LangCode 2027 — The First Language and Code Workshop

2 shared tasks

100+ languages in Task 1

~250K prompts in Task 2

25 committee members

1 day, hybrid

##About

Research at the intersection of natural language and code is among the most active areas in NLP, yet the ACL community has no dedicated workshop on the topic.

Code models are prompted overwhelmingly in English, evaluated almost exclusively on Python and a few high-resource programming languages, and guarded by safety mechanisms rarely tested beyond English inputs. LangCode treats linguistic variation and safety as intertwined: guardrails calibrated on English degrade sharply on other languages, so a code model cannot be safe for all users unless it is robust for all languages.

AXIS I

Natural languages

Generation, explanation, and documentation from prompts in any language.

AXIS II

Programming languages

Low-resource, emerging, legacy, and domain-specific languages.

AXIS III

Adversaries

Secure generation, guardrails, and jailbreak and prompt-injection defenses.

##Call for Papers

We invite submissions across five areas, including but not limited to:

01Secure Code Generation

Vulnerability-aware training and decoding
Guardrails: input filtering, output auditing, policy enforcement
Adversarial attacks: jailbreaks, prompt injection, obfuscation
Safety of agentic coding systems

02Trustworthiness

Explainability and interpretability of code models
Calibration, uncertainty, and hallucination
Contamination-aware benchmarks and evaluation
Human–CodeLLM and agent alignment

03Robustness

Paraphrasing, dialectal variation, and code-mixed prompts
Low-resource, emerging, legacy, and domain-specific languages
Adaptation to new languages and distribution shift

04Multilinguality

Generation, explanation, and documentation from non-English prompts
Cross-lingual transfer to new natural languages
Data curation for underrepresented languages

05Applications

Efficient code models for low-compute settings
Deploying code models across languages and communities
CodeLLMs in computing education

Reviewing: OpenReview, with an option to commit pre-reviewed ARR papers.

Format: One day, in person, hybrid. Authors who cannot travel can present remotely.

Venue: TBD. Announced with the acceptance notification.

##Shared Tasks

Two tasks on CodaBench, with public baselines and starter kits. Participation is free, and constrained tracks keep low-compute teams competitive.

Task 1 / Multilingual Code Generation

100+ natural languages execution-based Pass@1 macro-averaged across tiers

Generate Python solutions to problems presented in more than 100 natural languages across resource tiers. Building on mHumanEval, the task uses newly authored held-out problems to limit contamination. Systems are ranked by Pass@1 on hidden tests, macro-averaged across tiers so that high-resource performance alone cannot win.

CodaBench page: coming soon

Task 2 / Adversarial Prompt Detection

18 languages ~250K prompts macro-F1 surprise-language leaderboard

Classify CodeLLM prompts as benign or adversarial, with a subtrack for attack type. The task builds on MALICE (Findings of EMNLP 2026), a ~250K-prompt benchmark spanning 18 languages with code-mixed and transliterated variants. Ranked by macro-F1, with a cross-lingual leaderboard on surprise languages.

CodaBench page: coming soon