State-of-the-art models are far more proficient in scientific coding than SciCode has suggested---the bottleneck was not model capability, but the quality of the evaluation instrument.
Abstract
SciCode is the standard measure of the scientific-coding ability of language models: research-level problems that demand both frontier scientific theory and its implementation as working numerical code. It is a component of the Artificial Analysis Intelligence Index and a standing evaluation in government and national-laboratory suites. Yet its scores have recently plateaued: the strongest 2026 models cluster tightly around 60\% subproblem accuracy, and a successor model ties its predecessor. We trace this stagnation to defects in the benchmark itself. A per-problem, domain-expert audit of all 65 test problems uncovers 263 defects; 192 of them, spread across 91\% of the main problems, cause correct, instruction-following solutions to be wrongly rejected---through non-reproducible gold answers, over-tight tolerances, or self-contradictory specifications. Critically, 78\% of these score-suppressing defects require specialized physics or mathematics knowledge to detect, not mere clerical proofreading. We corrected every confirmable defect to produce SciCode-Verified. The corrections add only the specifications a well-posed problem requires, repair grading, and tighten the tests that were too lenient; every change is recorded with its justification and independently re-checked by a second domain expert. We re-evaluate twelve frontier model snapshots on the corrected benchmark and find a substantial recovery: subproblem accuracy rises from 45--60\% to 84--98\%, and main-problem accuracy from 9--27\% to 69--92\%. State-of-the-art models are far more proficient in scientific coding than SciCode has suggested---the bottleneck was not model capability, but the quality of the evaluation instrument. We release SciCode-Verified with its complete audit trail as the corrected public standard.
Evaluating Large Language Model (LLM) code generation quality requires examining not just whether the generated code is correct, but whether it is maintainable, efficient, and stylistically sound, all of which are qualities of direct importance to software engineering practitioners. Existing benchmarks reduce evaluation to a single Pass@k metric, which obscures critical trade-offs between functional correctness and structural quality. A further limitation is the near-exclusive focus on Python, leaving enterprise-relevant ecosystems such as C# and .NET without dedicated evaluation. This paper presents an automated, multi-dimensional evaluation framework for C# code generation, applying it to four state-of-the-art LLMs: GPT, Gemini, Claude, and Grok. We conduct a controlled experiment across 85 algorithmic tasks derived from HumanEval, generating and evaluating 340 solutions in total, in which each solution is assessed across three independent dimensions: functional correctness via automated unit testing, static code quality via Roslyn AST analysis, and runtime efficiency via adversarial BenchmarkDotNet profiling. Our central finding reveals a substantial gap between correctness and quality attributes (Pearson r = 0.075), demonstrating that Pass@k rankings systematically misrepresent the full LLM performance profile in software engineering contexts. We further characterize GPT's bimodal failure behavior.
Seyed Mohammad Mahdi Ghalandarian, Majid Bazargani, Masoumeh Taromirad· 0 citations
Code generated by modern language models often reads naturally. Yet, it also often fails to implement what was asked. This should be no surprise, as research shows the models'own confidence signals are poorly calibrated with actual correctness. A promising way to assess correctness looks inside the model: by contrasting the hidden states of correct and incorrect programs, recent work captured an internal signal of code correctness that is able to judge candidate solutions better than the model's token-level or stated confidence, with no test execution. However, this signal was captured under one particular way, leaving open an important question: whether it reflects a robust property of the model or an artifact of that choice. We study this question systematically, varying how the signal is extracted from the model internals. Besides this, we also ask if the signal's quality is limited by the data used to extract it, by constructing program pairs that differ only in the fault that makes them incorrect. Our results show that no single configuration is best, and that isolating the fault does not help.
Francisco Ribeiro, Sohaila Abdulsattar, Renata Gonzalez et al.· 0 citations
These findings show that reliable evaluation of LLM-generated code requires validated ground truth, protected tests, and multiple explicitly interpreted measures, and that CodeAssay provides a reproducible basis for evidence-based model evaluation in AI-augmented software development.
