Autonomous Research & Recursive Self‑Improvement

The question is no longer speculative. Research agents now synthesize literature, propose hypotheses that can be tested in laboratories, and run closed research loops: proposing an idea, implementing it, running an experiment, evaluating the result, and choosing the next attempt.

Self-improvement also reaches training: models generate practice data, rewards, curricula, and critiques through self-play, self-reward, and verifiable task generation, and agentic data systems improve the datasets they create. This broadens AI-for-AI to data curation, pretraining, midtraining, post-training, and systems optimization. Yet full recursive self-improvement remains unproven; evaluation must expose overfitting, hidden human labor, and reward tampering.

ARRSI fits AACL-IJCNLP because language is how research agents form hypotheses, retrieve multilingual evidence, call tools, critique claims, and explain results. The workshop brings together NLP, agentic systems, AI for Science, AI for AI, evaluation, and governance to establish definitions, benchmarks, reporting standards, and safety boundaries for systems that do not merely produce research-like text, but accumulate auditable evidence and use it to become better researchers.

Submission opens on OpenReview together with the call for papers.

Long papers run up to 8 pages and short, position, or resource papers up to 4 pages, excluding references, in *ACL format. Every accepted paper presents a poster, and 4 to 6 receive talk slots. Watch this page or the GitHub organization for the announcement.