Daniel Scalena

101 posts

Daniel Scalena banner
Daniel Scalena

Daniel Scalena

@daniel_sc4

PhDing @unimib 🇮🇹 & @GroNlp 🇳🇱, interpretability et similia

Katılım Şubat 2015
777 Takip Edilen134 Takipçiler
Sabitlenmiş Tweet
Daniel Scalena
Daniel Scalena@daniel_sc4·
You can easily save up to 65% of compute while improving performance on reasoning tasks 🤯 👀 Meet EAGer: We show that monitoring token-level uncertainty lets LLMs allocate compute dynamically - spending MORE on hard problems, LESS on easy ones. 🧵👇
Daniel Scalena tweet media
English
1
4
19
1.9K
Daniel Scalena
Daniel Scalena@daniel_sc4·
@GoodfireAI Nice work! I wonder, probe trained on answer choices needs known options. What if you probe model confidence and early exit there regardless of the answer it's thinking? I feel like after some t the model already knows and the rest is just overthinking
English
0
0
0
192
Goodfire
Goodfire@GoodfireAI·
LLMs often reason “performatively” well after deciding on a final answer - something that CoT monitors are slow to catch. Our new paper finds that: - probes can help monitor for this - it seems to track with task difficulty - probes enable early CoT exit, saving tokens! (1/7)
Goodfire tweet media
English
8
39
328
41.7K
Daniel Scalena
Daniel Scalena@daniel_sc4·
@paradigmainc Ok I was trying to cook something to improve model’s scientific creativity, throwing the repo into flywheel feels like the next logical step
English
0
0
4
172
Paradigma
Paradigma@paradigmainc·
introducing Flywheel: the infrastructure for autonomous research.
English
28
70
513
91.9K
Gabriele Sarti
Gabriele Sarti@gsarti_·
Surely a good omen, thanks Gemini
Gabriele Sarti tweet media
English
1
0
10
285
giulio
giulio@thelokasiffers·
I think the city of Rome vibecoded their public transport system
English
2
1
16
1K
Daniel Scalena retweetledi
Gabriele Sarti
Gabriele Sarti@gsarti_·
Happy to announce I will be mentoring a SPAR project this Spring! ✨Check out the programme and apply by Jan 14th to work with me on understanding and mitigating implicit personalization in LLMs, i.e. how models form hidden beliefs about users that shape their responses.
Gabriele Sarti tweet media
English
1
2
15
807
Daniel Scalena
Daniel Scalena@daniel_sc4·
Want models to translate in the style you actually like? Our paper just got accepted at EACL Main 🚀, check out our work on using interpretability for MT personalization! And, see you in Morocco! 🇲🇦
Daniel Scalena@daniel_sc4

📢 New paper: Applied interpretability 🤝 MT personalization! We steer LLM generations to mimic human translator styles on literary novels in 7 languages. 📚 SAE steering can beat few-shot prompting, leading to better personalization while maintaining quality. 🧵1/

English
0
0
3
253
Daniel Scalena
Daniel Scalena@daniel_sc4·
@thelokasiffers Woo big congrats on the launch and shoutout for starting in Rome. Can’t wait to hear more!
English
0
0
1
51
giulio
giulio@thelokasiffers·
Arriviamo
tensorqt@tensorqt

announcement: I will be founding a new company with @thelokasiffers and @EmanueleRodola. it seems very clear to us that we're on the verge of completely re-imagining many of the institutions humans have consolidated across history. one of these is the way we do, interpret, review and utilize science. we will be laying the foundations to build a breakthrough factory, and will approach the problem from a research-heavy perspective, while at the same time offering a product we hope many of you will use soon. crucially, we will do this in Europe, starting from Rome. over the next few weeks we will bring together a small number of angels to support our already well underway efforts, before making both our product and research available to the public. in the near future, we will also be expanding the team, with the specific purpose of building the single most talent-dense company in this space. if automating research taste sounds like an inevitable challenge, if you can feel that the current way we're doing science will change, if you think we don't have enough frontier labs in Europe, please reach out to any of us in DMs.

Italiano
5
1
62
9.4K
Daniel Scalena
Daniel Scalena@daniel_sc4·
@Turn_Trout @GladiaLab SVs are approximate directions in the latent space. They look for exact matches in the latent space. This could make things harder, but I’m still curious to know!
English
0
0
1
119
Alex Turner
Alex Turner@Turn_Trout·
@GladiaLab I'm curious what "prompt" you'd recover if you add steering vectors to a representation!
English
2
0
25
4.9K
GLADIA Research Lab
GLADIA Research Lab@GladiaLab·
LLMs are injective and invertible. In our new paper, we show that different prompts always map to different embeddings, and this property can be used to recover input tokens from individual embeddings in latent space. (1/6)
GLADIA Research Lab tweet media
English
280
1.3K
10.9K
5M
Cohere Labs
Cohere Labs@Cohere_Labs·
We are committed to making meaningful progress in machine learning research through open collaboration. Follow this 🧵to stay on top of our research contributions.
English
23
58
368
0
Daniel Scalena
Daniel Scalena@daniel_sc4·
You can easily save up to 65% of compute while improving performance on reasoning tasks 🤯 👀 Meet EAGer: We show that monitoring token-level uncertainty lets LLMs allocate compute dynamically - spending MORE on hard problems, LESS on easy ones. 🧵👇
Daniel Scalena tweet media
English
1
4
19
1.9K
Daniel Scalena
Daniel Scalena@daniel_sc4·
@thelokasiffers as a non citizen, best way to discover Rome ruins: getting lost through the ATAC connection graph
English
0
0
1
41
giulio
giulio@thelokasiffers·
a fun cope of living in a city devoid of proper mobility solutions is that each time you wanna go somewhere you get to play a game coming up with creative ways of getting to destination
English
5
0
12
700