← NewBrain research overview

Kira Labs / Public research report / October 26, 2026

NewBrain learns some new facts.
Value selection, retention and memory use remain limited.

Three distinct small synthetic-learning studies documented limited new-fact acquisition and unfamiliar-form answers alongside forgetting, with exact saved-state restarts. The later value-weighted NB14 intervention is a separate experiment with its own current public report. None improves accepted Original12 (2/12) or establishes general conversation or Qwen equivalence.

6 / 12New facts after teaching in the three-world study, versus 0/12 frozen
10 / 24New facts after both replay schedules in the six-world study
2 / 12Accepted Original12 unchanged, a different benchmark and model-state boundary

Study 1 · Three-world ordinary learning and external context

Completed October 25 local / October 26 UTC. The frozen-reference and taught arms used three fixed synthetic worlds with a 39-token vocabulary, 64 hidden units and 8,343 parameters. The experiment completed 15,360 SGD updates and 122,880 example presentations. Results below are exact complete answers with true EOS and without external retrieval.

Question partitionFrozenTaughtInterpretation
New associations0/126/12Some newly taught complete facts acquired
New subject, alternate form0/124/12Narrow combination transfer
Earlier questions15/2416/24Net gain masks four lost and five regained correct cases
Untaught unknown controls0/60/6Abstention was not explicitly trained

Context-use experiment

The same taught learner was also evaluated with different external contexts. The selected evidence was treated as context, not counted as knowledge newly acquired by the model weights.

Taught-model contextNew /12Transfer /12Earlier /24
No external context6416
Raw event history429
Latest event per subject429
Correct single oracle event329
Wrong-subject event109

Even supplying a correct oracle event did not guarantee correct output. That is an observed context-use limitation, but these comparisons cannot by themselves distinguish representational binding failure from context length, training distribution, formatting or decoding effects.

All 540 of 540 planned final restart answers reproduced in a distinct ordinary process with zero added updates. The saved states, parameters, counters and source identities matched. All 28 wrong taught no-context answers first differed at the third word—the fact value, including six untaught unknown controls. A separate earlier numerical precision diagnostic does not override these measured complete-answer results.

Study 2 · Six-world matched replay-form scheduling

Completed October 26. A different fixed experiment compared frozen, synchronized old-question scheduling and staggered scheduling across six fresh synthetic worlds. The two taught conditions used matched updates, row ordering and fact/form exposure; the change altered old-question form co-occurrence and phase together, so it cannot isolate a general diversity effect.

ConditionNew facts /24Alternate form /24Old cases /48Untaught unknown /12
Frozen00260
Synchronized replay109260
Staggered replay109250

Both taught schedules retained 19 of the 26 previously correct old answers and lost the same seven. Synchronized replay recovered seven previously incorrect old answers; staggering recovered six. Therefore the single old-case difference is a missed recovery, not extra loss of a formerly correct case. New-fact and transfer success sets were identical across schedules. These results show no observed benefit from this specific intervention, not equivalence among all replay methods.

The experiment completed 36,864 SGD updates and 294,912 example presentations. A separate process reproduced all 324 of 324 planned final answers from 18 states, with unchanged saved state, parameters, counters and identities and no further updates. Total recorded free answers included 432 during learning and 324 at restart; read-only expected-path observations were performed separately after free generation.

Value selection, not missing EOS

Every wrong final answer first differed at the third, value-bearing token. All complete answers ended with true EOS. In synchronized replay, the correct value ranked second or third in 40 of 51 wrong known-fact cases; in staggered replay, 41 of 52. Across both schedules, the correct-value probability improved for all 48 new/transfer questions, but 14 new-association and 15 transfer answers were still wrong. Higher probability is not the same as a correctly selected complete answer.

A proposed next controlled test would hold a trained state, question form and other input tokens fixed while changing only the subject in fresh preregistered cases, then compare free output and value-logit distributions. This subject-sensitivity probe is unrun and would not by itself establish internal memory retrieval.

Study 3 · Separate completed six-world subject-sensitivity learner

October 26, 2026. This tiny 8,343-parameter, 39-token, 64-hidden-unit recurrent reference decoder completed 30,720 learning updates in six new synthetic worlds. Scored answers were freely generated without external retrieval and contained three exact words plus true EOS.

Complete free-answer partitionFrozenTaught
New direct facts /2406
New alternate-form answers /2405
Older cases, both forms /483330

Only three of six worlds gained correct new answers. Teaching lost four formerly correct old answers and recovered one old error. All wrong outputs first diverged at the value token despite correct framing and EOS. A distinct process reproduced 192/192 full answers from saved states, with no additional training.

Boundary: these small synthetic results are not an accepted NewBrain checkpoint, broadly generalizable learning, or an NB14 training-weight comparison. Original12 remains 2/12. The subject-pair observations are dependent, not statistical proof of a retrieval mechanism. Read the later separate matched value-weighting study →

Evidence limitations and next step

Original12 unchanged: the accepted 2/12 result, earlier failed trained 0/12 candidate and unbound accepted checkpoint are separate from both successful small synthetic-reference studies. They cannot be treated as upgraded versions of the accepted learner.

All three studies used small, closed-vocabulary worlds; those worlds and seeds are consumed and must not be rerun or tuned. The six-world study was internally reviewed against saved journals, answer rows and source identities, not externally replicated across independent laboratories. Peak RAM and full physical I/O were not measured to a complete host-level standard in these studies. Unknown/abstention control behavior was not explicitly taught.

Later update: a separate six-world subject-sensitivity test and then a distinct NB14 value-weighted study were completed. A new varied-form teaching comparison is proposed only and remains unrun. Separate acquisition, context benefits and answer selection in future tests without changing accepted Original12.