How much do language models memorize? · HackerPapers discussions

How much do language models memorize? — mentioned by cpldcpu

Sat, 12 Sep 2026 22:12:09 GMT

>Notably, the neural network component behaves like a compressed index of the training data. The "Compression" part is indeed about the model finding order in the training data. But this is not about applying some predefined compression algorithm on it, but by actually learning an algorithmic representation of the data. This has nothing to do with "consulting the training data", because the training data cannot be reconstructed based on this information. Roughly speaking, there are two types of information being stored in the NN: Memorization and generalization, or, shannon entropy and…

Paper and discussion history. A mention is not an endorsement.

How much do language models memorize? — mentioned by adsharma

Sun, 05 Oct 2025 01:49:22 GMT

Recent: 3.6 bits per param https://arxiv.org/abs/2505.24832

Paper and discussion history. A mention is not an endorsement.

How much do language models memorize? — mentioned by kouteiheika

Tue, 30 Sep 2025 08:31:11 GMT

> Why is it okay for a model to serve up paraphrased books but verbatim copies from the Pirate Bay are illegal? Because they are not actually memorizing those books (besides few isolated pathological cases due to imperfect training data deduplication), and whatever they spit out is in no way a replacement for the original? Here's some back-of-the-envelope math: Harry Potter and the Philosopher's Stone is around ~460KB of text and equivalent to ~110k Qwen3 tokens, which gives us ~0.24 tokens per byte. Qwen3 models were trained on 36 trillion tokens, so this gives us a dataset of ~137TB. The…

Paper and discussion history. A mention is not an endorsement.

How much do language models memorize? — mentioned by jychang

Wed, 23 Jul 2025 10:31:42 GMT

You definitely want to use 4bit quants at minimum. https://arxiv.org/abs/2505.24832 LLMs usually have about 3.6 bits of data per parameter. You're losing a lot of information if quantized to 2 bits. 4 bit quants are the sweet spot where there's not much quality loss.

Paper and discussion history. A mention is not an endorsement.

How much do language models memorize? — mentioned by kouteiheika

Mon, 16 Jun 2025 06:08:34 GMT

No, assuming that just because it was in the training data it must be memorized is hare brained. LLMs have limited capacity to memorize, under ~4 bits per parameter[1][2], and are trained on terabytes of data. It's physically impossible for them to memorize everything they're trained on. The model memorized chunks of Harry Potter not just because it was directly trained on the whole book, which the article also alludes to: > For example, the researchers found that Llama 3.1 70B only memorized 0.13 percent of Sandman Slim, a 2009 novel by author Richard Kadrey. That’s a tiny fraction of the 42…

Paper and discussion history. A mention is not an endorsement.

How much do language models memorize? — mentioned by unconstrastive

Sat, 14 Jun 2025 19:19:47 GMT

This is why GenAI exists. GenAI took the good people, the money, the resources and the scope. Zuck tolerates entertains ideas until he doesn’t. It’s clear blue sky research is going to be pushed even further into the background I agree with most this, I just think we have different meanings of failure. FAIR has "failed" in the eyes of Meta leadership in that they have not converted into a "frontier AI lab" like Deepmind, and as a result they are being sidelined (much like Google Research, which I admit was a bad example). But the orgs were founded to pursue basic research and I think it's not…

Paper and discussion history. A mention is not an endorsement.

How much do language models memorize? — mentioned by Xmd5a

Thu, 05 Jun 2025 11:43:53 GMT

Here's my recommendation: What you're attempting to do, integrating an agent in your business, is difficult. It is however relatively easy to fake. Just setup a quick RAG tool, plug it into your LLM, and you're done. From the outside, the only difference between a quick-n-dirty integration and a much more robust approach will be in numbers. One will be more accurate than the other, but you need to actually measure and count performance to establish it as a fact and not just a vibe. First advice: build up a dataset and measure performance as you develop your agent. Or just don't, and deliver…

Paper and discussion history. A mention is not an endorsement.

How much do language models memorize? — mentioned by JimDabell

Thu, 05 Jun 2025 05:29:34 GMT

There’s a related paper that Meta published a couple of days ago that is worth looking at: > How much do language models memorize? — https://arxiv.org/abs/2505.24832 — https://news.ycombinator.com/item?id=44171363 It shows that models are limited in how much they can memorise (~3.6 bits per parameter), and once that threshold is reached, the model starts to generalise instead of memorise.

Paper and discussion history. A mention is not an endorsement.

How much do language models memorize? — HN submission

Tue, 03 Jun 2025 15:54:32 GMT

How much do language models memorize?

58 points; 1 discussion reply on HN. Paper and source history. A submission is not an endorsement.