Design

This page merges the LLM-side smoke tests. These runs check whether plain EGGROLL and decomposed EGGROLL can patch causal-LM target modules, stay finite, and optimise generated synthetic objectives before moving to a real reward benchmark.

The sequence was:

  1. random tiny Llama-style causal-LM plumbing smoke;
  2. small pretrained causal-LM generated-arithmetic logprob smoke;
  3. small pretrained causal-LM seed/rank logprob robustness sweep.

None of these is yet a reward-only behavioural benchmark. The pretrained arithmetic objective still rewards answer-token likelihood, not sampled answer correctness.

Results

20260731-174147-tiny_llm_smoke_jean_zay-3a903d5

commit: 3a903d5

metricvalue
eggroll_initial_nll4.177872180938721
eggroll_final_nll4.177436351776123
decomposed_initial_nll4.177872180938721
decomposed_final_nll4.176585674285889
nll_gap_decomposed_minus_eggroll-0.000850677490234375
eggroll_finite_epoch_fraction1
decomposed_finite_epoch_fraction1
num_target_modules7

20260731-192758-pretrained_llm_logprob_smoke_jean_zay-ccf6982

commit: ccf6982

metricvalue
eggroll_initial_nll0.8726494312286377
eggroll_final_nll0.7472426891326904
eggroll_nll_improvement0.12540674209594727
decomposed_initial_nll0.8726505041122437
decomposed_final_nll0.6978709697723389
decomposed_nll_improvement0.17477953433990479
nll_gap_decomposed_minus_eggroll-0.04937171936035156
population16
eval_samples32
num_target_modules7

20260731-234003-pretrained_llm_robustness_sweep_jean_zay-43386a3

commit: 43386a3

metricvalue
num_seeds5
num_ranks2
num_pairs10
eggroll_success_rate1
decomposed_success_rate1
eggroll_mean_final_nll0.8384017646312714
decomposed_mean_final_nll0.8348180532455445
mean_nll_gap_decomposed_minus_eggroll-0.0035837113857269285
eggroll_mean_nll_improvement0.1068449079990387
decomposed_mean_nll_improvement0.11042933464050293
decomposed_win_rate0.6
rank_1_mean_nll_gap_decomposed_minus_eggroll0.0049618840217590336
rank_2_mean_nll_gap_decomposed_minus_eggroll-0.01212930679321289

Interpretation

The LLM tests cleared the intended engineering gates. Both methods ran through causal-LM target-module patching, stayed finite, and improved synthetic NLL objectives. The decomposed variant was slightly ahead in the random-LM smoke, clearly ahead in the single pretrained logprob smoke, and mildly ahead overall in the pretrained seed/rank sweep.

The sweep is deliberately interpreted conservatively. Both methods had success rate 1; decomposed EGGROLL won 0.6 of paired comparisons, with the advantage concentrated at rank 2 and not present at rank 1. This is encouraging stability evidence, not a strong empirical win.

The next experiment should stop using answer-token likelihood as the reward. A tiny GSM8K-style or generated-arithmetic exact-answer benchmark should sample model outputs and give ES only scalar correctness or minimally shaped parseability/correctness feedback.