Articulate but Wrong: Self-Review Failures in LLM-Based Code Modernization

• URL: http://arxiv.org/abs/2605.21537v1
• Date: Wed, 20 May 2026 05:00:31 GMT
• Title: Articulate but Wrong: Self-Review Failures in LLM-Based Code Modernization
• Abstract summary: Large language model (LLM) agents are increasingly used to migrate legacy code to modern stacks.<n>We run 1,980 real modernization calls across 11 production LLMs from 7 distinct families on a balanced 60-snippet legacy-Python-2 corpus.<n>We find that semantic-preservation drift is prevalent and sharply separable from a cleanly-controlled baseline.
• Score: 1.0164694825170502
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language model (LLM) agents are increasingly used to migrate legacy code to modern stacks. We ask a deceptively simple question: when an LLM modernizes legacy code, can the same model be relied upon to recognize when its own output silently changes observable behavior? We run 1,980 real modernization calls across 11 production LLMs from 7 distinct families on a balanced 60-snippet legacy-Python-2 corpus, evaluate every output with a type-strict behavioral oracle, and then ask each model to judge whether its own output preserves behavior. We report four findings. (1) Semantic-preservation drift is prevalent and sharply separable from a cleanly-controlled baseline: semantic-trap snippets drift in 39.7% of attempts versus 7.0% on benign-control code that requires no real modernization (+32.7 percentage points; n=660 each). (2) Drift concentrates on specific snippets that fail across models: pairwise model agreement on which snippets are hard is high (mean Pearson r=0.52), and a small core of numeric-semantics snippets fails for nearly every model and every prompt phrasing. (3) Self-review by the producing model is not a reliable safety net: across all semantic drift cases, 31.7% are silently endorsed by the same model that produced them (83/262), and the per-model self-miss rate is strongly bimodal -- ranging from 0% on five models to 100% on one widely deployed model -- with several models explicitly articulating the very Py2/Py3 semantic distinction that broke their output, then declaring behavior preserved. (4) Drift rate is non-monotone in model capability and price: per-model rates range 5.6%-46.7% and do not track model capability cleanly, indicating the failure is task-structural rather than driven by model scale. All code, prompts, the 60-snippet corpus, the behavioral oracle, the output extractor, and the raw model outputs are released.

