Pau Rodriguez

The best part? LinEAS works on LLMs & T2I models.

Huge thanks to the team: Michal Klein, Eleonora Gualdoni, Valentino Maiorca, Arno Blaas, Luca Zappella, Marco Cuturi, & Xavier Suau (who contributed like a 1st author too🥇)!

💻https://github.com/apple/ml-lineas
📄https://arxiv.org/abs/2503.10679

Sparsity improves utility while mitigating toxicity. Toxicity results on Qwen2.5-7B using only 32 sentences, at different levels of sparsity γ that result in different support sizes (x axis). At 1K optimization steps, with a support of about 1% we maintain similar toxicity (left, center-left) while PPLWIK decreases (center-right) and MMLU increases (right). Note that too long optimizations (10k steps) might harm utility, due to overfitting. Similarly, short optimizations (e.g., 100 steps) and strong sparsity leads to low conditioning (mild toxicity mitigation).

LinEAS globally 🌐 optimizes all 1D-Wasserstein distances between source and target activation distributions at multiple layers via backprop. ✨ Bonus: we can now add a sparsity objective. The result? Targeted 🎯 interventions that preserve fluency with strong conditioning!

Existing methods estimate layer-wise 🥞 interventions. While powerful, layer-wise methods have some approximation error since the optimization is done locally, without considering multiple layers at once 🤔. We circumvent this problem in LinEAS with an end-to-end optimization ⚙️!

LinEAS learns lightweight maps to steer pretrained model activations. With LinEAS, we gain fine-grained control on text-to-image generation to induce precise styles (in the figure) or remove objects. The same procedure also allows controlling LLMs.

🦊Activation Steering modifies a model's internal activations to control its output. Think of a slider 🎚️ that gradually adds a concept, like art style 🎨 to the output. This is also a powerful tool for safety, steering models away from harmful content.

🚀 Excited to share LinEAS, our new activation steering method accepted at NeurIPS 2025! It approximates optimal transport maps e2e to precisely guide 🧭 activations achieving finer control 🎚️ with ✨ less than 32 ✨ prompts!

💻https://github.com/apple/ml-lineas
📄https://arxiv.org/abs/2503.10679

Our two phenomenal interns, Alireza Mousavi-Hosseini and Stephen Zhang @syz.bsky.social have been cooking some really cool work with Michal Klein and me over the summer.

Relying on optimal transport couplings (to pick noise and data pairs) should, in principle, be helpful to guide flow matching

🧵

Our work on fine-grained control of LLMs and diffusion models via Activation Transport will be presented @iclr_conf as spotlight✨Check out our new blog post machinelearning.apple.com/research/tra...

YouTube video by Deep Learning Barcelona Què és l'aprenentatge profund ? - La Dimoni de Maxwell #deeplearning #ciencia #català #barcelona

Què és l’aprenentatge profund ?

La @marionamec.bsky.social de @neurofregides.bsky.social ens ho explica en motiu del Deep Learning Barcelona Symposium 2024 (@dlbcn.ai), aquest dijous 19 de desembre.

#deeplearning #ciencia #català #barcelona

www.youtube.com/shorts/R4u_Z...

Evaluating your LLM uncertainties with Rougle-L will show clear winners... except that they aren't actually good. We find that Rouge-L spuriously favors some methods over others. 🧵1/4

📄 openreview.net/forum?id=jGt...
NeurIPS: Sunday, East Exhibition Hall A, Safe Gen AI workshop

Kudos to all co-authors 👏 Arno Blaas, Michal Klein, Luca Zappella, Nicholas Apostoloff, Marco Cuturi, and Xavier Suau.

Extra 👏 to Xavi for making this so great! Like a friend would say, he's the Rolls-Royce of the co-authors, and he should be regarded the first author too!

Summary:
🤝 Unifying activation steering w/ OT.
✨ Linear-AcT preserves distributions w/ interpretable ([0, 1]) strength.
💪 Robust: models/layers/modalities
💬 LLMs: toxicity mitigation, truthfulness and concept induction,
🌄 T2I: style induction and concept negation.
🚀 Negligible cost!

8/9 T2I models tend to generate negated concepts 😮

In the image, StableDiffusion XL prompted with: “2 tier cake with multicolored stars attached to it and no {white bear, pink elephant, gorilla} can be seen.”

✨Linear-AcT makes the negated concept disappear✨

7/9 And here we induce Cyberpunk 🤖 for the same prompt!

6/9 Amazingly, we can condition Text-to-Image (T2I) Diffusion with the same exact method we used for LLMs! 🤯

In this example, we induce a specific style (Art Nouveau 🎨), which we can accurately control with our λ parameter.

5/9 With Linear-AcT, we achieve great results in LLM 👿 toxicity mitigation and 👩🏼‍⚖️ truthfulness induction.

And the best result is always obtained at λ=1, as opposed to vector-based steering methods!

4/9 Linear-AcT preserves target distributions, with interpretable strength λ 🌈

🍰 All we need is two small sets of sentences {a},{b} from source and target distributions to estimate the Optimal Transport (OT) map 🚚

🚀 We linearize the map for speed/memory, thus ⭐Linear-AcT⭐

3/9 An activation has a different output distributions per behavior, eg. 🦠 toxic (source) and 😊 non-toxic (target). i) Vector-based AS moves activations OOD 🤯, with catastrophic consequences 💥 harming model utility. ii) The strength λ is unbounded and non-interpretable 🤨!

2/9 🤓 Activation Steering (AS) is a fast and cheap alternative for alignment/control.

Most AS techniques perform a vector addition such as a* = a + λv, where v is some estimated vector and λ the conditioning strength. How v is estimated differs for each method.

1/9 🤔 How do we currently align/control generative models?
- Pre-prompting
- Fine-tuning
- RLHF
However, these techniques can be slow/expensive! 🐢

Thrilled to share the latest work from our team at
@Apple
where we achieve interpretable and fine-grained control of LLMs and Diffusion models via Activation Transport 🔥

📄 arxiv.org/abs/2410.23054
🛠️ github.com/apple/ml-act

0/9 🧵

Thank you to the @neuripsconf.bsky.social for this recognition of the Generative Adversarial Nets paper published ten years ago with @ian-goodfellow.bsky.social, Jean Pouget-Abadie, @memimo.bsky.social, Bing Xu, David Warde-Farley, Sherjil Ozair and Aaron Courville.
blog.neurips.cc/2024/11/27/a...

Apple will be a platinum sponsor of the Deep Learning Barcelona Symposim 2024. This is the first time that Apple sponsors the event. #DLBCN

Bring stats to LM evals!!

open.substack.com/pub/desiriva...

Watching Frieren can’t stop thinking that demons are evil LLMs 😅