Tongran Liu

Chenglong Wang, Hang Zhou, Yimin Hu et al.

Applying Reinforcement Learning (RL) to sequence generation models enables the direct optimization of long-term rewards (\textit{e.g.,} BLEU and human feedback), but typically requires large-scale sampling over a space of action sequences. This is a computational challenge as presented by the practice of sequence generation problems, such as machine translation, where we often deal with a large action space (\textit{e.g.,} a vocabulary) and a long action sequence (\textit{e.g.,} a translation). In this work, we introduce two-stage sampling and dynamic sampling approaches to improve the sampling efficiency during training sequence generation models via RL. We experiment with our approaches on the traditional sequence generation tasks, including machine translation and abstractive summarization. Furthermore, we evaluate our approaches in RL from human feedback (RLHF) through training a large language model using the reward model. Experimental results show that the efficient sampling-based RL, referred to as ESRL, can outperform all baselines in terms of both training efficiency and memory consumption. Notably, ESRL yields consistent performance gains over the strong REINFORCE, minimum risk training, and proximal policy optimization methods.

Chenglong Wang, Hang Zhou, Kaiyan Chang et al.

Automatic evaluation of sequence generation, traditionally reliant on metrics like BLEU and ROUGE, often fails to capture the semantic accuracy of generated text sequences due to their emphasis on n-gram overlap. A promising solution to this problem is to develop model-based metrics, such as BLEURT and COMET. However, these approaches are typically hindered by the scarcity of labeled evaluation data, which is necessary to train the evaluation models. In this work, we build upon this challenge by proposing the Customized Sequence Evaluation Metric (CSEM), a three-stage evaluation model training method that utilizes large language models to generate labeled data for model-based metric development, thereby eliminating the need for human-labeled data. Additionally, we expand the scope of CSEM to support various evaluation types, including single-aspect, multi-aspect, reference-free, and reference-based evaluations, enabling the customization of metrics to suit diverse real-world scenarios. Experimental results on the SummEval benchmark demonstrate that CSEM can effectively train an evaluation model without human-labeled data. Further experiments in reinforcement learning and reranking show that metrics developed through CSEM outperform traditional evaluation metrics, leading to substantial improvements in sequence quality as evaluated by both commonly used metrics and ChatGPT.

Chenglong Wang, Yang Gan, Yifu Huo et al.

Large vision-language models (LVLMs) often fail to align with human preferences, leading to issues like generating misleading content without proper visual context (also known as hallucination). A promising solution to this problem is using human-preference alignment techniques, such as best-of-n sampling and reinforcement learning. However, these techniques face the difficulty arising from the scarcity of visual preference data, which is required to train a visual reward model (VRM). In this work, we continue the line of research. We present a Robust Visual Reward Model (RoVRM) which improves human-preference alignment for LVLMs. RoVRM leverages auxiliary textual preference data through a three-phase progressive training and optimal transport-based preference data selection to effectively mitigate the scarcity of visual preference data. We experiment with RoVRM on the commonly used vision-language tasks based on the LLaVA-1.5-7B and -13B models. Experimental results demonstrate that RoVRM consistently outperforms traditional VRMs. Furthermore, our three-phase progressive training and preference data selection approaches can yield consistent performance gains over ranking-based alignment techniques, such as direct preference optimization.

Yifu Huo, Chenglong Wang, Ziming Zhu et al.

Reinforcement learning (RL) has emerged as a promising paradigm for training reasoning-oriented models by leveraging rule-based reward signals. However, RL training typically tends to improve single-sample success rates (i.e., Pass@1) while offering limited exploration of diverse reasoning trajectories, which is crucial for multi-sample performance (i.e., Pass@k). Our preliminary analysis reveals that this limitation stems from a fundamental squeezing effect, whereby probability mass is excessively concentrated on a narrow subset of high-reward trajectories, restricting genuine exploration and constraining attainable performance under RL training. To address this issue, in this work, we propose Steering Probability Squeezing (SPS), a training paradigm that interleaves conventional RL with inverse reinforcement learning (IRL). SPS treats on-policy rollouts as demonstrations and employs IRL to explicitly reshape the induced trajectory distribution, thereby enhancing exploration without introducing external supervision. Experiments on five commonly used reasoning benchmarks demonstrate that SPS can enable better exploration and improve Pass@k. Beyond algorithmic contributions, we provide an analysis of RL learning dynamics and identify an empirical upper bound on Pass@k, shedding light on intrinsic exploration limits in RL-based reasoning models. Our findings suggest that alternating between RL and IRL offers an effective pathway toward extending the exploration capacity of reasoning-oriented large language models.

Yanyang Li, Qiang Wang, Tong Xiao et al.

