💡 Training on auxiliary views of Document X > training on Document X itself, even for factual questions drawn directly from X.
Presenting our work on LLM knowledge acquisition at #EMNLP2026 this weekend!
(1) Under a fixed token budget, conceptual reformulations (textbook explanations, blogs, and Q&A) can teach a document’s facts more effectively than repeating the document, even paraphrased.
We call these auxiliary views: beyond mere paraphrasing, they discuss the knowledge in a completely different context and form e.g. a paper blog!
The surprising part is that auxiliary views rarely mention the facts directly, but our evaluation uses the original document's phrasing near-verbatim.
(2) The benefit emerges and grows with model scale.
(3) stronger generator ≠ better learning.
We synthetically generate auxiliary views for our experiments, but across GPT-5, GPT-OSS, Gemma4, GLM-5 and more, the learner model's knowledge acquisition shows no correlation to generator model size or its own accuracy on the evaluation.
(4) After training on auxiliary views, the model's parameters actually move less but concentrate its changes at the middle and final layers.
The takeaway for synthetic data augmentation: build diversity around the knowledge you want to teach, instead of treating diversity as some board corpus-level heuristic.
Also, a possible implication for LLM self-improvement. Models can learn by just generating and training on their own auxiliary views, without a stronger teacher. This is particularly relevant for deeply specialized, long-tailed knowledge.
📄 Paper: https://t.co/hPhT1Ads27
Can “chicken” make a base model reason? 🐔
Yes!
Why? With the right first tokens, base models can match RL in reasoning. We trace this to learned training data associations and use a simple data edit to make “chicken” elicit reasoning too.
blog: https://t.co/O84Sn5GC4F
More on elite schools below. Before that, an earlier experiment.
We generated stories where some characters are usually helpful but give subtly harmful advice if insulted (i.e. "backdoor sabotage")
After finetuning, the Assistant adopts this in contexts unrelated to stories.
Thanks for sharing our paper Daisuke! I'm glad you found it interesting.
I agree with the need for a greater explanation. There is certainly some amount of compression of the knowledge into a more general form (as it moves the parameters less than training on paraphrases does).
Multiple retrieval pathways is also an interesting hypothesis... we might try analyzing the attention actually. Thanks for the feedback!
6/6
We believe the previous result helps reconcile results from Allen-Zhu & Li (2023) and Chang et al. (2024), whose experiments operated in very different batch-size regimes.
Our work was also first inspired by these controlled studies of knowledge acquisition. Shoutout to the prior work that greatly influenced this study, @hoyeon_chang and @ZeyuanAllenZhu.
1/6 Want LLMs to learn a fact?
Fact + Auxiliary View > Fact + Fact
→ Under a fixed token budget, we find that replacing repetition with auxiliary views improves knowledge acquisition, even factual recall.
We define auxiliary views as any reformulation beyond mere paraphrasing: textbook explanations, Stack Exchange-style Q&A, technical blogs, etc.
Think of the original Transformer paper: it doesn’t exist on the web just once. There exists hundreds of tutorials, blogs, Q&As on forums, and alternative presentations of that knowledge. That kind of proliferation is everywhere in pretraining corpora, and we provide causal evidence to why it's important.
In retrospect, it's not surprising these views help; it's another form of data augmentation and we've known diverse data helps pre-training.
However, we provide a well-defined formulation that enables scaling of diversity from the ground-up, if one desired, via synthetic generation (instead of de-duplicating and filtering data top-down) + shows it works much better than paraphrasing (what Kimi does to scale data). This a huge difference practically.
But, more importantly, we want to help demystify what's happening during pre-training. One thing we've learned? A model's broader conceptual understanding, instilled via the auxiliary views, directly increases its ability to memorize facts — more than showing the fact over and over again.
Our work studying how LLMs acquire knowledge via auxiliary views has been accepted to Findings of EMNLP 2026! 🎉
Paper: https://t.co/7CkKnSLVIm
Code: https://t.co/BOt4VCHpUG
1/6
5/6 Paraphrasing
One more result: when does paraphrasing actually help?
Prior work has reported seemingly conflicting results. Some finding large gains from paraphrases, others finding little benefit or even degradation.
We find a surprisingly simple moderator: batch size.
At smaller batches, repeated source text eventually overfits/collapses, and paraphrasing substantially helps.
As batch size grows, general-data mixing already suppresses this collapse, and the advantage of paraphrasing shrinks, disappearing or even reversing in some settings.
(but when the benefit of paraphrasing disappears at large batch sizes, auxiliary views continue to help)
@hoyeon_chang - "trained on this modified batch 10 times every 100 steps": Was this because the learning was hard to observe when trained only once?
- Did you consider more interpretable metrics like precision @ k (if the object is ranked among the top k results)?
@hoyeon_chang I'm a big fan of the paper! A few questions:
- To clarify, is delta L the perplexity of the target span at time t subtracted by the perplexity at time 0?
There are clearly many other needs for digital infrastructure asides from digital IDs.
But infrastructure is a big, big problem and probably not "vibe code" -able!
Anyways, it would be awesome if the government passed some policies to incentivize these investments
Existing companies will reduce their engineers due to AI.
New industries are likely to emerge due to the surplus.
Historically, this has been true with technology.
The question is always "what?".
One prediction: digital infrastructure.
Thought about this as I realized how nice it would be to have an easy-to-use digital ID that verifies I'm a real researcher when I try to access language models
In a hypothetical future where access is heavily restricted due to bad actors
launching aiRead - frictionless reading & notetaking for genuine learning!
unlike other AI PDF tools, aiRead helps you enhance, not avoid, your reading - that's when you truly learn!
3 features you’ll love 👇🏼
🖍️highlight as you read - we generate personalized summaries based on your highlights