SHELDON: Reasoning through Dual Search over Hypothesis and Observation Spaces
SHELDON searches for unexplained observations separately from the hypotheses it already has.
The paper argues that tying observation-gathering to current hypotheses keeps systems near local refinements. SHELDON instead explores hypothesis and observation spaces together, using observations that existing hypotheses cannot explain to drive new investigations. On diagnosis, abstract reasoning, and optimization benchmarks, it beats evaluated baselines by up to 26.7 points and produces up to 4.5x more novel solutions. HF Daily Papers' note
The paper argues that tying observation-gathering to current hypotheses keeps systems near local refinements. SHELDON instead explores hypothesis and observation spaces together, using observations that existing hypotheses cannot explain to drive new investigations. On diagnosis, abstract reasoning, and optimization benchmarks, it beats evaluated baselines by up to 26.7 points and produces up to 4.5x more novel solutions. HF Daily Papers' note
score 5