Though early successes of Statistical Machine Translation (SMT) systems are attributed in part to the explicit modelling of the interaction between any two source and target units, e.g., alignment, the recent Neural Machine Translation (NMT) systems resort to the attention which partially encodes the interaction for efficiency. In this paper, we employ Joint Representation that fully accounts for each possible interaction. We sidestep the inefficiency issue by refining representations with the proposed efficient attention operation. The resulting Reformer models offer a new Sequence-to- Sequence modelling paradigm besides the Encoder-Decoder framework and outperform the Transformer baseline in either the small scale IWSLT14 German-English, English-German and IWSLT15 Vietnamese-English or the large scale NIST12 Chinese-English translation tasks by about 1 BLEU point.We also propose a systematic model scaling approach, allowing the Reformer model to beat the state-of-the-art Transformer in IWSLT14 German-English and NIST12 Chinese-English with about 50% fewer parameters. The code is publicly available at https://github.com/lyy1994/reformer.

Chenglong Wang, Yang Gan, Yifu Huo et al.

In aligning large language models (LLMs), reward models have played an important role, but are standardly trained as discriminative models and rely only on labeled human preference data. In this paper, we explore methods that train reward models using both unlabeled and labeled data. Building on the generative models in LLMs, we develop a generative reward model that is first trained via large-scale unsupervised learning and then fine-tuned via supervised learning. We also show that by using label smoothing, we are in fact optimizing a regularized pairwise ranking loss. This result, in turn, provides a new view of training reward models, which links generative models and discriminative models under the same class of training objectives. The outcome of these techniques is a foundation reward model, which can be applied to a wide range of tasks with little or no further fine-tuning effort. Extensive experiments show that this model generalizes well across several tasks, including response ranking, reinforcement learning from human feedback, and task adaptation with fine-tuning, achieving significant performance improvements over several strong baseline models.

Yifu Huo, Chenglong Wang, Qiren Zhu et al.

Preference optimization methods like DPO have achieved remarkable performance in LLM alignment. However, the evaluation for these methods relies on a single response and overlooks other potential outputs, which could also be generated in real-world applications within this hypothetical space. To address this issue, this paper presents a \textbf{H}ypothesis-based Pr\textbf{E}ference-aware \textbf{A}na\textbf{L}ysis Framework (HEAL), a novel evaluation paradigm that formulates preference alignment as a re-ranking process within hypothesis spaces. The framework incorporates two complementary metrics: ranking accuracy for evaluating ordinal consistency and preference strength correlation for assessing continuous alignment. To facilitate this framework, we develop UniHypoBench, a unified hypothesis benchmark constructed from diverse instruction-response pairs. Through extensive experiments based on HEAL, with a particular focus on the intrinsic mechanisms of preference learning, we demonstrate that current preference learning methods can effectively capture preferences provided by proxy models while simultaneously suppressing negative samples. These findings contribute to preference learning research through two significant avenues. Theoretically, we introduce hypothesis space analysis as an innovative paradigm for understanding preference alignment. Practically, HEAL offers researchers robust diagnostic tools for refining preference optimization methods, while our empirical results identify promising directions for developing more advanced alignment algorithms capable of comprehensive preference capture.

Chenglong Wang, Yifu Huo, Yang Gan et al.

Previous methods evaluate reward models by testing them on a fixed pairwise ranking test set, but they typically do not provide performance information on each preference dimension. In this work, we address the evaluation challenge of reward models by probing preference representations. To confirm the effectiveness of this evaluation method, we construct a Multi-dimensional Reward Model Benchmark (MRMBench), a collection of six probing tasks for different preference dimensions. We design it to favor and encourage reward models that better capture preferences across different dimensions. Furthermore, we introduce an analysis method, inference-time probing, which identifies the dimensions used during the reward prediction and enhances its interpretability. Through extensive experiments, we find that MRMBench strongly correlates with the alignment performance of large language models (LLMs), making it a reliable reference for developing advanced reward models. Our analysis of MRMBench evaluation results reveals that reward models often struggle to capture preferences across multiple dimensions, highlighting the potential of multi-objective optimization in reward modeling. Additionally, our findings show that the proposed inference-time probing method offers a reliable metric for assessing the confidence of reward predictions, which ultimately improves the alignment of LLMs.

Chenglong Wang, Yang Gan, Hang Zhou et al.

Recent advances in diffusion language models (DLMs) have presented a promising alternative to traditional autoregressive large language models (LLMs). However, DLMs still lag behind LLMs in reasoning performance, especially as the number of denoising steps decreases. Our analysis reveals that this shortcoming arises primarily from the independent generation of masked tokens across denoising steps, which fails to capture the token correlation. In this paper, we define two types of token correlation: intra-sequence correlation and inter-sequence correlation, and demonstrate that enhancing these correlations improves reasoning performance. To this end, we propose a Multi-Reward Optimization (MRO) approach, which encourages DLMs to consider the token correlation during the denoising process. More specifically, our MRO approach leverages test-time scaling, reject sampling, and reinforcement learning to directly optimize the token correlation with multiple elaborate rewards. Additionally, we introduce group step and importance sampling strategies to mitigate reward variance and enhance sampling efficiency. Through extensive experiments, we demonstrate that MRO not only improves reasoning performance but also achieves significant sampling speedups while maintaining high performance on reasoning benchmarks.

Chenglong Wang, Hang Zhou, Kaiyan Chang et al.

