Fudan, Tencent Hunyuan and Tsinghua unveil ExplorationBench to test AI rule discovery
jiqizhixin · x · 2026-10-31
Researchers from Fudan University, Tencent Hunyuan, and Tsinghua University introduced ExplorationBench, testing whether AI systems can discover unfamiliar rules rather than recall memorized ones. The benchmark builds verifiable worlds where familiar programming and logic conventions are deliberately altered: 31 changed coding rules and 24 changed logic rules create 55 discovery targets, with 140 held-out tasks checking whether exploration yields transferable understanding. Models interact with feedback over four rounds with frozen weights, while separate evaluation copies prevent held-out answer leakage; ten models are tested across three independent exploration paths. On AlienCode, the best initial score is just 15.7%, with the strongest explored trajectory improving substantially—suggesting current models lean on memorization and that genuine exploration remains a bottleneck.
More from Research
- SCOPE workshop proposed for ICLR 2027 asks what counts as an AI scientific discovery — mariyaivasileva · 2026-10-31
- New MolmoSpaces-CommonSense benchmark: no robot policy tops 30% on manipulation tasks — RanjayKrishna · 2026-10-31
- Mischief launches video model trained from lab-grown brain cells at 0.5¢ per second — evilsocket · 2026-10-31
- Stanford CS336 notes: Percy Liang on scaling recipes and hyperparameter transfer — stanfordnlp · 2026-10-31
- Composable LoRA directions for texture, geometry and lighting make 3D fine-tuning cheaper — burkov · 2026-10-31
- Moondream releases RefCOCO-M with pixel-accurate masks and cleaned prompts for grounding evals — victormustar · 2026-10-31