The ability of zero-shot translation emerges when we train a multilingual model with certain translation directions; the model can then directly translate in unseen directions. Alternatively, zero-shot translation can be accomplished by pivoting through a third language (e.g., English). In our work, we observe that both direct and pivot translations are noisy and achieve less satisfactory performance. We propose EBBS, an ensemble method with a novel bi-level beam search algorithm, where each ensemble component explores its own prediction step by step at the lower level but all components are synchronized by a "soft voting" mechanism at the upper level. Results on two popular multilingual translation datasets show that EBBS consistently outperforms direct and pivot translations, as well as existing ensemble techniques. Further, we can distill the ensemble's knowledge back to the multilingual model to improve inference efficiency; profoundly, our EBBS-distilled model can even outperform EBBS as it learns from the ensemble knowledge.

Final-answer-based metrics are commonly used for evaluating large language models (LLMs) on math word problems, often taken as proxies for reasoning ability. However, such metrics conflate two distinct sub-skills: abstract formulation (capturing mathematical relationships using expressions) and arithmetic computation (executing the calculations). Through a disentangled evaluation on GSM8K and SVAMP, we find that the final-answer accuracy of Llama-3 and Qwen2.5 (1B-32B) without CoT is overwhelmingly bottlenecked by the arithmetic computation step and not by the abstract formulation step. Contrary to the common belief, we show that CoT primarily aids in computation, with limited impact on abstract formulation. Mechanistically, we show that these two skills are composed conjunctively even in a single forward pass without any reasoning steps via an abstract-then-compute mechanism: models first capture problem abstractions, then handle computation. Causal patching confirms these abstractions are present, transferable, composable, and precede computation. These behavioural and mechanistic findings highlight the need for disentangled evaluation to accurately assess LLM reasoning and to guide future improvements.

We present substantial evidence demonstrating the benefits of integrating Large Language Models (LLMs) with a Contextual Multi-Armed Bandit framework.Contextual bandits have been widely used in recommendation systems to generate personalized suggestions based on user-specific contexts.We show that LLMs, pre-trained on extensive corpora rich in human knowledge and preferences, can simulate human behaviours well enough to jumpstart contextual multi-armed bandits to reduce online learning regret.We propose an initialization algorithm for contextual bandits by prompting LLMs to produce a pre-training dataset of approximate human preferences for the bandit.This significantly reduces online learning regret and data-gathering costs for training such models.Our approach is validated empirically through two sets of experiments with different bandit setups: one which utilizes LLMs to serve as an oracle and a real-world experiment utilizing data from a conjoint survey experiment.

We investigate the unsupervised constituency parsing task, which organizes words and phrases of a sentence into a hierarchical structure without using linguistically annotated data. We observe that existing unsupervised parsers capture different aspects of parsing structures, which can be leveraged to enhance unsupervised parsing performance. To this end, we propose a notion of "tree averaging," based on which we further propose a novel ensemble method for unsupervised parsing. To improve inference efficiency, we further distill the ensemble knowledge into a student model; such an ensemble-then-distill process is an effective approach to mitigate the over-smoothing problem existing in common multi-teacher distilling methods. Experiments show that our method surpasses all previous approaches, consistently demonstrating its effectiveness and robustness across various runs, with different ensemble components, and under domain-shift conditions.

Despite large neural networks demonstrating remarkable abilities to complete different tasks, they require excessive memory usage to store the optimization states for training. To alleviate this, the low-rank adaptation (LoRA) is proposed to reduce the optimization states by training fewer parameters. However, LoRA restricts overall weight update matrices to be low-rank, limiting the model performance. In this work, we investigate the dynamics of LoRA and identify that it can be approximated by a random projection. Based on this observation, we propose Flora, which is able to achieve high-rank updates by resampling the projection matrices while enjoying the sublinear space complexity of optimization states. We conduct experiments across different tasks and model architectures to verify the effectiveness of our approach.

How do large language models (LLMs) encode the state of the world, including the status of entities and their relations, as described by a text? While existing work directly probes for a complete state of the world, our research explores whether and how LLMs abstract this world state in their internal representations. We propose a new framework for probing for world representations through the lens of state abstraction theory from reinforcement learning, which emphasizes different levels of abstraction, distinguishing between general abstractions that facilitate predicting future states and goal-oriented abstractions that guide the subsequent actions to accomplish tasks. To instantiate this framework, we design a text-based planning task, where an LLM acts as an agent in an environment and interacts with objects in containers to achieve a specified goal state. Our experiments reveal that fine-tuning as well as advanced pre-training strengthens LLM-built representations' tendency of maintaining goal-oriented abstractions during decoding, prioritizing task completion over recovery of the world's state and dynamics.

