iiinova — where your digital self keeps evolving

Matthew Finlayson @mattf.nl · 17/10/2025

This opens the door to a verification system analogous to cryptographic message authentication—where the model ellipse functions as a secret key. Providers could verify outputs to trusted third parties without revealing model parameters. 6/

1161

Matthew Finlayson @mattf.nl · 17/10/2025

The forgery resistance comes from the excessive complexity of extracting an ellipse from an API: O(d³ log d) queries and O(d⁶) time to fit. For a 70B model, that's ~$16M in API costs and millennia of computation time 💸⏰5/

191

Matthew Finlayson @mattf.nl · 17/10/2025

We tested this on models like Llama 3.1, Qwen 3, and GPT-OSS. Even when we copied their linear signatures onto other models' outputs, the ellipse signature cleanly identified the true source by orders of magnitude. 4/

1121

Matthew Finlayson @mattf.nl · 17/10/2025

The key insight is that LLMs with normalization layers produce outputs that lie on the surface of a high-dimensional ellipse. This geometric constraint acts as a signature unique to each model. 2/

1213

Matthew Finlayson @mattf.nl · 23/06/2025

The project was led by Murtaza Nazir, an independent researcher with serious engineering chops. It's his first paper. He's a joy to work with and is applying to PhDs. Hire him! It's great to finally collab with Jack Morris, and a big thanks to @swabhs.bsky.social and Xiang Ren for advising.

040

Matthew Finlayson @mattf.nl · 23/06/2025

Our technical insight is that logprob vectors can be linearly encoded as a much smaller vector. We make prompt stealing both *more accurate* and *cheaper*, by compactly encoding logprob outputs over multiple generation steps, resulting in massive gains over previous SoTA methods.

140

Matthew Finlayson @mattf.nl · 23/06/2025

We noticed that existing methods don't fully use LLM outputs: either they ignore logprobs (text only), or they only use logprobs from a single generation step. The problem is that next-token logprobs are big--the size of the entire LLM vocabulary *for each generation step*.

130

Matthew Finlayson @mattf.nl · 23/06/2025

When interacting with an AI model via an API, the API provider may secretly change your prompt or inject a system message before feeding it to the model. Prompt stealing--also known as LM inversion--tries to reverse engineer the prompt that produced a particular LM output.

110

Matthew Finlayson @mattf.nl · 23/06/2025

I didn't believe when I first saw, but: We trained a prompt stealing model that gets >3x SoTA accuracy. The secret is representing LLM outputs *correctly* 🚲 Demo/blog: mattf1n.github.io/pils 📄: arxiv.org/abs/2506.17090 🤖: huggingface.co/dill-lab/pi... 🧑‍💻: github.com/dill-lab/PILS

1110

Matthew Finlayson @mattf.nl · 16/03/2025

If you are writing a paper for #colm2025 and LaTeX keeps increasing your line height to accommodate things like superscripts, consider using $\smash{2^d}$, but beware of character overlaps.

Screenshot of inconsistent line height to make way for a superscript.

Screenshot of text with consistent line height.

2120

Matthew Finlayson @mattf.nl · 25/02/2025

6/ Our method is general, and we are excited to see how it might be used to better adapt LLMs to other tasks in the future. A big shout-out to my collaborators at Meta: Ilia, Daniel, Barlas, Xilun, and Aasish (of whom only @uralik.bsky.social is on Bluesky)

000

Matthew Finlayson @mattf.nl · 25/02/2025

5/ Training on self-demos, our model learns to better leverage the context to answer questions, and to refuse questions that it is likely to answer incorrectly. This results in consistent, large improvements across several knowledge-intensive QA tasks.

100

Matthew Finlayson @mattf.nl · 25/02/2025

4/ To obtain self-demos we generate candidate responses with an LLM, then use the same LLM to compare these responses to the gold one, choosing the one that best matches (or refuses to answer). Thus we retain the gold supervision from the original responses while aligning the training data.

111

Matthew Finlayson @mattf.nl · 25/02/2025

3/ OOD responses encourage the model to answer questions it does not know the answer to, and since retrievals are added post-hoc, the responses tend ignore or even contradict the retrieved context. Instead of training on these low-quality responses, we use the LLM to generate "self-demos".

100

Matthew Finlayson @mattf.nl · 25/02/2025

2/ A popular recipe for adapting LLMs for RAG involves adding retrievals post-hoc to an existing instruction-tuning dataset. The hope is that the LLM learns to leverage the added context to respond to instructions. Unfortunately, the gold responses in these datasets tend to be OOD for the model.

