Conditions for Unguided Reachability of Guidance-Like Computation

Investigating whether unguided rollouts from a base language model can reach states similar to those induced by oracle-provided guidance prefixes (POPE, Qu et al. 2026)

0.287
R-squared for Exponential Decay Fit
0.0108
Decay Constant (alpha)
0.559
Min Curriculum Stage Hit Rate
0.3045
Max Spectral Gap (tau=10.0)

Experiment 1: Reachability vs. Guidance Length

Experiment 1b: Log Hit Rate vs. Effective Information

Experiment 2: Spectral Gap vs. Temperature

Experiment 3: Curriculum Hit Rates by Stage

Experiment 3b: Per-Step Information Profile

Experiment 4: Spectral Characterization

Experiment 5: Hit Rate vs. Effective Information (Parameter Sweep)

Key Findings