aiqbench analysis
The paper tests whether small transformers trained on mixed correct and incorrect math text prefer the correct solution. It argues that the standard pairwise accuracy test overstates the effect, and that scoring against the full space of possible errors shows a modest preference (about 59% at 86M parameters) that disappears below that scale. For readers, the main takeaway is methodological: how a "preference" is measured can change the conclusion, and the paper finds that these models mostly leave consistent false patterns unfiltered rather than filtering falsehood as such.
Truth as a Compression Artifact in Language Model Training
arXiv:2603.11749v4 Announce Type: replace-cross Abstract: When a transformer is trained from scratch on corpora that solve the same mathematical problem both correctly and incorrectly, which solution does it prefer? The usual test is pairwise, and it is not neutral: when the error can take several forms, a model that merely reproduces training frequencies wins it without preferring anything. We instead score the correct completion against every string the error generator can produce, a share on which frequency matching scores exactly 0.500. At 86M parameters the correct completion holds 59.0% +/- 0.8% of the mass against random errors (59.4% on an independent corpus), where the single realized error gives 77.6% and pair accuracy 87.4%, a three- to fourfold overstatement of the excess above chance; below 86M the primary corpus shows no preference. The coherent false rules of the main corpora are not filtered at any size (43.5--48.8%; the shortfall below 0.5 is a completion-length effect). Two experiments built to identify a compression mechanism raise the pairwise measures while the summed share stays flat or falls: both are negative controls, and compression remains motivation, not a measured mechanism. With the scored false strings held fixed, a false rule whose selector the model can learn moves the correct share from 5 points above the null to 7 below (418M--3.08B), though the correct completion still outweighs the selected string. What these models leave unfiltered is consistent or learnable false structure (one single-rule corruption excepted), not falsehood as such.
Abstract:When a transformer is trained from scratch on corpora that solve the same mathematical problem both correctly and incorrectly, which solution does it prefer? The usual test is pairwise, and it is not neutral: when the error can take several forms, a model that merely reproduces training frequencies wins it without preferring anything. We instead score the correct completion against every string the error generator can produce, a share on which frequency matching scores exactly 0.500. At 86M parameters the correct completion holds 59.0% +/- 0.8% of the mass against random errors (59.4% on an independent corpus), where the single realized error gives 77.6% and pair accuracy 87.4%, a three- to fourfold overstatement of the excess above chance; below 86M the primary corpus shows no preference. The coherent false rules of the main corpora are not filtered at any size (43.5--48.8%; the shortfall below 0.5 is a completion-length effect). Two experiments built to identify a compression mechanism raise the pairwise measures while the summed share stays flat or falls: both are negative controls, and compression remains motivation, not a measured mechanism. With the scored false strings held fixed, a false rule whose selector the model can learn moves the correct share from 5 points above the null to 7 below (418M--3.08B), though the correct completion still outweighs the selected string. What these models leave unfiltered is consistent or learnable false structure (one single-rule corruption excepted), not falsehood as such.
v4: 30 pages, 2 figures. Revised evaluation using probability mass over the full false support; added negative controls and learnable-selector experiments up to 3.08B parameters. Updated claims and discussion
Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
arXiv:2603.11749 [cs.CL]
(or arXiv:2603.11749v4 [cs.CL] for this version)
Full text under the publisher's open licence · 283 words
Related coverage
VoRTeC: Taming Foundation Flow for One-step Real time Video Compression
arXiv · Sep 3, 2026
Beyond Retraining-Free MoE Compression: A Cost-Normalized Study of Post-Compression Adjustment
arXiv · Sep 9, 2026
Persian Pixel: A large-scale synthetic OCR dataset for Persian language
arXiv · Jul 23, 2026
Benchmarking Composable Compression Techniques in Mixture-of-Experts LLMs
arXiv · Aug 25, 2026