Indeed, but we also show the other side of the coin: personalized generation and its evaluation remain extremely challenging, and IMO professional human translators are still essential to produce a truly original, publication-ready final work as of today.
Want models to translate in the style you actually like?
Our paper is here! See you in Marocco! 🇲🇦
Takeaway: EAGer shows we can be MORE efficient & MORE effective by letting models focus compute where it matters most.
📄Paper: arxiv.org/abs/2510.11170
💻Code: github.com/DanielSc4/EA...
✨Huge thanks to my mentors and collaborators @leozotos.bsky.social E. Fersini @malvinanissim.bsky.social A. Üstün
Results: Across 3B-20B models, EAGer cuts budget by up to 80%, boosts perf 13% w/o labels & 37% w/ labels on AIME.
As M scales, EAGer consistently:
🚀 Achieves HIGHER Pass@k,
✂️ Uses FEWER tokens than baseline,
🕺 Shifts the Pareto frontier favorably across all tasks.
🧵5/
The fun part: EAGer-adapt reallocates saved budget to "saturating" prompts hitting the M cap, no labels needed! – Training & Verification-Free 🚀
Full EAGer uses labels to catch failing prompts, lowering threshold to branch or add sequences. Great for verifiable pipelines!
🧵4/
EAGer works by monitoring token entropy during generation. High entropy token → It branches to explore new paths (reusing prefixes). Token with low entropy → It continues a single path.
We cap at M sequences/prompt, saving budget on easy ones without regen. Training-free!
🧵3/
Why? Reasoning LLMs shine with CoTs, but full parallel sampling—generating multiple paths per prompt—is inefficient 😤.
It wastes compute on redundant, predictable tokens, esp. for easy prompts. Hard prompts need more exploration but get the same budget. Enter EAGER🧠!
🧵2/
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.
🧵👇
I’ll be attending the NEMI 2025 workshop this Friday and presenting a poster👇.
Happy to chat about cool interpretability stuff there!
📝 Paper: arxiv.org/abs/2505.16612
🔗 Code: github.com/DanielSc4/st...
Thanks to my amazing co-authors:
@gsarti.com , @arianna-bis.bsky.social , Elisabetta Fersini, @malvinanissim.bsky.social
7/7
🔍 What’s happening in the model?
We find that SAE steering and multi-shot prompting impact internal representations similarly, suggesting insight from user examples are summarized with extra interpretability potential (look at latents) and better efficiency (no long context) 6/
🌍 Across 7 languages, our SAE-based method matches or outperforms traditional prompting methods! Our method obtains better human-like translations (H) personalization accuracy (P), and maintains translation quality (Comet ☄️ @nunonmg.bsky.social) especially for smaller LLMs. 5/
💡 We compare prompting (zero and multi-shot + explanations) and inference-time interventions (ActAdd, REFT and SAEs).
Following SpARE (@yuzhaouoe.bsky.social @alessiodevoto.bsky.social), we propose ✨ contrastive SAE steering ✨ with mutual info to personalize literary MT by tuning latent features 4/
📈 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/
📘 Literary translation isn't just about accuracy, but also creatively conveying meaning across languages. But LLMs prompted for MT are very literal. Prompting & steering to the rescue!
Can we personalize LLM’s MT when few examples are available, without further tuning? 🔍 2/
📢 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/
Hellooo 👀
Hey hello! 👋
Now on 🦋!
Hello!
👋
It was great, I'm starting to get tickets for next year!
👀🙋♂️