Alignment training is crucial for enabling large language models (LLMs) to cater to human intentions and preferences. It is typically performed based on two stages with different objectives: instruction-following alignment and human-preference alignment. However, aligning LLMs with these objectives in sequence suffers from an inherent problem: the objectives may conflict, and the LLMs cannot guarantee to simultaneously align with the instructions and human preferences well. To response to these, in this work, we propose a Hybrid Alignment Training (Hbat) approach, based on alternating alignment and modified elastic weight consolidation methods. The basic idea is to alternate between different objectives during alignment training, so that better collaboration can be achieved between the two alignment tasks.We experiment with Hbat on summarization and dialogue tasks. Experimental results show that the proposed \textsc{Hbat} can significantly outperform all baselines. Notably, Hbat yields consistent performance gains over the traditional two-stage alignment training when using both proximal policy optimization and direct preference optimization.

Yongyu Mu, Peinan Feng, Zhiquan Cao et al.

In this study, we reveal an in-context learning (ICL) capability of multilingual large language models (LLMs): by translating the input to several languages, we provide Parallel Input in Multiple Languages (PiM) to LLMs, which significantly enhances their comprehension abilities. To test this capability, we design extensive experiments encompassing 8 typical datasets, 7 languages and 8 state-of-the-art multilingual LLMs. Experimental results show that (1) incorporating more languages help PiM surpass the conventional ICL further; (2) even combining with the translations that are inferior to baseline performance can also help. Moreover, by examining the activated neurons in LLMs, we discover a counterintuitive but interesting phenomenon. Contrary to the common thought that PiM would activate more neurons than monolingual input to leverage knowledge learned from diverse languages, PiM actually inhibits neurons and promotes more precise neuron activation especially when more languages are added. This phenomenon aligns with the neuroscience insight about synaptic pruning, which removes less used neural connections, strengthens remainders, and then enhances brain intelligence.

Ye Lin, Yanyang Li, Tengbo Liu et al.

8-bit integer inference, as a promising direction in reducing both the latency and storage of deep neural networks, has made great progress recently. On the other hand, previous systems still rely on 32-bit floating point for certain functions in complex models (e.g., Softmax in Transformer), and make heavy use of quantization and de-quantization. In this work, we show that after a principled modification on the Transformer architecture, dubbed Integer Transformer, an (almost) fully 8-bit integer inference algorithm Scale Propagation could be derived. De-quantization is adopted when necessary, which makes the network more efficient. Our experiments on WMT16 En<->Ro, WMT14 En<->De and En->Fr translation tasks as well as the WikiText-103 language modelling task show that the fully 8-bit Transformer system achieves comparable performance with the floating point baseline but requires nearly 4x less memory footprint.

Bei Li, Hui Liu, Ziyang Wang et al.

In encoder-decoder neural models, multiple encoders are in general used to represent the contextual information in addition to the individual sentence. In this paper, we investigate multi-encoder approaches in documentlevel neural machine translation (NMT). Surprisingly, we find that the context encoder does not only encode the surrounding sentences but also behaves as a noise generator. This makes us rethink the real benefits of multi-encoder in context-aware translation - some of the improvements come from robust training. We compare several methods that introduce noise and/or well-tuned dropout setup into the training of these encoders. Experimental results show that noisy training plays an important role in multi-encoder-based NMT, especially when the training data is small. Also, we establish a new state-of-the-art on IWSLT Fr-En task by careful use of noise generation and dropout methods.

Yinqiao Li, Chi Hu, Yuhao Zhang et al.

Neural architecture search (NAS) has advanced significantly in recent years but most NAS systems restrict search to learning architectures of a recurrent or convolutional cell. In this paper, we extend the search space of NAS. In particular, we present a general approach to learn both intra-cell and inter-cell architectures (call it ESS). For a better search result, we design a joint learning method to perform intra-cell and inter-cell NAS simultaneously. We implement our model in a differentiable architecture search system. For recurrent neural language modeling, it outperforms a strong baseline significantly on the PTB and WikiText data, with a new state-of-the-art on PTB. Moreover, the learned architectures show good transferability to other systems. E.g., they improve state-of-the-art systems on the CoNLL and WNUT named entity recognition (NER) tasks and CoNLL chunking task, indicating a promising line of research on large-scale pre-learned architectures.

Tong Xiao, Yinqiao Li, Jingbo Zhu et al.

Recently, the Transformer machine translation system has shown strong results by stacking attention layers on both the source and target-language sides. But the inference of this model is slow due to the heavy use of dot-product attention in auto-regressive decoding. In this paper we speed up Transformer via a fast and lightweight attention model. More specifically, we share attention weights in adjacent layers and enable the efficient re-use of hidden states in a vertical manner. Moreover, the sharing policy can be jointly learned with the MT model. We test our approach on ten WMT and NIST OpenMT tasks. Experimental results show that it yields an average of 1.3X speed-up (with almost no decrease in BLEU) on top of a state-of-the-art implementation that has already adopted a cache for fast inference. Also, our approach obtains a 1.8X speed-up when it works with the \textsc{Aan} model. This is even 16 times faster than the baseline with no use of the attention cache.