The shared task is divided into two complementary subtasks, each targeting a different aspect of the title provenance problem.
Find It
Original Title Retrieval & Ranking
Participants are given a reformulated title and must generate a ranked list of up to ten candidate titles ordered by their likelihood of being the true source title. Systems are judged on how high they place the correct original.
Fix It
Original Title Reconstruction
Participants receive a reformulated title together with its stylistic cue (Technical, Accessible, or Catchy). The goal is to reconstruct the original research title — evaluating a system's ability to recover underlying source content despite LLM-introduced stylistic transformations.
Input & Output Format
- Input (Subtask 1): A reformulated title (no style cue provided)
- Output (Subtask 1): Ranked list of up to 10 candidate original titles
- Input (Subtask 2): A reformulated title + its style label (Technical / Accessible / Catchy)
- Output (Subtask 2): A reconstructed original title
Disclaimer
To ensure fairness and reproducibility, participants are not permitted to use proprietary frontier language models, including but not limited to GPT-5, Claude Sonnet-5, and Gemini 3 series models, for system development, inference, or result generation. Participants must submit the complete source code, scripts, model configurations, and relevant documentation along with their result submission. The final rank list will be released only after successful verification of the submitted code. Failure to provide the required materials may result in disqualification.