Milos Gligoric

Research profile

Research summary

Milos Gligoric is a researcher at The University of Texas at Austin whose work sits at the intersection of software engineering, programming languages, and large language models for code. The OpenAlex-resolved record spans 2008–2026 and lists an h-index of 28 with an i10-index of 58, alongside a sustained publication cadence that peaks at 14 works in 2022 and again in 2026; the last three years account for 28 works. Topic clusters are led by software testing and debugging techniques (94 records) and software engineering research (82), with logic, programming, and type systems (55) and parallel computing and optimization techniques (27) also represented. Recent field attribution splits evenly between NLP/software engineering for LLMs and code and LLM-based code generation. Evolve has reviewed three 2026 arXiv papers: retrofitting Java code with exception behavior learned from tests (consistent gains across 5 models), CROCODIL's GRPO-based reward that halves excessive edits by the Olmo3 7B model, and the PrEx benchmark probing whether LLMs use supplied operational semantics rather than priors. Two of the three carry high reproducibility readiness, supporting a reproducibility mean of 78 across two scored papers. Radar scores flag very high activity (98) and strong reproducibility (78), moderate collaboration breadth (39, with 6 distinct recent co-authors), and a focus score recorded as 0.

Research keywords

Highlights

• Retrofitting Code with LLM-Supported Exceptional Behavior

  This 2026 study uses tests to retrofit exception logic into existing code, reporting consistent LLM gains across 5 models — while noting that roughly 25% of test-passing code still differs from intent.

• CROCODIL: Curbing Excessive Cross-Model Code Edits

  LLMs tend to over-edit code they did not write; CROCODIL's GRPO-based reward halves edits by Olmo3 7B while preserving success, though gains lack statistical tests and results come from a single base model.

• PrEx: Do LLMs Actually Use Supplied Semantics?

  The 2026 PrEx benchmark pairs LLMs with programming language semantics to predict program exit codes; open-source models answer from priors, with sharp drops under semantic shifts — evaluated with means-only statistics and no heuristic baseline.

Timeline

1. 2008First indexed works appear (2 records), opening a continuous publication record that runs through 2026.
2. 2018Output accelerates: 10 works in 2018, rising to 13 in each of 2019 and 2020.
3. 2022A peak year with 14 indexed works, the highest annual count in the 2008–2026 record.
4. 2026Three Evolve-reviewed arXiv papers appear on LLM code retrofitting, cross-model editing, and exit-code prediction; 14 works are already indexed for the year.

Papers in the Evolve catalog

• 2026-09-09

  Retrofitting Code Using LLMs to Support Exceptional Behavior

• 2026-09-03

  CROCODIL: Cross-Model Code Editing with LLMs

• 2026-09-01

  Predicting Program Exit Code with LLMs and Programming Language Semantics

Frequent co-authors