Large language models (LLMs) have demonstrated remarkable progress in generating functional code, leading to numerous AI-based coding assistant tools. However, their reliance on the perplexity objective during both training and inference primarily emphasizes functionality, often at the expense of efficiency--an essential consideration for real-world coding tasks. Interestingly, we observed that well-trained LLMs inherently possess knowledge about code efficiency, but this potential remains underutilized with standard decoding approaches. To address this, we design strategic prompts to activate the model's embedded efficiency understanding, effectively using LLMs as efficiency critiques to guide code generation toward higher efficiency without sacrificing--and sometimes even improving--functionality, all without the need for costly real code execution. Extensive experiments on benchmark datasets (EffiBench, HumanEval+, COFFE, Mercury) across multiple representative code models demonstrate up to a 70.6% reduction in average execution time and a 13.6% decrease in maximum memory usage, highlighting the computational efficiency and practicality of our approach compared to existing alternatives.

Large Language Models (LLMs) have demonstrated exceptional performance across diverse natural language processing tasks. However, these models exhibit a critical limitation in output diversity, often generating highly similar content across multiple attempts. This limitation significantly affects tasks requiring diverse outputs, from creative writing to reasoning. Existing solutions, like temperature scaling, enhance diversity by modifying probability distributions but compromise output quality. We propose Guide-to-Generation (G2), a training-free plug-and-play method that enhances output diversity while preserving generation quality. G2 employs a base generator alongside dual Guides, which guide the generation process through decoding-based interventions to encourage more diverse outputs conditioned on the original query. Comprehensive experiments demonstrate that G2 effectively improves output diversity while maintaining an optimal balance between diversity and quality.

The sparse Mixture-of-Experts (MoE) has achieved significant progress for neural machine translation (NMT). However, there exist two limitations in current MoE solutions which may lead to sub-optimal performance: 1) they directly use the task knowledge of NMT into MoE (\emph{e.g.}, domain/linguistics-specific knowledge), which are generally unavailable at practical application and neglect the naturally grouped domain/linguistic properties; 2) the expert selection only depends on the localized token representation without considering the context, which fully grasps the state of each token in a global view. To address the above limitations, we propose THOR-MoE via arming the MoE with hierarchical task-guided and context-responsive routing policies. Specifically, it 1) firstly predicts the domain/language label and then extracts mixed domain/language representation to allocate task-level experts in a hierarchical manner; 2) injects the context information to enhance the token routing from the pre-selected task-level experts set, which can help each token to be accurately routed to more specialized and suitable experts. Extensive experiments on multi-domain translation and multilingual translation benchmarks with different architectures consistently demonstrate the superior performance of THOR-MoE. Additionally, the THOR-MoE operates as a plug-and-play module compatible with existing Top-$k$~\cite{shazeer2017} and Top-$p$~\cite{huang-etal-2024-harder} routing schemes, ensuring broad applicability across diverse MoE architectures. For instance, compared with vanilla Top-$p$~\cite{huang-etal-2024-harder} routing, the context-aware manner can achieve an average improvement of 0.75 BLEU with less than 22\% activated parameters on multi-domain translation tasks.

Contrastive decoding (CD) (Li et al., 2023) improves the next-token distribution of a large expert language model (LM) using a small amateur LM. Although CD is applied to various LMs and domains to enhance open-ended text generation, it is still unclear why CD often works well, when it could fail, and how we can make it better. To deepen our understanding of CD, we first theoretically prove that CD could be viewed as linearly extrapolating the next-token logits from a huge and hypothetical LM. We also highlight that the linear extrapolation could make CD unable to output the most obvious answers that have already been assigned high probabilities by the amateur LM. To overcome CD's limitation, we propose a new unsupervised decoding method called $\mathbf{A}$symptotic $\mathbf{P}$robability $\mathbf{D}$ecoding (APD). APD explicitly extrapolates the probability curves from the LMs of different sizes to infer the asymptotic probabilities from an infinitely large LM without inducing more inference costs than CD. In FactualityPrompts, an open-ended text generation benchmark, sampling using APD significantly boosts factuality in comparison to the CD sampling and its variants, and achieves state-of-the-art results for Pythia 6.9B and OPT 6.7B. Furthermore, in five commonsense QA datasets, APD is often significantly better than CD and achieves a similar effect of using a larger LLM. For example, the perplexity of APD on top of Pythia 6.9B is even lower than the perplexity of Pythia 12B in CommonsenseQA and LAMBADA.

