The Invisible Lottery: How Subtle Cues Steer Algorithm Choice in LLM Code Generation

Large language models (LLMs) are increasingly used to implement algorithms from research manuscripts, but papers often leave implementation choices implicit. This study examines how the written format of an algorithm specification affects first-pass LLM implementation accuracy. We compare ordinary prose, LaTeX algorithm-style pseudocode, PDF-like extracted pseudocode, Markdown fields, YAML-like specifications, JSON-like specifications, and Python code stubs across five machine learning tasks, three models, and four experimental settings, yielding 4,020 generated implementations. Hidden tests evaluate details that often determine correctness, including tie-breaking, array shapes, numerical rules, return structures, and invalid-input behavior. Under the core-information setting, LaTeX algorithm-style pseudocode has the largest average format effect, with YAML-like specifications and ordinary prose close behind. Under complete information, GPT-5.4 mini shows no format differences in the matched comparisons, whereas Gemma 3 4B and Llama 3.2 3B still do. Code stubs do not consistently improve correctness despite specifying the function signature. The results support a writing recommendation: authors should state the interface, computation steps, numerical rules, and boundary-case behavior explicitly, instead of relying on a particular surface format to carry those details.

High pass rates on established programming benchmarks such as HumanEval and LiveCodeBench do not always show whether a model can reason about algorithms. Many fixed benchmarks eventually become part of the public training ecosystem through released problem statements, editorials, and generated solutions, allowing later models to improve partly by exposure rather than by stronger algorithmic ability. We introduce ALGOBENCH, a framework that automatically builds novel algorithmic problems from known competitive-programming problems through structured constraint-shifting transformations. Each accepted ALGOBENCH variant is traceable to a source problem, but must make the original reference algorithm fail. Beyond pass@$k$, we introduce complexity-aware metrics -- including OPTT, OPTS, TRAPRATE, GAPT, and CONSENS -- to test whether a solution is not only functionally correct but also asymptotically suitable for the generated problem. Experiments across multiple LLMs and prompting strategies show that performance drops sharply on ALGOBENCH variants, retrieval can increase reuse of the old algorithm, and many correct-looking solutions fail to meet the required complexity. Error analysis shows that failures are mainly algorithmic rather than implementation-level, suggesting that ALGOBENCH evaluates adaptation beyond functional correctness.

This study empirically evaluates whether cost-efficient Large Language Models (LLMs) can be trusted to generate enterprise code to a written specification. Three models (Gemini Flash 3, GPT-5.4 mini and Claude Haiku 4.5) were asked to solve 992 algorithmic problems as Java Spring Boot service methods conforming to a mandated signature and data-transfer-object specification, crossing four model and agentic coding tool combinations with two prompt variants to yield eight configurations, with iteration forbidden and hardcoded answers explicitly prohibited. Eight problem statements were withheld to probe how models respond to missing input. The 7,593 resulting methods were classified by an eight-class outcome taxonomy describing what each does about producing an answer, then deployed and executed, giving 7,936 measured requests joined to that classification. Structural conformance approached ceiling, yet 38.4% of methods do not compute the value they returned and only 12.9% of returned answers were correct. Conditioning on outcome class shows that response reliability and correctness are inversely related, whereas genuinely computing methods answered least often and were correct 19.3%. Limitations include single generation runs per configuration, partial harness coverage, single-pass timing, syntactic classification, and probable corpus contamination.