Datasets:

Papajams
/

beam-audit-bench

Tasks:

Text Generation

Size:

ArXiv:

Tags:

reflection-beam

agentic-science

reproducibility

License:

Dataset card Files Files and versions

beam-audit-bench

1.72 MB

Ctrl+K

Ctrl+K

1 contributor

History: 11 commits

Papajams's picture

figures fix: episode ladder was missing the Kimi K3 row (provider field mismatch) — regenerated with all four models

63161cf verified 6 days ago

figures
figures fix: episode ladder was missing the Kimi K3 row (provider field mismatch) — regenerated with all four models 6 days ago
model_runs
v2: add RL environment + rollouts, Beam self-improvement round (C1 flip, verify_gain-flagged), improved README 6 days ago
references
Beam Audit Bench v1: 13 claims, 2 papers, Beam clean rounds + recorded 27B/Kimi comparison 6 days ago
rl_env
v7: closed-frontier comparison — Claude Opus 5.5, 8 episodes, reward 1.0 in 1/8 (keyed C1: 0/4). Failure analysis in README: contract failure (physics computed, metrics reported outside the graded alias list), not science failure. Opus recovered the harder ca_c2 D-ball series once from the excerpt alone. 6 days ago
tasks
Beam Audit Bench v1: 13 claims, 2 papers, Beam clean rounds + recorded 27B/Kimi comparison 6 days ago
.gitattributes

2.5 kB
initial commit 6 days ago
README.md

15.2 kB
v8: presentation — stats strip, three data-driven dark-mode figures embedded in README, figures/ shipped in the dataset 6 days ago
bakeoff_results.json

1.07 kB
Beam Audit Bench v1: 13 claims, 2 papers, Beam clean rounds + recorded 27B/Kimi comparison 6 days ago
grader.py

47.3 kB
v8: presentation — stats strip, three data-driven dark-mode figures embedded in README, figures/ shipped in the dataset 6 days ago
scores.json

94.9 kB
v2: add RL environment + rollouts, Beam self-improvement round (C1 flip, verify_gain-flagged), improved README 6 days ago