issues, we propose a dual-space knowledge distillation (DSKD) framework that unifies the output spaces of the two models for KD. On the basis of DSKD, we further develop a cross-model attention mechanism, which can automatically align the representations of the two models with different vocabularies. Thus, our framework is not only compatible with various distance functions for KD (e.g., KL divergence) like the current framework, but also supports KD between any two LLMs regardless of their vocabularies. Experiments on task-agnostic instruction-following benchmarks show that DSKD significantly outperforms the current white-box KD framework with various distance functions, and also surpasses existing KD methods for LLMs with different vocabularies.

Generating adversarial examples for Neural Machine Translation (NMT) with single Round-Trip Translation (RTT) has achieved promising results by releasing the meaning-preserving restriction. However, a potential pitfall for this approach is that we cannot decide whether the generated examples are adversarial to the target NMT model or the auxiliary backward one, as the reconstruction error through the RTT can be related to either. To remedy this problem, we propose a new definition for NMT adversarial examples based on the Doubly Round-Trip Translation (DRTT). Specifically, apart from the source-target-source RTT, we also consider the target-source-target one, which is utilized to pick out the authentic adversarial examples for the target NMT model. Additionally, to enhance the robustness of the NMT model, we introduce the masked language models to construct bilingual adversarial pairs based on DRTT, which are used to train the NMT model directly. Extensive experiments on both the clean and noisy test sets (including the artificial and natural noise) show that our approach substantially improves the robustness of NMT models.