writing2026-09-12

KRAFTON AI R&D Hackathon

writingbloghackathonkraftontransformerlpnnumba

Summary

Published 2026-03-29 · slug krafton-ai-rnd-hackathon-2026-preliminary · category "Projects, hackathons, and competitions / ML and data competitions" · source article recorded in Korean and English · related slugs: equilaw-gemini3-hackathon-retro, lg-aimers-5th-comprehensive-retro, softbank-hackathon-2025-cicd-pipeline.
The article describes tasks and solutions from the two-day online preliminary round, March 28–29, 2026, and why the submission missed the final-round cutoff. The poster listed an application deadline of March 27 at noon, online preliminaries on March 28–29, and an in-person final on April 4–5. This was external participation before joining KRAFTON.
Day 1, MultiplierBoard: predict the 12-bit product of two six-bit numbers autoregressively under the organizers' training settings, including AdamW and 200 epochs with a cosine schedule. Minimize trainable parameter count P2 while attaining validation accuracy Acc2 of at least 99%, and separately submit theoretical lower bound P1. Problems included exposure bias from a zero-padding shortcut and vanishing gradients from threshold activations, with accuracy stuck around 3%. The solution combined input alignment, weight tying, learned positional encodings, and reduced capacity (d_model=16, nhead=2, two layers, FFN 64). Submitted figures: P1 36, P2 6,978, Acc2 99.98%.
Day 2, SparseTap: recover LFSR/parity (LPN) structure from noisy sequences. After continuous relaxation with SIREN and FCA-style brute-force attempts, I used ISD, bit packing, and Numba. Logs showed convergence at approximately attempt 3,612,183, about 95 seconds on one core with Numba, and approximately 79.99% validation accuracy against a theoretical ceiling of 80%.
Retrospective on missing the final: selection considered more than code, including the report, P1 explanation, and tie-breakers; time spent on SIREN on Day 2; an undisclosed overall scoring formula; and reproducibility, which needed multiple-seed verification and a stated worst-case runtime. The next plan was to secure a deterministic solution first and write the report alongside the code.

Mentioned

KRAFTON: the hackathon organizer.
KRAFTON AI R&D Hackathon 2026: the participation record discussed in the article, also recorded as a project note.

Connections

Writing · Projects
Atomic notes: Selection requires more than a working implementation; secure the deterministic solution first and write the report alongside the code · Deliver research as a repository, technical report, and quantitative results table
Concepts: AWS Lambda · Gemini API · LPN and ISD · SMOTE-ENN · Stacking ensembles · Transformer · i18n and multilingual support · Running a technical blog · Random-seed effects · Ensembles · Reproducibility audits · Reproducibility verification · Statistical inference · Isolated and offline environments

Local graph

Notes 4 · Links 3

Backlinks

Google DeepMind Gemini3 Hackathonwriting

Source posts

KRAFTON AI R&D Hackathon Prelims: Teaching Multiplication with 7,000 Parameters, Recovering a Rule from Noise