Self-critique mechanisms significantly improve the performance of language models in complex reasoning tasks by giving them the ability to correct errors, conduct induction and deduction, and switch thinking insights. However, synthetic data methods often require human-introduced errors or sampling of the model’s reasoning results from the previous moment, and the current output distribution of the model cannot be obtained, makes the data for critique and reasoning face the problem of distribution shifts. In this work, we propose an on-policy reinforcement learning framework to synchronize the reasoning and critique capabilities of language models. To alleviate reward hacking caused by outcome-based supervision, we design a deliberate reward framework for different purposes. The reward framework not only supervises the model reasoning process based on the results, but also uses Monte Carlo sampling to give appropriate rewards to the critique content according to the success rate of the model’s correction after critique. In addition, we introduce a rule-based reward function to impose penalties on the model when it generates hallucinatory critiques. When our approach is applied to the DeepSeek-Math-7B-Base and Qwen2.5-7B-Base models, model performance improves 5.40 and 3.66 points, respectively, compared to the best baseline approach. This validates the significant advantages of our method in improving model’s reasoning and self-critique capability. Code will be made available at https://github.com/rbao2018/SCOP

Despite recent progress in Speech Translation (ST) research, the challenges posed by inherent speech phenomena that distinguish transcribed speech from written text are not well addressed. The informal and erroneous nature of spontaneous speech is inadequately represented in the typical parallel text available for building translation models. We propose to address these issues through a text rewrite approach that aims to transform transcribed speech into a cleaner style more in line with the expectations of translation models built from written text. Moreover, the advantages of the rewrite model can be effectively distilled into a standalone translation model. Experiments on several benchmarks, using both publicly available and in-house translation models, demonstrate that adding a rewrite model to a traditional ST pipeline is a cost-effect way to address a variety of speech irregularities and improve speech translation quality for multiple language directions and domains.

When the complete source sentence is provided, Large Language Models (LLMs) perform excellently in offline machine translation even with a simple prompt “Translate the following sentence from [src lang] into [tgt lang]:”. However, in many real scenarios, the source tokens arrive in a streaming manner and simultaneous machine translation (SiMT) is required, then the efficiency and performance of decoder-only LLMs are significantly limited by their auto-regressive nature. To enable LLMs to achieve high-quality SiMT as efficiently as offline translation, we propose a novel paradigm that includes constructing supervised fine-tuning (SFT) data for SiMT, along with new training and inference strategies. To replicate the token input/output stream in SiMT, the source and target tokens are rearranged into an interleaved sequence, separated by special tokens according to varying latency requirements. This enables powerful LLMs to learn read and write operations adaptively, based on varying latency prompts, while still maintaining efficient auto-regressive decoding. Experimental results show that, even with limited SFT data, our approach achieves state-of-the-art performance across various SiMT benchmarks and different evaluation metrics, and preserves the original capabilities of offline translation. Moreover, our approach generalizes well to document-level SiMT setting without requiring specific fine-tuning, even beyond the offline translation model.

The recent release of GPT-4o showcased the potential of end-to-end multimodal models, not just in terms of low latency but also in their ability to understand and generate expressive speech with rich emotions. While the details are unknown to the open research community, it likely involves significant amounts of curated data and compute, neither of which is readily accessible. In this paper, we present BLSP-Emo (Bootstrapped Language-Speech Pretraining with Emotion support), a novel approach to developing an end-to-end speech-language model capable of understanding both semantics and emotions in speech and generate empathetic responses. BLSP-Emo utilizes existing speech recognition (ASR) and speech emotion recognition (SER) datasets through a two-stage process. The first stage focuses on semantic alignment, following recent work on pretraining speech-language models using ASR data. The second stage performs emotion alignment with the pretrained speech-language model on an emotion-aware continuation task constructed from SER data. Our experiments demonstrate that the BLSP-Emo model excels in comprehending speech and delivering empathetic responses, both in instruction-following tasks and conversations.