Large Language Models (LLMs) have shown remarkable performance in various natural language tasks, but they often struggle with planning problems that require structured reasoning. To address this limitation, the conversion of planning problems into the Planning Domain Definition Language (PDDL) has been proposed as a potential solution, enabling the use of automated planners. However, generating accurate PDDL files typically demands human inputs or correction, which can be time-consuming and costly. In this paper, we propose a novel approach that leverages LLMs and environment feedback to automatically generate PDDL domain and problem description files without the need for human intervention. Our method introduces an iterative refinement process that generates multiple problem PDDL candidates and progressively refines the domain PDDL based on feedback obtained from interacting with the environment. To guide the refinement process, we develop an Exploration Walk (EW) metric, which provides rich feedback signals for LLMs to update the PDDL file. We evaluate our approach on $10$ PDDL environments. We achieve an average task solve rate of 66\% compared to a 29\% solve rate by GPT-4's intrinsic planning with chain-of-thought prompting. Our work enables the automated modeling of planning environments using LLMs and environment feedback, eliminating the need for human intervention in the PDDL translation process and paving the way for more reliable LLM agents in challenging problems. Our code is available at https://github.com/BorealisAI/llm-pddl-planning

Peng Xu, Dhruv Kumar, Wei Yang, Wenjie Zi, Keyi Tang, Chenyang Huang, Jackie Chi Kit Cheung, Simon J.D. Prince, Yanshuai Cao. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

In this work, we develop a novel regularizer to improve the learning of long-range dependency of sequence data. Applied on language modelling, our regularizer expresses the inductive bias that sequence variables should have high mutual information even though the model might not see abundant observations for complex long-range dependency. We show how the “next sentence prediction (classification)" heuristic can be derived in a principled way from our mutual information estimation framework, and be further extended to maximize the mutual information of sequence variables. The proposed approach not only is effective at increasing the mutual information of segments under the learned model but more importantly, leads to a higher likelihood on holdout data, and improved generation quality. Code is releasedat https://github.com/BorealisAI/BMI.

The field of deep generative modeling has succeeded in producing astonishingly realistic-seeming images and audio, but quantitative evaluation remains a challenge. Log-likelihood is an appealing metric due to its grounding in statistics and information theory, but it can be challenging to estimate for implicit generative models, and scalar-valued metrics give an incomplete picture of a model’s quality. In this work, we propose to use rate distortion (RD) curves to evaluate and compare deep generative models. While estimating RD curves is seemingly even more computationally demanding than log-likelihood estimation, we show that we can approximate the entire RD curve using nearly the same computations as were previously used to achieve a single log-likelihood estimate. We evaluate lossy compression rates of VAEs, GANs, and adversarial autoencoders (AAEs) on the MNIST and CIFAR10 datasets. Measuring the entire RD curve gives a more complete picture than scalar-valued metrics, and we arrive at a number of insights not obtainable from log-likelihoods alone.

The variational autoencoder (VAE) can learn the manifold of natural images on certain datasets, as evidenced by meaningful interpolating or extrapolating in the continuous latent space. However, on discrete data such as text, it is unclear if unsupervised learning can discover similar latent space that allows controllable manipulation. In this work, we find that sequence VAEs trained on text fail to properly decode when the latent codes are manipulated, because the modified codes often land in holes or vacant regions in the aggregated posterior latent space, where the decoding network fails to generalize. Both as a validation of the explanation and as a fix to the problem, we propose to constrain the posterior mean to a learned probability simplex, and performs manipulation within this simplex. Our proposed method mitigates the latent vacancy problem and achieves the first success in unsupervised learning of controllable representations for text. Empirically, our method outperforms unsupervised baselines and strong supervised approaches on text style transfer, and is capable of performing more flexible fine-grained control over text generation than existing methods.

Peng Xu, Hamidreza Saghir, Jin Sung Kang, Teng Long, Avishek Joey Bose, Yanshuai Cao, Jackie Chi Kit Cheung. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019.

Learning by contrasting positive and negative samples is a general strategy adopted by many methods.Noise contrastive estimation (NCE) for word embeddings and translating embeddings for knowledge graphs are examples in NLP employing this approach.In this work, we view contrastive learning as an abstraction of all such methods and augment the negative sampler into a mixture distribution containing an adversarially learned sampler.The resulting adaptive sampler finds harder negative examples, which forces the main model to learn a better representation of the data.We evaluate our proposal on learning word embeddings, order embeddings and knowledge graph embeddings and observe both faster convergence and improved results on multiple metrics.

We propose an efficient optimization algorithm to select a subset of training data as the inducing set for sparse Gaussian process regression. Previous methods either use different objective functions for inducing set and hyperparameter selection, or else optimize the inducing set by gradient-based continuous optimization. The former approaches are harder to interpret and suboptimal, whereas the latter cannot be applied to discrete input domains or to kernel functions that are not differentiable with respect to the input. The algorithm proposed in this work estimates an inducing set and the hyperparameters using a single objective. It can be used to optimize either the marginal likelihood or a variational free energy. Space and time complexity are linear in training set size, and the algorithm can be applied to large regression problems on discrete or continuous domains. Empirical evaluation shows state-of-art performance in discrete cases, competitive prediction results as well as a favorable trade-off between training and test time in continuous cases.

We propose an efficient discrete optimization algorithm for selecting a subset of training data to induce sparsity for Gaussian process regression. The algorithm estimates this inducing set and the hyperparameters using a single objective, either the marginal likelihood or a variational free energy. The space and time complexity are linear in the training set size, and the algorithm can be applied to large regression problems on discrete or continuous domains. Empirical evaluation shows state-of-art performance in the discrete case and competitive results in the continuous case.