Generating code is an important application of Large Language Models (LLMs) and the task of function completion is one of the core open challenges in this context. Existing approaches focus on either training, fine-tuning or prompting LLMs to generate better outputs given the same input. We propose a novel and complementary approach: to optimize part of the input, the docstring (summary of a function’s purpose and usage), via reformulation with an LLM, in order to improve code generation. We develop two baseline methods for optimizing code generation via docstring reformulation and test them on the original HumanEval benchmark and multiple curated variants which are made more challenging by realistically worsening the docstrings. Our results show that, when operating on docstrings reformulated by an LLM instead of the original (or worsened) inputs, the performance of a number of open-source LLMs does not change significantlyThis finding demonstrates an unexpected robustness of current open-source LLMs to the details of the docstrings. We conclude by examining a series of questions, accompanied by in-depth analyses, pertaining to the sensitivity of current open-source LLMs to the details in the docstrings, the potential for improvement via docstring reformulation and the limitations of the methods employed in this work.

We explore how we can build accurate world models, which are partially specified by language, and how we can plan with them in the face of novelty and uncertainty. We propose the first model-based reinforcement learning approach to tackle the environment Read To Fight Monsters (Zhong et al., 2019), a grounded policy learning problem. In RTFM an agent has to reason over a set of rules and a goal, both described in a language manual, and the observations, while taking into account the uncertainty arising from the stochasticity of the environment, in order to generalize successfully its policy to test episodes. We demonstrate the superior performance and sample efficiency of our model-based approach to the existing model-free SOTA agents in eight variants of RTFM. Furthermore, we show how the agent’s plans can be inspected, which represents progress towards more interpretable agents.

We provide a study of how induced model sparsity can help achieve compositional generalization and better sample efficiency in grounded language learning problems. We consider simple language-conditioned navigation problems in a grid world environment with disentangled observations. We show that standard neural architectures do not always yield compositional generalization. To address this, we design an agent that contains a goal identification module that encourages sparse correlations between words in the instruction and attributes of objects, composing them together to find the goal. The output of the goal identification module is the input to a value iteration network planner. Our agent maintains a high level of performance on goals containing novel combinations of properties even when learning from a handful of demonstrations. We examine the internal representations of our agent and find the correct correspondences between words in its dictionary and attributes in the environment.