TL;DR of the paper, generated automatically.

SciGame is a benchmark for whether AI agents can rediscover scientific laws, with memorized answers filtered out.

• How it works: Papers are turned into "scientific cards", each yielding an Original environment that follows the published law and a Counterfactual twin with a mutated but dimensionally consistent law.
• Scope: A Historical collection of canonical laws and a Temporal collection from papers published after model knowledge cutoffs, tested on five frontier models under Prior-only, Fixed and Adaptive protocols.
• Findings: Familiar priors hurt once the governing relation is altered, and adaptive experimentation helps the strongest models but harms weaker ones under limited budgets. The main bottleneck is recovering the exact symbolic form, even when agents interpret the evidence correctly.