Design

This experiment tests the real-dataset version of the exact-answer reward
gate from the draft empirical sketch proof. It uses a small GSM8K subset,
sampled model answers, integer parsing, and scalar exact-answer reward. There
is still no solid theory note, so this is an empirical plumbing and sparsity
check rather than a settled-theory experiment.

Results

20260801-145407-pretrained_llm_gsm8k_reward_smoke_jean_zay-0cabfb5

commit: 0cabfb5

metricvalue
both_success_rate0.5
decomposed_mean_accuracy_improvement0.02083333395421505
decomposed_mean_best_accuracy0.0416666679084301
decomposed_mean_final_accuracy0.02083333395421505
decomposed_mean_final_parseable_rate0.9791666865348816
decomposed_mean_finite_epoch_fraction1
decomposed_mean_parseable_rate_improvement0
decomposed_mean_shaped_epoch_fraction0.8125
decomposed_median_final_accuracy0.02083333395421505
decomposed_success_rate0.5
decomposed_win_rate0
eggroll_mean_accuracy_improvement0.02083333395421505
eggroll_mean_best_accuracy0.031250000931322575
eggroll_mean_final_accuracy0.02083333395421505
eggroll_mean_final_parseable_rate0.9791666865348816
eggroll_mean_finite_epoch_fraction1
eggroll_mean_parseable_rate_improvement0
eggroll_mean_shaped_epoch_fraction0.75
eggroll_median_final_accuracy0.02083333395421505
eggroll_success_rate0.5
epochs4
eval_samples24
mean_accuracy_gap_decomposed_minus_eggroll0
mean_improvement_gap_decomposed_minus_eggroll0
mean_parseable_gap_decomposed_minus_eggroll0
num_pairs4
num_ranks2
num_seeds2
num_target_modules7
population8
rank_1_decomposed_success_rate0.5
rank_1_decomposed_win_rate0
rank_1_eggroll_success_rate0.5
rank_1_mean_accuracy_gap_decomposed_minus_eggroll0
rank_2_decomposed_success_rate0.5
rank_2_decomposed_win_rate0
rank_2_eggroll_success_rate0.5
rank_2_mean_accuracy_gap_decomposed_minus_eggroll0

Interpretation

Run 20260801-145407-pretrained_llm_gsm8k_reward_smoke_jean_zay-0cabfb5 is a
useful plumbing check, not evidence for a decomposed advantage. Both methods
stayed finite and improved from the zero-accuracy baseline on the GSM8K subset,
with high parseability, but their final accuracy and improvement were tied.
The decomposed variant reached a slightly better best checkpoint during
training, while the paired win rate stayed flat, so the fair conclusion is that
the GSM8K reward path works and remains too small/noisy for a method comparison.

The boring alternative explanation is reward granularity. This subset is small,
the model is tiny, and exact GSM8K correctness moves in coarse jumps. The run
does not rule out sampling noise or a too-small reward budget.

The next useful experiment should keep the real GSM8K reward path but increase
statistical resolution: more evaluation examples, more seeds, and enough reward
budget to see whether either method consistently moves beyond parseability.