←
Daniel Scalena126 posts
Daniel Scalena avatar

Daniel Scalena

@daniel_sc4

interpretability et similia | prev: PhDed @unimib 🇮🇹 & @GroNlp 🇳🇱, intern @cohere

◷ Joined February 2015▧ 24 Photos and videos
Daniel Scalena avatar
Daniel Scalena@daniel_sc4·16 Oct 2025

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. 🧵👇

Post attachment
Daniel Scalena avatar
Daniel Scalena@daniel_sc4·Oct 25

I’m ending my PhD exactly how it started: going to @emnlpmeeting If you have any suggestions on how to avoid writing a thesis, happy to chat I’ll also be co-presenting the mega-important paper below ⤵️

Post attachment

Does entropy actually track what's important in a chain of thought? Our paper (accepted to the BlackBoxNLP repr. track) says mostly no, and we found the spurious cause behind why it looked like it did 🧵

Daniel Scalena avatar
Daniel Scalena@daniel_sc4·Oct 3

probably because writing style is incredibly rich and feature dense -> extremely complex to decode. We even tried, it was fun and tough

📈 But can models recognize and replicate individual translator styles?: ✓ Classifiers can find styles with high acc. (humans kinda don’t) ✓ Multi-shot prompting boosts style a lot ✓ We can detect strong style traces in activations (esp. mid layers) 3/

Post attachment
Daniel Scalena avatar
Daniel Scalena@daniel_sc4·Oct 3

astra/fable and co. may be good at hacking HF and winning prizes, but these models are extremely shit at writing I respect your time, so here’s the solution: style transfer. But I really want to see companies start optimising in this direction in the future

Opus 5.5 communicates more naturally, addressing some of the most common feedback we heard on Opus 5. It puts the most important information up front and follows the writing rules you give it, which makes long sessions easier to follow.

Post attachment
Daniel Scalena avatar
Daniel Scalena@daniel_sc4·Sep 9

agents are taking over swe tasks, math, biology and it just keeps going while we're losing ground on their monitorability (no, cot aint good). mech interp is undoubtedly cool at explaining narrow things about those agents but something in between is missing this is it:

Introducing Parallax 🔍 We're a new UK/EU-based nonprofit research lab building scalable methods and infrastructure to audit the beliefs, goals, and plans that shape model behaviours in long-horizon agentic evaluations. parallx.ai — Thread 🧵 1/

Post attachment
Daniel Scalena avatar
Daniel Scalena@daniel_sc4·Jul 6

EAGer mentioned

We’re thrilled to share that research from Cohere Labs and @Cohere will be presented at ICML this week! 🥳 📄ModernVBERT: Towards Smaller Visual Document Retrievers. Congrats to authors Paul Teiletche Quentin Mace, Max Conti, Antonio Loison, Gautier Viaud, @PierreColombo6, Manuel Faysse (arxiv.org/pdf/2510.01149) 🔎Investigating continual pretraining in large language models: Insights and implications. Congrats to authors Çağatay Yıldız, Nishaanth Kanna Ravichandran, Nitin Sharma, Matthias Bethge, @beyzaermis (openreview.net/pdf?id=aKjJoE…) ⌚EAGER: Entropy-Aware GEneRation for Adaptive Inference-Time Scaling. Congrats to authors @daniel_sc4, Leonidas Zotos, Elisabetta Fersini, and Malvina Nissim, @ahmetustun89 (arxiv.org/abs/2510.11170) 📈When AI Benchmarks Plateau: A Systematic Study of Benchmark Saturation. Congrats to all 35+ authors including Cohere's @YiyangNan (arxiv.org/abs/2602.16763)

Post attachment
Daniel Scalena avatar
Daniel Scalena@daniel_sc4·Jul 4

I’ll be presenting this work at ICML on Wednesday. If you’re curious, the paper is already available: arxiv.org/abs/2510.11170 Happy to chat about reasoning, RL, agents, interpretability, or anything adjacent!

Daniel Scalena avatar
Daniel Scalena@daniel_sc4·Jul 4

Heading to ICML 🇰🇷 and finding Seoul’s metro map a bit too branchy? Check out my metro poster instead! We branch only when an LM is uncertain, improving both performance and efficiency on reasoning tasks. (applying the same strategy to public transit is not guaranteed to help)

Post attachmentPost attachment
Daniel Scalena avatar
Daniel Scalena@daniel_sc4·Jul 1

I'd never have guessed models commit to their final answer this early, often within the first 20% of reasoning, across math/logic tasks and model families. The rest is mostly hedging that doesn't change their mind. And turns out they encode this internally, we can decode it! 🧵👇

Are all the CoT steps necessary? In our latest paper, we find evidence for the existence of a commitment boundary, marking a sharp transition from no/mid guesses to the model final answer across various reasoning tasks and model families. Thread 🧵👇

0:47
Cohere avatar
⟳ Daniel Scalena retweeted
Cohere@cohere·Jun 9

Introducing Cohere's first open-source coding model: North Mini Code Small & efficient, designed for agentic performance and built for community input.

6:10
Daniel Scalena avatar
Daniel Scalena@daniel_sc4·May 26

personal update: I'll be starting an internship at @cohere, working on code agents, one of the most interesting things happening in AI right now. feeling really grateful for this one and genuinely excited to see where it goes. here we go! 👀

Daniel Scalena avatar
Daniel Scalena@daniel_sc4·May 2

Really glad to see this work accepted at ICML 2026🎉 Can’t wait to present this work in Seoul!

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. 🧵👇

Post attachment
Daniel Scalena avatar
Daniel Scalena@daniel_sc4·Mar 26

Presenting this tomorrow at EACL2026, poster session at 9am If you’re around come say hi, happy to chat about the work and ideas More details in the thread 👇

📢 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/

Post attachment
Gabriele Sarti avatar
⟳ Daniel Scalena retweeted
Gabriele Sarti@gsarti_·Jan 9

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.

Post attachment
Gabriele Sarti avatar
⟳ Daniel Scalena retweeted
Gabriele Sarti@gsarti_·Jan 4

Now accepted at EACL main! Check it out! ⬇️

📢 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/

Post attachment