Large Language Models (LLMs) are widely used by software engineers for programming tasks. However, research shows that LLMs often lack a deep understanding of program semantics. Even minor changes to syntax, such as renaming variables, can significantly degrade performance across various tasks. In this work, we examine the task of *type prediction*: given a partially typed program, can a model predict a missing type annotations such that the resulting program is more typed? We construct a dataset of adversarial examples where models initially predict the correct types, but begin to fail after semantically irrelevant edits. This is problematic, as models should ideally generalize across different syntactic forms of semantically equivalent code. This lack of robustness suggests that models may have a shallow understanding of code semantics.Despite this, we provide evidence that LLMs do, in fact, learn robust mechanisms for type prediction—though these mechanisms often fail to activate in adversarial scenarios. By using *activation steering*, a method that manipulates a model’s internal activations to guide it toward using latent knowledge, we restore accurate predictions on adversarial inputs. We show that steering successfully activates a type prediction mechanism that is shared by both Python and TypeScript, and is more effective than prompting with in-context examples. Across five different models, our comprehensive evaluation demonstrates that LLMs can learn generalizable representations of code semantics that transfer across programming languages.

Although LLMs are increasing the productivity of professional programmers, existing work shows that beginners struggle to prompt LLMs to solve text-to-code tasks (Nguyen et al., 2024; Prather et al., 2024b; Mordechai et al., 2024). Why is this the case? This paper explores two competing hypotheses about the cause of student-LLM miscommunication: (1) students simply lack the technical vocabulary needed to write good prompts, and (2) students do not understand the extent of information that LLMs need to solve code generation tasks. We study (1) with a causal intervention experiment on technical vocabulary and (2) by analyzing graphs that abstract how students edit prompts and the different failures that they encounter. We find that substance beats style: a poor grasp of technical vocabulary is merely correlated with prompt failure; that the information content of prompts predicts success; that students get stuck making trivial edits; and more. Our findings have implications for the use of LLMs in programming education, and for efforts to make computing more accessible with LLMs.

Code LLMs have the potential to make it easier for non-experts to understand and write code. However, current CodeLLM benchmarks rely on a single expert-written prompt per problem, making it hard to generalize their success to non-expert users. In this paper, we present a new natural-language-to-code benchmark of prompts written by a key population of non-experts: beginning programmers. StudentEval contains 1,749 prompts written by 80 students who have only completed one introductory Python course. StudentEval contains numerous non-expert prompts describing the same problem, enabling exploration of key factors in prompt success. We use StudentEval to evaluate 12 Code LLMs and find that StudentEval is a better discriminator of model performance than existing benchmarks. Our analysis of student prompting strategies reveals that nondeterministic LLM sampling can mislead students about the quality of their descriptions, a finding with key implications for Code LLMs in education.