International Workshop on Large Language Models for Code (LLM4Code) 2026 - Rio de Janeiro, Brazil (Nathan Rutherford)

 By Nathan Rutherford


In April I had the pleasure of travelling to Rio de Janeiro in Brazil to present our paper, **[An Empirical Study of C to Rust Translation using Local Large-Language Models](https://pure.royalholloway.ac.uk/files/71513373/llm4code26.pdf)**, at the 3rd International Workshop on Large Language Models for Code (LLM4Code) 2026, co-located with the International Conference on Software Engineering (ICSE).

The aim of the workshop was to share ideas and results related to the application of large language models (LLMs) to code-related problems within software engineering, while also enabling discussion around emerging challenges in the area. The workshop covered topics including benchmark development for evaluating LLMs, dataset creation, and code generation, with a particularly strong focus on agentic AI systems. The keynote talks provided useful insights into how both academia and industry are leveraging these systems, from improving artefact evaluation for researchers to reducing the cost of code generation through prompt engineering techniques such as “Dynamic Turn Reminder” in TRAE agents.

At the workshop, we presented results from an empirical evaluation of C-to-Rust translation over the Transcoder-IR testing dataset using our [Oxidation pipeline](https://doi.org/10.5281/zenodo.18341646) with local LLMs (models run locally rather than through cloud-based services like ChatGPT). Our experiments showed that leveraging code summarisation could help improve translation quality for computationally cheaper local models, narrowing the gap with more advanced reasoning-capable systems. The discussion following the talk also highlighted interest in how the technique could be adapted to produce different forms of code summarisation for translation tasks.

Beyond our own work, a recurring theme across the workshops and sessions I attended was the growing interest in agentic systems for software engineering. Broadly speaking, these systems are designed around multiple LLMs being assigned specialised tasks and collaborating to solve larger problems. For example, one AI agent might analyse source code and produce an implementation plan, while another agent uses that plan to generate or refactor code.

Two of the major research challenges currently attracting attention in this area are:

1. **Context management** — how can we represent problems succinctly and effectively for AI agents without incurring significant computational or monetary costs?

2. **The verification bottleneck** — how can we efficiently verify that AI-generated code is correct, safe, and functionally equivalent at scale?

One of the benefits of attending a conference like ICSE is the opportunity to hear a wide range of perspectives from both academia and industry research labs tackling these challenges in different ways. Some of my personal highlights included:

- [Automated Software Test Generation at Industry Scale Using a Multi-Agent Architecture and Workflow Integration](https://homes.cs.washington.edu/~rjust/publ/auto_cover_icse_2026.pdf) by Rastenis et al.

- [EvoC2Rust: A Skeleton-guided Framework for Project-Level C-to-Rust Translation](https://arxiv.org/abs/2508.04295) by Wang et al.

- [HoarePrompt: Structural Reasoning About Program Correctness in Natural Language](https://arxiv.org/abs/2503.19599) by Bouras et al.

- [SmartC2Rust: Iterative, Feedback-Driven C-to-Rust Translation via Large Language Models for Safety and Equivalence](https://arxiv.org/abs/2409.10506) by Shiraishi et al.

 

Outside of the paper sessions, the conference also provided a great opportunity to meet other researchers working on software engineering, code refactoring, and AI-assisted development while networking over pizza. I also had the opportunity to visit one of the seven wonders of the world, Christ the Redeemer, and take in some of the incredible views of Rio de Janeiro (including witnessing a marriage proposal overlooking the city).

Overall, attending LLM4Code and ICSE was an extremely positive experience. The conference highlighted just how rapidly research into AI-assisted software engineering is evolving, particularly around agentic systems. Presenting our work at LLM4Code and engaging with researchers across academia and industry was both valuable and motivating for future research.

Next year, ICSE will be held in Dublin, Ireland, much closer to home for those interested in seeing the current state of the art in software engineering research!





Comments

Post a Comment

Popular posts from this blog

Remote working and Cyber Security: Georgia Crossland and Amy Ertan

Introducing the project The shift to remote working represents a huge shift for employees and organisations alike, introducing significant implications for cyber security. Amy Ertan and Georgia Crossland are currently working on a research project, in collaboration with the Research Institute for Sociotechnical Cyber Security (RISCS) and funded by the UK National Cyber Security Centre (NCSC), on the impact of remote working on cyber security and wellbeing. This blog post will give an overview of a section of this project, how remote working impacts cyber security. In response to COVID-19, in March 2020, organisations across the world made the rapid shift away from physical office working to remote working, wherever possible. This shift has had a significant impact across many aspects of organisations. This blog post will offer a high-level overview of how the cyber threat landscape was impacted for many organisations and may further change organisational cyber resilience. Remote workin...
Read more

New Publication: Remote Working and (In)Security?: Amy Ertan

Image
A recent research report by CDT researchers Amy Ertan and Georgia Crossland has been published with the Research Institute for Sociotechnical Cyber Security (RISCS). The report examines ‘The Impact of Pandemic-Driven Remote Working on Employee Wellbeing, the Psychological Contract and Cyber Security’, with research carried out through a series of interviews with senior cyber security colleagues between January and April 2021. The full report can be found on the RISCS website: New publication: Remote Working and (In)Security | RISCS. Context In the UK and around the world, the impact of COVID-19 represented a total shift away from what was considered ‘normal living’. One aspect of this change was the rapid transition to remote working as people were encouraged to work from home and limit contact with others as far as possible through repeated lockdown conditions. Fellow CDT colleague Georgia Crossland and I were curious to explore how remote working has impacted the experiences - and se...
Read more

The Artificial Intelligence Monster: Nicola Bates

Image
Christopher Booker in 2004 wrote of the seven basic plots to any story, one of these being ‘overcoming the monster’. Within Hollywood, such ‘monsters’ often reflects the zeitgeist balancing commercial requirements with the political and societal preoccupations of the age. The monster evolves over time - from Russia in the Cold War years through to lone actors and terrorists and now to seemingly shadowy figures controlling intelligence. Whilst generally lighter in tone, such monsters also exist within children’s cinema. In recent years this monster has become increasingly defined by their use of highly sophisticated technology, finally becoming the technology in ‘The Mitchells vs. the Machine’*. Lethal technology and subversive machines aren’t new concepts in children’s movies. They can, for example, be found in ‘Short Circuit’ back in 1986. This movie begins with five United States military robots in a training exercise against tanks and focusses upon one of the robots becoming ‘alive’...
Read more