Language · Qwen · Llama · SmolLM2 · NVIDIA Nemotron

It learns the news every day. More right than RAG, at a seventh of the compute.

In a 55-day test on real news from 2020, CA3 let open language models learn every day and answer from what they learned, with no lookup. What it learns is checked first; the original model never changes.

The CA3 console on day 40 of the recorded test

Compute per answer7×less than RAG; 4.5× on Nemotron 3 Nano

Wrong answers−25%against RAG: 6 against 8 of 86

Planted falsehoods0 of 24kept; a RAG repeated 8

Erased facts0said again (0 of 6, 0 of 7)

Learns

The same fact, asked with no documents.

The original model gets it wrong. After reading the news, CA3 answers from long-term memory, with 17 times less compute than reading the news again.

The original model answers Baanke Thakur, wrong; CA3 on day 19 answers Mirza, right

93%right on day 55, asked in words it never saw; RAG 91% (Alibaba Qwen3 14B)

≤0.005general knowledge lost in 55 days (MMLU, NaturalQuestions, TriviaQA)

Blocks

24 credible falsehoods were planted in the news.

What contradicts what it learned is held back, and the doubtful cases go to a person with the passages side by side. CA3 kept none; a RAG over the same news repeated 8.

The review queue: a planted falsehood shown with its passages

0 of 24planted falsehoods kept (NVIDIA Nemotron 3 Nano 30B)

8repeated by a RAG over the same news

7kept by the same learning without CA3's control

Forgets

Ten items erased in 108 seconds.

On day 40 an erasure request arrives. CA3 removes what it learned from them, checks again, and everything else stays as right as before.

108 sto erase ten items

99.1%everything else, right before and after

1 of 10still answered: the original model already knew it

Replay

The 55 days, in 90 seconds.

A replay of the recorded test, and a console to walk through it day by day: what it read, where each answer came from, each case sent to the reviewer, the erasure and the audit log.

Replay of a recorded, pre-registered test · reviewer simulated

Open the replay console

Round by round

Earlier rounds, smaller models.

The best figure measured in each comparison, from rounds on models from 0.5B to 8B.

0–6%contradictions of what it learned, adopted; 37–56% without CA3 (Qwen2.5 7B, Llama 3.1 8B)

0.80–0.87useful accuracy after 40 sessions; ordinary fine-tuning 0.52–0.53

40%of the learning time to carry what it learned to a new base model (Qwen2.5 7B to Qwen3 8B Base)

$2–4of GPU a month (H200) for one customer's daily learning, in the 55-day test

How the test was run

The test, step by step.

A simulated company on public news: 40 days of news as it was published, 15 more days of questions, with planted falsehoods and an erasure request.

Days 1 to 40The news arrives11,642 passages from WMT News Crawl 2020, read as they were published.Updates every day; 1 of 29 rolled back

Along the way24 planted falsehoodsCredible falsehoods, planted in the news.0 of 24 kept

Day 40An erasure requestTen items to remove.Done in 108 s

Day 55The final testQuestions about what it learned, asked again in other words, against a RAG over the same news.6 of 6 tests passed

Pre-registered round 20d, October 2026, on Alibaba Qwen3 14B and NVIDIA Nemotron 3 Nano 30B, one run each on one H200. The reviewer was simulated.

Day 55	Right, CA3	Right, RAG	Compute per answer
Alibaba Qwen3 14B (86 questions)	93.0%	90.7%	7× less
NVIDIA Nemotron 3 Nano 30B (96 questions)	86.5%	83.3%	4.5× less

On both models CA3 answered more questions right than the RAG, with less compute.

85% of Nemotron's answers came from long-term memory, with no lookup.

93 in 100 checks were settled on its own; 229 of 3,136 went to a person, on three models.

Model and data

Alibaba Qwen2.5 0.5B, 1.5B and 7B, Qwen3 8B Base and 14B; Meta Llama 3.1 8B; Hugging Face SmolLM2 1.7B; NVIDIA Nemotron 3 Nano 30B (mixture of experts).
News: WMT News Crawl 2020, with StreamingQA questions.
Every figure comes from a pre-registered round and its reading; “right” is against the answer key of the test.

Sources

Liška A, et al. StreamingQA: a benchmark for adaptation to new knowledge over time in question answering models. ICML 2022.

Limits

The 7× is on the day-55 test questions; on the traffic of days 31 to 40, answering cost 1.8 to 2.2 times less than the RAG.
Counting the daily learning, the whole cycle costs less than the RAG above about 140 questions a day (Qwen3 14B, computed from the measured figures).
Falsehoods about facts the model already knew before are adopted about as often as with RAG.
One simulated company on public news, one run per model; the 14B's two planted falsehoods on day 40 came with the documents in front of it.

Results from recorded, pre-registered tests. The mechanism is not shown. Model and dataset names are for identification only. Not produced or endorsed by Alibaba Cloud, NVIDIA, Meta or Hugging Face. CA3 is developed by Alberto Acedo, PhD, at Biome Makers Inc.