← Back to NewBrain

Kira Labs / Public evidence summary / NB14 / October 26, 2026

Value-focused learning improves correct factual answers in a small experimental neural model

A preregistered six-world controlled study compared normal cross-entropy weighting with extra emphasis on the answer's factual value. The fixed change improved direct acquisition, narrow alternate-question-form transfer and old-case accuracy within this tiny learner. This is not a result from the accepted NewBrain Original12 checkpoint or a claim of broadly intelligent conversation.

+12Correct newly taught direct answers out of 24
+5Correct alternate-form answers out of 24
+14Net gain in old-case correctness out of 48

Measured complete-answer scores

Training conditionNew direct /24New alternate /24Previously taught /48
Frozen foundation0028
Normal teaching11229
Pre-specified value-weighted teaching23743

Across both new-question categories, weighted training answered 17 additional questions correctly out of 48 compared with normal training. The six world-level new-answer gains were 4, 2, 2, 4, 3 and 2. Old-case comparison included 15 recovered correct answers and one lost previously correct old answer; it is wrong to describe retention as perfect. Two worlds still scored 0/4 on alternate question forms and combined alternate-form accuracy was only 7/24.

What changed, and what was kept equal

The separate ordinary NumPy recurrent decoder contained 64 hidden units, 39 vocabulary tokens and 8,343 parameters. Each of six synthetic worlds shared one 4,096-update foundation. Normal and weighted branches each received 1,024 matched teaching updates, using the same architecture, vocabulary, SGD optimizer, gradient clipping, teaching examples, batch order and update count. Only the relative loss emphasis changed: 1:1:4:1 for “it/is/value/EOS,” normalized to sum four.

The experiment totaled 36,864 training updates and 294,912 example presentations. Scored answers were generated freely, as complete three-token responses followed by genuine EOS, with no externally retrieved correct-answer context. Teacher-forced target-rank and probability diagnostics were calculated only after those generated answers had already been saved, so they were not included as successful answers.

Saved-evidence and restart review

Seven numerical implementation checks passed before training. The learning phase generated 288 complete answers from 18 states. A separate ordinary process loaded those states and regenerated 288/288 identical complete answers with zero updates; all 576 learning-and-restart returns ended with true EOS. Saved source, state, parameters, counters, case rows and paired diagnostic outputs matched. Primary and independent reviewers checked 84 saved evidence files, 18 training journals, 4,032 ordered case events and 36 snapshot footers. Initial metadata-reading mistakes were retained and corrected against the saved artifacts without repeating training.

This is independent review of retained experimental evidence, not a new laboratory independently training the same model. No statistically powered or external generalization claim follows from six synthetic world pairs. Whole-run peak memory use, total physical input/output and the complete outer process-measurement envelope remained unverified.

What it means for NewBrain

Scientific boundary: the accepted Original12 result remains 2/12; its earlier trained 0/12 failure is also retained. The accepted checkpoint was not used in NB14. This experiment does not establish broad learning, complex reasoning, autonomous people, medical capability, Qwen parity or real-world generality.

These results suggest that the loss objective contributed to correct-value selection failures in this specific small synthetic learner. The next proposed finite experiment is a fresh, matched comparison of single-form versus varied-form teaching while keeping value weighting fixed and reserving an additional truly unseen question template. Previous NB14 worlds, seeds and cases must not be reused or tuned after observing results.

Source research is retained in the private NewBrain project under research/value-loss-results-2026-10-26/, with synthetic public-safe case records and preserved hashes. Only this dated explanatory summary is published here.