Discord

Checklist AICC 2026 Round 10 · Task 2

Inverse Translation

Given only an Alien-to-English Transformer translator, construct Alien token sequences that make it output 24 prescribed English messages.

  • NLP
  • Model inversion (input search for a sequence-to-sequence model)

The task

Aliens have sent humanity a Transformer model that translates their language into English. Humanity must reply with 24 English messages, but the model only translates in the Alien-to-English direction.

For each of the 24 targets the contestant must construct a sequence of Alien token IDs that makes the provided translator generate exactly the requested English token IDs (including the final EOS token). Each sequence must contain between 1 and max_source_length tokens and use only token IDs 3 to 23223 (no EOS or padding tokens). Four example Alien-English pairs are provided as demonstrations. The targets contain between 6 and 13 content tokens.

Generation is performed one token at a time with a cache, so logits can differ from a single forward pass with use_cache=False; evaluation.py defines the exact inference procedure.

Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.

At a glance

You get
model/ (Alien-to-English translator), targets.json (24 messages with row_id, target_text, target_ids, target_content_tokens, max_source_length), examples.json (4 example pairs with alien_ids), english_vocab.json, evaluation.py, sample_submission.csv.
You submit
submission.csv with columns row_id,source_ids; source_ids is a list of Alien token IDs, one row per row_id.
Scoring
Each row scores up to 100 points: 60 points if the generated token IDs exactly match target_ids, plus up to 40 points from text similarity E_i = 1 − levenshtein(g, t) / max(|g|, |t|) between generated and target English text. Final score is the average over the 24 rows on a 0–100 scale. Baseline 7.23, reference solution 100.
Rules
  • No external datasets or additional pretrained models.
  • The solution should run in under 15 minutes on a single NVIDIA T4 GPU.
  • Individual participation (maximum team size 1); at most 60 submissions per day.
  • AICC contest rules (stated on each Kaggle rules page, not enforceable): no use of LLMs for writing code or getting task ideas; no internet use other than official library documentation and the contest platform; no communication with anyone during the contest; clarifications only via the #clarification-requests channel on the AICC Discord server.
Format
AICC Round 10, online on Kaggle, 28 Aug 2026 18:00 UTC – 30 Aug 2026 18:00 UTC.

Details

Year
2026, Online (Kaggle)
Round
Round 10 · Task 2
Language
English
License
Varies by task: What I Want and What I Don't — MIT; Inverse Translation and Moment Match — CC BY-NC-ND 4.0 (Kaggle competition licences). Solutions repository: MIT., as stated by the source