Shahbaz Siddeeq, Muhammad Waseem, Umar Subhan Malhi et al.· 0 citations
Large language models (LLMs) are increasingly applied to reverse-engineering tasks, and recent threat-intelligence reporting shows them operating inside live offensive-security workflows. Claims about their capability, however, outpace our ability to measure it. Existing benchmarks for LLM-assisted binary analysis treat the construction of function-level ground truth as a solved pre-processing step and report accuracy without disclosing how many functions were reliably evaluable. We argue that the principal obstacle to fair evaluation is not model capability but the reliability of binary-to-source alignment under compiler optimization. This paper presents Reforge, a provenance-tracked pipeline that constructs function-level ground truth from C source through compilation, DWARF and syntactic extraction, alignment, and decompilation, and that operationalizes alignment uncertainty as an eight-gate confidence funnel with three-tier stratification. On a controlled micro-benchmark, high-confidence yield falls from 87.2% to 65.9% across optimization levels, and unpaired comparisons overstate optimization-induced performance decay through survivorship bias. A proof-of-concept evaluation of seven contemporary LLMs on function naming demonstrates the validity of the concept and generally motivates an uncertainty-aware benchmarking practice.
Vibe coding - AI-assisted software development in which the developer describes intent in natural language and validates results by running rather than reading the generated code - was named by Andrej Karpathy in February 2025 and produced its first body of empirical evidence within seventeen months. This state-of-the-art review assembles that evidence across a cross-disciplinary corpus spanning software engineering, human-computer interaction, labour economics, security research, governance, and education. We survey the model landscape, the tool ecosystem, and the performance record by task type, finding the early benchmarks saturated but task-level capability uneven: reliable code generation alongside weak fault detection and hard-to-audit documentation. The productivity record is at first contradictory: peer-reviewed field experiments report +26% more tasks per week, independent randomised trials measure a 19% slowdown, and team-level telemetry shows code-review time up +441%. We argue these readings are consistent once measurement method, scope, and time horizon are held constant, and identify six patterns behind the dispersion, among them effect-shrinkage under broader measurement, self-report diverging from independent measurement, output volume conflated with productivity, and bold claims walked back once tested over longer horizons. We further document security failures in deployed applications, code-quality degradation visible in large-scale code and developer telemetry, unsettled copyright exposure, and evidence of skill atrophy. The review closes with the open research questions and one falsifiable conjecture: that the gains are real on new code and shrink or reverse on mature codebases, which would account for most of the disagreement in the record.
D. Michels, Mutaz Abu Ghazaleh, Francois Lazzari et al.· 0 citations
Large Language Models (LLMs) are predominantly assessed based on their common sense reasoning, language comprehension, and logical reasoning abilities. While models trained in specialized domains like mathematics or coding have demonstrated remarkable advancements in logical reasoning, there remains a significant gap in evaluating their code generation capabilities. Existing benchmark datasets fall short in pinpointing specific strengths and weaknesses, impeding targeted enhancements in models’ reasoning abilities to synthesize code.
To bridge this gap, this thesis introduces two novel contributions: CodeEval and CodeQual. CodeEval is an innovative, pedagogical benchmarking method that mirrors the evaluation processes encountered in academic programming courses. It comprises a multi-dimensional benchmark dataset of 602 hand-crafted problems designed to rigorously evaluate LLMs across 24 distinct aspects of Python programming, covering three proficiency levels—beginner, intermediate, and advanced—and includes both class-based and function-based problem types with detailed problem specifications and comprehensive test suites achieving 99.1% coverage. To facilitate widespread adoption, we developed RunCodeEval, an open-source execution framework that provides researchers with a ready-to-use evaluation pipeline. Our evaluation of 15 state-of-the-art LLMs revealed consistent performance degradation with increasing complexity (validated statistically, Cohen’s d = 0.790) and universal struggles with advanced concepts like concurrency.
Code quality is inherently subjective, encompassing dimensions like readability, efficiency, and adherence to language idioms that traditional static metrics fail to capture adequately. While large language models can assess these subjective qualities, lightweight models offer practical advantages: seamless CI/CD pipeline integration, lower operational costs, and full control over model behavior. We investigate whether such models can learn to assess code quality by training on synthetic LLM annotations. We introduce CodeQual, a dataset of 5,819 code samples derived from five established sources spanning diverse domains—competitive programming, pedagogical problems, software engineering, and general benchmarks—scored by LLMs across five quality dimensions, with 655 human-annotated samples for evaluation. Our fine-tuned model, CodeQualBERT, not only matches LLM performance but exceeds inter-human agreement on all five dimensions, achieving 16–100% improvement over the inter-human agreement baseline.
Together, these contributions provide a comprehensive framework for evaluating and improving LLMs in software engineering contexts, encompassing both functional correctness assessment and subjective code quality evaluation.