Related papers

• ReDraft, Don't Just Distill: Reference-Driven Revision for Continual VLLM Post-Training [66.72708893922605]
  Post-training of large multimodal models should add new capabilities while preserving those from pre-training.<n>SFT gives explicit target supervision that learns a task from near-zero accuracy, but its off-policy targets move the model far enough to cause forgetting.<n>We introduce ReDraft, which obtains both from the model's own failures.
  arXiv  Detail & Related papers  (2026-09-15T04:59:03Z)
• SemVerBench: Benchmarking LLM Comprehension of Version-Constraint Resolution Semantics [15.233613140316791]
  SemVerBench is the first benchmark of version-constraint resolution semantics across three ecosystems.<n>We evaluate six frontier models and find no statistically significant self-favoritism.<n>Because the task is verifiable and a free, 100%-correct resolver exists, tool delegation reaches 100%.
  arXiv  Detail & Related papers  (2026-09-10T07:34:44Z)
• Solve the Loop: Attractor Models for Language and Reasoning [4.8720589853137435]
  Looped Transformers offer a promising alternative to purely feed-forward computation.<n>We introduce Attractor Models, in which a backbone module first proposes output embeddings, then an attractor module refines them by solving for the fixed point.<n>We show that Attractor Models outperform existing models across two regimes, large-scale language-model pretraining and reasoning with tiny models.
  arXiv  Detail & Related papers  (2026-05-12T17:51:26Z)
• When Correct Isn't Usable: Improving Structured Output Reliability in Small Language Models [2.064923532131528]
  Deployed language models must produce outputs that are both correct and format-compliant.<n>We study this structured-output reliability gap using two mathematical benchmarks -- GSM8K and MATH.<n>We develop AloLab, an iterative system-prompt requiring only black-box API access to the target model.
  arXiv  Detail & Related papers  (2026-05-04T09:07:44Z)
• The Surprising Universality of LLM Outputs: A Real-Time Verification Primitive [0.0]
  CPU-only scoring primitive runs at 2.6 microseconds per token.<n> token rank-frequency distributions converge to the same two- parameter Mandelbrot ranking distribution.<n>We derive a single-pass scoring primitive that composes with model log probabilities when available and degrades to a rank-only mode usable on closed APIs.
  arXiv  Detail & Related papers  (2026-04-28T13:35:31Z)
• Semantic Layers for Reliable LLM-Powered Data Analytics: A Paired Benchmark of Accuracy and Hallucination Across Three Frontier Models [0.0]
  LLMs deployed for natural-language querying of analytical databases suffer from two intertwined failures.<n>We benchmark three frontier LLMs on 100 natural-language questions over the Cleaned Contoso Retail dataset in ClickHouse.
  arXiv  Detail & Related papers  (2026-04-28T02:53:23Z)
• How Robustly do LLMs Understand Execution Semantics? [3.2717315277334706]
  LLMs demonstrate remarkable reasoning capabilities, yet whether they utilize internal world models or rely on sophisticated pattern matching remains open.<n>We study LLMs through the lens of robustness of their code understanding using a standard program-output prediction task.<n>Our findings both point to limitations in the way all models understand code, and establish the value of using perturbation to evaluate code models.
  arXiv  Detail & Related papers  (2026-02-24T19:07:25Z)
• AdapTrack: Constrained Decoding without Distorting LLM's Output Intent [53.75852524070165]
  Language model-based code generation and completion tools can sometimes produce code that does not meet necessary constraints.<n> Constrained decoding techniques are developed to help the model generate code adhering to the constraints.<n>AdapTrack avoids distorting the output intent of the model, thereby producing results that are not only constraint-compliant but also semantically aligned with model's output intent.
  arXiv  Detail & Related papers  (2025-10-20T10:11:34Z)
• Goedel-Prover-V2: Scaling Formal Theorem Proving with Scaffolded Data Synthesis and Self-Correction [95.91743732150233]
  Goedel-Prover-V2, a series of open-source language models, set a new state-of-the-art in automated theorem proving.<n>We generate synthetic tasks of increasing difficulty to train the model to master increasingly complex theorems.<n>Goedel-Prover-V2-32B achieves 88.1% on MiniF2F at pass@32 in standard mode and 90.4% in self-correction mode.
  arXiv  Detail & Related papers  (2025-08-05T16:28:22Z)
• Promises and Pitfalls of Generative Masked Language Modeling: Theoretical Framework and Practical Guidelines [74.42485647685272]
  We focus on Generative Masked Language Models (GMLMs) We train a model to fit conditional probabilities of the data distribution via masking, which are subsequently used as inputs to a Markov Chain to draw samples from the model. We adapt the T5 model for iteratively-refined parallel decoding, achieving 2-3x speedup in machine translation with minimal sacrifice in quality.
  arXiv  Detail & Related papers  (2024-07-22T18:00:00Z)
• The False Promise of Imitating Proprietary LLMs [158.65692029352584]
  An emerging method to cheaply improve a weaker language model is to finetune it on outputs from a stronger model. This approach looks to cheaply imitate the proprietary model's capabilities using a weaker open-source model. We first finetune a series of LMs that imitate ChatGPT using varying base model sizes. We then evaluate the models using crowd raters and canonical NLP benchmarks.
  arXiv  Detail & Related papers  (2023-05-25T05:00:12Z)
• TACRED Revisited: A Thorough Evaluation of the TACRED Relation Extraction Task [80.38130122127882]
  TACRED is one of the largest, most widely used crowdsourced datasets in Relation Extraction (RE) In this paper, we investigate the questions: Have we reached a performance ceiling or is there still room for improvement? We find that label errors account for 8% absolute F1 test error, and that more than 50% of the examples need to be relabeled.
  arXiv  Detail & Related papers  (2020-04-30T15:07:37Z)
• AvgOut: A Simple Output-Probability Measure to Eliminate Dull Responses [97.50616524350123]
  We build dialogue models that are dynamically aware of what utterances or tokens are dull without any feature-engineering. The first model, MinAvgOut, directly maximizes the diversity score through the output distributions of each batch. The second model, Label Fine-Tuning (LFT), prepends to the source sequence a label continuously scaled by the diversity score to control the diversity level. The third model, RL, adopts Reinforcement Learning and treats the diversity score as a reward signal.
  arXiv  Detail & Related papers  (2020-01-15T18:32:06Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.