New preprint: arxiv.org/abs/2603.20133
Does performance on reasoning benchmarks transfer to real-world settings such as task-oriented dialogue?
Not necessarily: our new benchmark tests LLMs on problems framed in both standalone and dialogue settings, and shows that dialogue makes reasoning harder.
Endure: Mind body and the curiously elastic limits of human performance by Alex Hutchinson.
Good theory before to learn anything from any Dummies series book
Interesting did not know that arxiv requires peer-reviewed papers
arxiv.org/abs/2601.17036
Our Pulse News App -- our first app is live on iOS today.
We also announce star investors cap.
See the announcement video to see our plans.
If you are LLM guru at research or product building, join us to make the plans a reality
youtu.be/cA8-XJeoZAQ?...
Thank you to all coauthors to push the article to acceptance!
In general using LLM annotations for self-improvement is great future direction
Annotations using LLM are pretty under researched topics.
I would be interested how it performs on coding tasks and if it helps during CoT for coding tasks
Well done @zdenekkasner.bsky.social et al!
LLMs as Span Annotators: A Comparative Study of LLMs and Humans is accepted to multilingual-multicultural-evaluation.github.io 🎉
See paper arxiv.org/abs/2504.08697
What you plan to read on vacation?
Dr.LLM: Adaptive Layer Routing for Efficient Inference
Avoiding computation is pretty obvious way how to be more efficient.
The question is how to maintain quality. MoE and router is an efficient approach but suffer from exposure bias fixed thresholds.
arxiv.org/abs/2510.12773
so what it is?
What do you want to do with the deeper layers
Do you think future is MoE or dense models?
Qwen3 Embedder is amazing
but it is tricky to use.
Left padding confused lot of early adopters.
However, clearly it is the future. Reuse of pre-trained LLM for everything.
Including vector search where BERT-like models ruled.
1/N
arxiv.org/abs/2506.05176
huggingface.co/Qwen/Qwen3-E...
So would you like to prune them?
The no free lunch idea is here again.
Saw it in TTS models where Guided Attention worked very well. Idea is simple: For speech conversion prefer guide the attention to focus on narrow context. It learns faster.
Looking forward to deep dive on this one - expect similarities
arxiv.org/abs/2601.15165
It can be used in RAG, so for fetching relevant documents in clawbot it makes very much sense
The last token pooling is compressing the embeddings too much.
However it is engineering marvel.
Including tricks like the spherical last checkpoint averaging is awesome.
👀 WE ARE LOOKING FOR YOU! (yes, you!)
🚨NoCap test deadline 11/11/2025
🏆 $3K / $2K / $1K prizes
Our new team members shared tips&tricks which helped them to go through our test!
TRY IT NOW! 👉 github.com/BottleCapAI/...
The right words at the right time with the right style for the venue. Proud of our president 🇨🇿
youtu.be/d3wT84egi-g?...
Me: Scaling algorithms is just much more fun, than just burning electricity on more and more GPUs.
Official BottleCapAI: we believe AI shouldn’t cost tens of millions to train.
We are now opening challenge for those who want to help with that!
🏆 $3K / $2K / $1K prizes
⏰ Deadline: 11/11/2025
🖥️ 1 GPU. Your ideas.
👉 Join the NoCap Test:
github.com/BottleCapAI/...
📣Take part in 3rd Terminology shared task @WMT!📣
This year:
👉5 language pairs: EN->{ES, RU, DE, ZH},
👉2 tracks - sentence-level and doc-level translation,
👉authentic data from 2 domains: finance and IT!
www2.statmt.org/wmt25/termin...
Don't miss an opportunity - we only do it once in two years😏
I just found awesome channel on YouTube.
It has 77k subscribers with just 7 videos.
Because the videos are just awesome!
I wish I could explain stuff that simply!
Does anybody know how long take to prepare such video?
www.youtube.com/@algorithmic...
Thanks to all my awesome collaborators!
👉️ Vilém @zouharvi.bsky.social
👉️ Patrícia @patuchen.bsky.social
👉️ Ivan @ivankartac.bsky.social
👉️ Kristýna Onderková
👉️ Ondřej P. @oplatek.bsky.social
👉️ Dimitra @dimitrag.bsky.social
👉️ Saad @saad.me.uk
👉️ Ondřej D. @tuetschek.bsky.social
👉️ Simone Balloccu
How do LLMs compare to human crowdworkers in annotating text spans? 🧑🤖
And how can span annotation help us with evaluating texts?
Find out in our new paper: llm-span-annotators.github.io
Arxiv: arxiv.org/abs/2504.08697
We've been making the media rounds!
👉📺 @hajicjan.bsky.social talked about the new OpenEuroLLM project on Czech TV's Studio 6 www.ceskatelevize.cz/porady/10969...
👉📻 @tuetschek.bsky.social discussed #LLMs on Czech Radio radiozurnal.rozhlas.cz/proc-umela-i...).
