LLMs Struggle to Convert Research Papers into Code

The IdeaAMBIG benchmark demonstrates that current LLMs achieve only a 9.6% success rate in Defect Localization, the process of identifying implementation gaps in research paper methodologies. The dataset evaluates 660 instances to determine their codification readiness, defined as the ability of a system to implement a method without making unsupported assumptions. Researchers constructed the benchmark using 163 real-world gaps sourced from reproducibility reports and GitHub issues, alongside 497 synthetic gaps injected into functional references.

The Performance Gap in Codification

Existing models struggle to determine exactly which part of a research specification is insufficient for implementation. When provided only with a specification, the best-performing model captures fewer than one in ten defects. However, performance shifts when the missing information is explicitly provided. When the benchmark gives the identified defect to the model, it achieves an 80.6% Macro Clarification Action Success Rate. The models demonstrate a strong ability to act on defined information, yet they consistently fail to recognize when their inputs are incomplete.

The Architectural Incentive to Hallucinate

If you are building tools to automate research-to-code pipelines, the bottleneck is the model's inability to detect ambiguity. The transformer architecture is optimized for next-token prediction, which treats every prompt as a mandate to complete a sequence. During training, the model is penalized for failing to predict the next token in the distribution, but it receives no explicit signal for flagging insufficient information. Consequently, the model is mathematically incentivized to fill gaps with statistically probable text rather than signaling an error. Because the attention mechanism must assign weights to existing text, it effectively forces the model to project its internal training distribution onto the gaps in the paper. Identifying a defect requires a meta-cognitive check that the model's objective function—which rewards plausible continuation—actively suppresses.

Consider an engineer building a tool to translate a machine learning paper into PyTorch. If the paper omits the specific normalization constant used in a custom layer, a robust system should identify this as a missing constraint. Instead, an LLM typically guesses a default value, such as a standard weight initialization, because that output is a common token sequence in its training corpus. It bridges the gap with a plausible, yet functionally incorrect, implementation because it lacks a built-in mechanism for uncertainty quantification. True uncertainty quantification would require the model to output a 'null' or 'unknown' token when probability mass is too dispersed across potential fillers. Current architectures make this difficult because they lack a dedicated verification head; they are designed to produce text, not to perform formal logical validation on the content they receive. In the IdeaAMBIG oracle study, supplying the model with the gold resolution raised the codification-ready rate from 14% to 98%, confirming that the failure is a result of performance bottlenecks in detecting missing context rather than a lack of coding syntax knowledge. What remains unknown is how to effectively fine-tune these models to prioritize 'I don't know' responses without degrading their ability to follow complex but well-specified instructions.