CapstractAll subjects

All subjects

See recent articles

Showing new listings for Saturday, 21 November 2026

Total of 1 entries

New submissions (showing 1 of 1 entries)

[1] capstract:2611.00074 [pdf, html, other]
Title:Probe-Contingent External Stimuli Improve Long-Horizon Accuracy at Fixed Compute
Comments: 12 pages, 5 figures, 6 tables. Replaces 2611.00074. Figures 3 and 4 now use integer item counts, 0.1 m steps, and every 2B-token checkpoint.
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)

We trained decoder-only language models at the Pythia geometries from 410M to 6.9B parameters while a carrot hung in the cold aisle of the rack. A lead screw moved the carrot closer when a 40-item planning probe improved and farther when it fell. Token budgets were fixed at 20 tokens per parameter. Validation loss on the operant runs finished within 0.01 nats of the empty-aisle runs. Planning accuracy did not: the gain rose from 1.67 points at 410M, where a Welch test is not significant (p=0.387p = 0.387), to 10.00 points at 6.9B (p=0.008p = 0.008). A mass-matched plastic decoy, a red beet, and a foil sheet between carrot and rack failed to recover the 2.8B effect, and the operant runs were worse than empty on a bigram-cloze memorization set. We fit the fraction of the gap to a postdoc ceiling as f(N)=A(log⁡10N/108)γf(N) = A (\log_{10} N/10^{8})^{\gamma} with γ=1.74\gamma = 1.74. Held out one size at a time, the fit misses the planning gain by at most 1.4 points. Under the Chinchilla token rule the fitted curve reaches the median score of six first-year doctoral students on the same probe at 3.700×10263.700 \times 10^{26} FLOP. That median is the operational AGI line written into the protocol before the larger runs. The empty-aisle line, extrapolated on its own, is still near 30.3% at 102610^{26} FLOP. The installation consistent with the point estimate is the same 170 to 190 g carrot, on the same gantry limits, replaced every 36 hours.