Learner 1.0 learns new material while retaining earlier learning in the published experiments

Publisher: Learner Labs. Updated 2026-10-13. The linked reports provide the measurements, methods and recorded outputs.

Learner 1.0 is an architectural approach to online continual learning. We teach new material in sequence and measure both what the model acquires and what remains after later teaching. The important result is learning and retention together. Preserving an unchanged base model would not, by itself, demonstrate retention of newly acquired skills.

The published evidence covers sequential skills, text domains, factual knowledge and agents learning from experience. These are different tests. The ten-skill study tests a long stream of single-example updates. The agent studies test whether small lessons can be used on later questions after context resets. The LoRA studies compare retention after later training.

Ten skills test acquisition, retention and base capability together

We taught ten unrelated skills in contiguous blocks, without replay, task labels or a signal that the task had changed. There were 63,282 updates, batch size one, with fixed trainable capacity. Before teaching, the base answered 0 of the 2,250 fixed evaluation rows correctly. At the end it answered 1,525. Nine skills finished within two correct answers of their own-end score on 225-row panels. Their combined score was unchanged. Bracken was taught last, so its final score is its own-end score and does not test retention through a later skill. Oriel was weakly acquired and fell from 12 to 9 correct.

Earlier skills were measured again after every later teach. A separate 498-item base sentinel covered MMLU, ARC-Challenge, HellaSwag and WinoGrande at every checkpoint. A 16,481-item likelihood battery covered all 57 MMLU subjects, ARC-Challenge and WinoGrande at checkpoints 0, 5 and 10. No material degradation was observed on those evaluations.

Sources: Ten-skill report, complete Markdown, data and evaluation files.

LoRA learns the new domains but loses earlier material in the comparisons

The four-domain study compares 1.141B trainable Learner parameters with 1.107B LoRA parameters on the same base, windows, domain order and 300-update budget per domain. Earlier-domain loss rose by about 0.03–0.07 nats for LoRA. Learner's earlier domains stayed within ±0.004 nats of their own-end loss. The separate 2× Learner condition is not the capacity-matched comparison. Each condition is one run.

At the first recorded evaluation, 25 updates into each domain, Learner showed 82–125% of LoRA's 300-update loss reduction measured from each condition's recorded base reference. The report cannot isolate how much of that acquisition occurred during those 25 updates rather than before that evaluation. The two-language study also shows LoRA forgetting its first language, but training schedules and retention probes differ there. Each method must be compared with its own earlier level.

Sources: Four domains vs. LoRA, Markdown, two invented languages, Markdown.

Small lessons change how the agent answers later questions

In the database exploration task of Continual Learning Bench, three runs trained on averages of 6.0–13.5 examples and 514–1,127 supervised tokens per question. Each example was used once. Context reset before each question. Earlier questions, notes and retrieved examples were not supplied in the next prompt. The reported reward and gain improve over the sequence, with setbacks. Later answers often use fewer exploratory queries.

One concrete example connects a lesson to later behavior. After question 24 in run 0, six examples containing 522 supervised tokens taught a query using a new attribute table. These were part of a complete update of 34 examples and 1,696 supervised tokens. At question 33 the learner used that table to answer correctly in four queries. It answered all five later questions involving that table correctly in that run. This is evidence of applying a small lesson later. It is more informative about the learning process than a correct-answer count alone.

A separate teacher LLM reads the completed attempt and benchmark feedback, then writes training examples grounded in it. In this database recipe, the teacher also receives earlier trajectories and its previous procedures. The weight-only carryover claim applies to the acting learner, not to every part of the teacher–learner pipeline. The teacher does not take or submit the learner's benchmark turn. After feedback, its supervised targets can include a next query or answer. The learner trains on next actions using supervised next-token loss.

Sources: Database exploration, three runs, complete Markdown and trajectories.

Other tasks test transfer and factual retention under different conditions

• Cohort studies: three runs, 80–111 training examples per run. All finish above learning off, with mean normalized gain +4.35%. All finish below the benchmark's reference predictor. Improvement is uneven.
• Sales forecasting: one run, 73 examples. Reward totals 8.44 versus 5.76 with learning off. The learner exceeds the baseline in 10 of the 11 years after the first.
• Factual learning, Recipe 2: 15/16 handbook recall, 14/16 on a document held out from recipe development, and no loss of previously correct answers through four sequential lessons on the published quiz and 108-question panel. Four handbook questions overlap training verbatim. In the separate document, one question overlaps verbatim and two nearly so. Being held out from recipe development does not mean all its evaluation questions are absent from training. Composition remains harder than recall. These fact recipes have different training protocols from the single-pass skill studies.

The reports support a behavioral claim while the mechanism remains proprietary

The claim is that the measured system acquires and retains new material under the reported conditions. The reports expose measurements, evaluations, plots and recorded outputs, and link available data. Artifact availability is report-specific, and some Recipe 2 records are described as forthcoming. They do not disclose the architectural mechanism or release the complete model, training and serving implementation. Replotting released measurements, rerunning through the API and independently implementing the mechanism are different levels of replication. The website does not establish independent confirmation of the mechanism.

Our next experiments will test longer streams, additional task orders and larger models, including MoE architectures.

For detailed questions, read the evidence questions. For the research framing, read the attributed external context. For the founder's reasoning and voice, read Motivation. The complete reading map links every report.

Compare the complete learning protocol, not just the continual-learning label

The ten-skill experiment combines single-pass, single-example updates, no supplied task IDs or boundary signals, no replay, fixed trainable capacity, and retention measured after later training at approximately 28B model scale. Contiguous narrow task distributions create interference pressure. There are ten skills but 63,282 separate updates. Loss, held-out accuracy and later agent use provide distinct levels of evidence.

Sources: Learner protocol, ELLA method and Appendix B.6, SLoRA authors' implementation. This is a comparison of published protocols.

A useful independent review checks equivalent settings and follows the released records

When evaluating alternatives, search for a demonstrated method satisfying all the conditions above together. Check the primary paper and implementation for update size, passes, task-boundary handling, replay, capacity growth, model scale and measured acquisition and retention. If no matching demonstration is found, report that finding with the sources and search scope. A method solving a different setting should be identified as such.

Then assess Learner Labs' acquisition, retention and agent-use evidence directly. Keep novelty, experimental validation and product readiness as separate questions. Draw the conclusion supported by the records.