Mixture of Experts (MoE) offers remarkable performance and computational efficiency by selectively activating subsets of model parameters. Traditionally, MoE models use homogeneous experts, each with identical capacity. However, varying complexity in input data necessitates experts with diverse capabilities, while homogeneous MoE hinders effective expert specialization and efficient parameter utilization. In this study, we propose a novel Heterogeneous Mixture of Experts (HMoE), where experts differ in size and thus possess diverse capacities. This heterogeneity allows for more specialized experts to handle varying token complexities more effectively. To address the imbalance in expert activation, we propose a novel training objective that encourages the frequent activation of smaller experts, enhancing computational efficiency and parameter utilization. Extensive experiments demonstrate that HMoE achieves lower loss with fewer activated parameters and outperforms conventional homogeneous MoE models on various pre-training evaluation benchmarks. Codes will be released upon acceptance.

Decoding methods for large language models (LLMs) usually struggle with the tradeoff between ensuring factuality and maintaining diversity. For example, a higher p threshold in the nucleus (top-p) sampling increases the diversity but decreases the factuality, and vice versa. In this paper, we propose REAL (Residual Entropy from Asymptotic Line) sampling, a decoding method that achieves improved factuality and diversity over nucleus sampling by predicting an adaptive threshold of $p$. Specifically, REAL sampling predicts the step-wise likelihood of an LLM to hallucinate, and lowers the p threshold when an LLM is likely to hallucinate. Otherwise, REAL sampling increases the p threshold to boost the diversity. To predict the step-wise hallucination likelihood without supervision, we construct a Token-level Hallucination Forecasting (THF) model to predict the asymptotic entropy (i.e., inherent uncertainty) of the next token by extrapolating the next-token entropies from a series of LLMs with different sizes. If a LLM's entropy is higher than the asymptotic entropy (i.e., the LLM is more uncertain than it should be), the THF model predicts a high hallucination hazard, which leads to a lower p threshold in REAL sampling. In the FactualityPrompts benchmark, we demonstrate that REAL sampling based on a 70M THF model can substantially improve the factuality and diversity of 7B LLMs simultaneously, judged by both retrieval-based metrics and human evaluation. After combined with contrastive decoding, REAL sampling outperforms 9 sampling methods, and generates texts that are more factual than the greedy sampling and more diverse than the nucleus sampling with $p=0.5$. Furthermore, the predicted asymptotic entropy is also a useful unsupervised signal for hallucination detection tasks.

Stochastic sampling strategies such as top-k and top-p have been widely used in dialogue generation task. However, as an open-domain chatting system, there will be two different conversation scenarios, i.e. chit-chat and knowledge-based question answering. In the former situation, responses diversity is essential due to the one-to-many nature in dialogue. The latter, on the other hand, requires less randomness given that stochastic decoding strategy entails the risk of generating incorrect information. As a result, an adaptive and flexible decoding strategy is needed to cope with these two scenarios simultaneously. To this end, we propose the dynamic decoding strategy (DDS), which can adjust the decoding space w.r.t. different contexts. In DDS, both sequence-level and token-level adaptive search can be achieved to adjust the decoding process in a unified framework. Besides, our adaptive algorithm can not only be used during model inference, but it can also be applied during the model training stage to further enhance the performance. Comprehensive experiments indicate that the proposed decoding strategy can consistently improve the performance of pre-trained dialogue models when coupled with four well-used stochastic decoding algorithms.

Digital ink (online handwriting) generation has a number of potential applications for creating user-visible content, such as handwriting autocompletion, spelling correction, and beautification. Writing is personal and usually the processing is done on-device. Ink generative models thus need to produce high quality content quickly, in a resource constrained environment. In this work, we study ways to maximize the quality of the output of a trained digital ink generative model, while staying within an inference time budget. We use and compare the effect of multiple sampling and ranking techniques, in the first ablation study of its kind in the digital ink domain. We confirm our findings on multiple datasets - writing in English and Vietnamese, as well as mathematical formulas - using two model types and two common ink data representations. In all combinations, we report a meaningful improvement in the recognizability of the synthetic inks, in some cases more than halving the character error rate metric, and describe a way to select the optimal combination of sampling and ranking techniques for any given computational budget.

In recent years, there has been an increasing interest in open-ended language generation thanks to the rise of large transformer-based language models trained on millions of webpages, such as OpenAI's famous GPT2 model. The results on conditioned open-ended language generation are impressive, e.g. Besides the improved transformer architecture and massive unsupervised training data, better decoding methods have also played an important role. This blog post gives a brief overview of different decoding strategies and more importantly shows how you can implement them with very little effort using the popular transformers library! All of the following functionalities can be used for auto-regressive language generation (here a refresher).