Large language models (LLMs) have significantly improved in understanding natural language but still lack in mathematical reasoning, a hurdle on the path to true artificial general intelligence. The training of large language models, based on next-token prediction, struggles to capture the precise nature of mathematical reasoning, presenting both practical and theoretical challenges. In this paper, we address this challenge by enriching the data landscape and introducing a reasonable data format, enhanced the text analysis of the LLM with a capability to utilize a Python code interpreter. This dataset is derived from GSM8K and MATH and has been further refined through a combination of GPT annotations, human review, and self-training processes. Additionally, we propose a tentative, easily replicable protocol for the fine-tuning of math-specific LLMs, which has led to a significant improvement in the performance of a 7B-parameter LLM on the GSM8K and MATH datasets. A solution generator and a value estimator are fine-tuned simultaneously in a multi-task fashion, while an outlier-free value model-based inference method is proposed to further boost the performance. We are committed to advancing the field of mathematical reasoning in LLMs and, to that end, we will make the source code and checkpoints publicly available.

Pre-trained speech models, such as wav2vec 2.0, have significantly advanced speech-related tasks, including speech recognition and translation. However, their applicability in streaming scenarios is limited because these models are trained on complete utterances, leading to a mismatch with incremental streaming inputs. This paper identifies three critical design aspects within the architecture of wav2vec 2.0 and proposes a novel model, wav2vec-S, which incorporates simple modifications to ensure consistent speech representations during both training and inference phases for streaming speech inputs. Furthermore, we demonstrate that wav2vec-S models can be efficiently adapted from pre-trained wav2vec 2.0 models through continued pre-training and effectively finetuned to meet various latency requirements in downstream applications. Experiments on speech recognition and translation tasks show that wav2vec-S outperforms strong baseline models and achieves a superior balance between quality and latency.

Direct Preference Optimization (DPO) using an implicit reward model has proven to be an effective alternative to reinforcement learning from human feedback (RLHF) for fine-tuning preference aligned large language models (LLMs). However, the overall preference annotations of responses do not fully capture the fine-grained quality of model outputs in complex multi-step reasoning tasks, such as mathematical reasoning. To address this limitation, we introduce a novel algorithm called Step-level Value Preference Optimization (SVPO). Our approach employs Monte Carlo Tree Search (MCTS) to automatically annotate step-level preferences for multi-step reasoning. Furthermore, from the perspective of learning-to-rank, we train an explicit value model to replicate the behavior of the implicit reward model, complementing standard preference optimization. This value model enables the LLM to generate higher reward responses with minimal cost during inference. Experimental results demonstrate that our method achieves state-of-the-art performance on both in-domain and out-of-domain mathematical reasoning benchmarks.

A popular approach to streaming speech translation is to employ a single offline model with a wait-k policy to support different latency requirements, which is simpler than training multiple online models with different latency constraints. However, there is a mismatch problem in using a model trained with complete utterances for streaming inference with partial input. We demonstrate that speech representations extracted at the end of a streaming input are significantly different from those extracted from a complete utterance. To address this issue, we propose a new approach called Future-Aware Streaming Translation (FAST) that adapts an offline ST model for streaming input. FAST includes a Future-Aware Inference (FAI) strategy that incorporates future context through a trainable masked embedding, and a Future-Aware Distillation (FAD) framework that transfers future context from an approximation of full speech to streaming input. Our experiments on the MuST-C EnDe, EnEs, and EnFr benchmarks show that FAST achieves better trade-offs between translation quality and latency than strong baselines. Extensive analyses suggest that our methods effectively alleviate the aforementioned mismatch problem between offline training and online inference.

One of the main problems in speech translation is the mismatches between different modalities. The second problem, scarcity of parallel data covering multiple modalities, means that the end-to-end multi-modal models tend to perform worse than cascade models, although there are exceptions under favorable conditions. To address these problems, we propose an end-to-end zero-shot speech translation model, connecting two pre-trained uni-modality modules via word rotator’s distance. The model retains the ability of zero-shot, which is like cascade models, and also can be trained in an end-to-end style to avoid error propagation. Our comprehensive experiments on the MuST-C benchmarks show that our end-to-end zero-shot approach performs better than or as well as those of the CTC-based cascade models and that our end-to-end model with supervised training also matches the latest baselines.