IOL-AI 2026 — Qwen3-14B-AWQ bounded thinking
Controlled follow-up to the preserved rank1-0.1470 non-thinking submission.
The script first generates the exact greedy, non-thinking baseline for every
problem and atomically checkpoints a complete submission. It then attempts one
bounded Qwen3 thinking generation per problem using the official sampling
settings (temperature=0.6, top_p=0.95, top_k=20) and a stable per-row
seed. A baseline answer is replaced only when generation:
- ends naturally at EOS;
- contains Qwen3's
</think> token (151668);
- has a non-empty final suffix after the last close token; and
- returns the same number of answer lines as the baseline.
Incomplete, timed-out, malformed, or failed thoughts retain the proven direct
answer. Each thinking call is capped at 1,536 tokens and at most 120 seconds,
with a global soft deadline 150 seconds below the evaluator's 30-minute limit.
The repository remains fully offline and reuses the pinned local model and
the bundled Transformers 4.51.3 / AutoAWQ 0.2.9 runtime.