100

Matthew Finlayson @mattf.nl · 25/02/2025

🧵 Adapting your LLM for new tasks is dangerous! A bad training set degrades models by encouraging hallucinations and other misbehavior. Our paper remedies this for RAG training by replacing gold responses with self-generated demonstrations. Check it out here: arxiv.org/abs/2502.10

2181

Matthew Finlayson @mattf.nl · 12/12/2024

Putting together an unofficial usc Beamer template, I noticed that the USC style guide lists 4 formats for “cardinal red” but each of them is different: PMS 201 C is #9D2235 CMYK: 7, 100, 65, 32 is #A1003D RGB: 135, 27, 30 is #991B1E HEX: #990000 Is this normal? The CMYK is especially egregious.

The usc style guide list of formats for “cardinal” (see main post for list)

The rgb and CMYK colors side by side. The CMYK is considerably pinker

000

Matthew Finlayson @mattf.nl · 09/12/2024

In Vancouver for NeurIPS but don't have Taylor Swift tickets? You can still spend the day going through our tutorial reading list: cmu-l3.github.io/neurips2024-... Tuesday December 10, 1:30-4:00pm @ West Exhibition Hall C, NeurIPS

A diagram demonstrating text generation with beam search. One of the paths reads “Taylor Swift is the only person to…”

0292

Matthew Finlayson @mattf.nl · 06/12/2024

Curious about all this inference-time scaling hype? Attend our NeurIPS tutorial: Beyond Decoding: Meta-Generation Algorithms for LLMs (Tue. 1:30)! We have a top-notch panelist lineup. Our website: cmu-l3.github.io/neurips2024-...

Panelist photos: Rishabh Agarwal (Google, McGill), Noam Brown (OpenAl), Beidi Chen (CMU), Nouha Dziri (AI2), Jakob Foerster (Oxford, Meta)

1273

Matthew Finlayson @mattf.nl · 22/11/2024

This is niche but the LLM360 logo always reminds me of the 2014 iOS game Oquonie

Screenshot from Oquonie with a long-necked character.

030

Matthew Finlayson @mattf.nl · 17/11/2024

I made a map! Thank you to my 2019 self for providing the code github.com/mattf1n/Reli...

A relief map of Los Angeles rendered in Blender giving it a 3D appearance.

2170

Matthew Finlayson @mattf.nl · 14/11/2024

I’m proud of this tikz drawing I made today for our upcoming NeurIPS tutorial on decoding (our paper: arxiv.org/abs/2406.16838)

A diagram of how beam search works. The graphic is a tree with “Taylor swift is” at the root and possible continuations branching off.

0151

Matthew Finlayson @mattf.nl · 23/01/2024

Link not working for me :(

100

Matthew Finlayson @mattf.nl · 25/10/2023

A cute TikZ diagram for your enjoyment: (I spent too much time making this for a presentation)

A TikZ diagram of the standard 3-simplex

040

Matthew Finlayson @mattf.nl · 11/10/2023

We translate this finding into a new class of threshold-free sampling methods. In pilot studies, our easy-to-implement method (BA-η) performs competitively against existing methods, and outperforms them in low-entropy (close to greedy) decoding 💪

100

Matthew Finlayson @mattf.nl · 11/10/2023

Knowing the set of possible model outputs, and assuming that the model minimizes its training loss w.r.t. the true distribution, we can actually deduce a set of distributions containing the true distribution, and thereby find tokens that must have nonzero true probability 🕵️‍♀️

100

Matthew Finlayson @mattf.nl · 11/10/2023

In particular, we identify the SOFTMAX BOTTLENECK as a source of the errors. The softmax bottleneck entails that model outputs are restricted to a subset of probability distributions. If the true distribution is not in this set, the model cannot output it 😱

120

Matthew Finlayson @mattf.nl · 11/10/2023

Nucleus and top-k sampling are ubiquitous, but why do they work so well? We explain the theory and give a method to address model errors at their source (the softmax bottleneck). 📄 arxiv.org/abs/2310.01693 🧑‍💻 github.com/mattf1n/basi... 1/🧵

Title and figure 1 for the paper "Closing the Curious Case of Neural Text Degeneration". The figure depicts the next-token distribution according to an LM. Our method is able to pick out high quality next-token candidates while discarding low-quality ones, even when the low-quality tokens have higher probability than the high-quality ones.

182