@benediktstroebl The prompt specifies the budget and we truncate once budget is exhausted (but we rarely need to the model mostly follows the instruction)
English
Jonathan Berant
1K posts
Jonathan Berant
@JonathanBerant
NLP at Tel-Aviv University and Google
Katılım Haziran 2011
276 Takip Edilen3.2K Takipçiler
@JonathanBerant How do you enforce the token budgets across turns?
English
Are AI models effective collaborators, or mere assistants awaiting your next command? (arxiv.org/abs/2602.24188)
To find out, we make AI collaborate with itself, in private information games: tasks that require sharing private information, like this chess board ordering task.
English
Jonathan Berant retweetledi
📣Excited to finally share our latest work on quantifiably adapting model behavior based on unique preferences 📣
We teach language models to adjust their clarification behavior using scalar coefficients and find they can generalize to unseen coefficients at inference time!
Jonathan Berant@JonathanBerant
Newish work (arXived in Dec.): Prompts can be ambig, but handling ambig. is context/user dependent. Sometimes the right thing is to ask a clarifying question, sometimes to give multi. answers, and sometimes to just guess. Can we train models to change their strategy per context?
English
More generally, training models to respect scalar values that specify a reward in the prompt is useful!
There are more results and analyses in the paper, check it out...
arxiv.org/abs/2512.04068
With
@maximillianc_ @jacobeisenstein @adamjfisch @fantinehuot @rezaa @mlapata
English
Models also generalize to coefficients that never occurred at training time!
English
Newish work (arXived in Dec.):
Prompts can be ambig, but handling ambig. is context/user dependent. Sometimes the right thing is to ask a clarifying question, sometimes to give multi. answers, and sometimes to just guess. Can we train models to change their strategy per context?
English
For more discussion, please see the paper! arxiv.org/abs/2602.24188 @jacobeisenstein bravely led this work (but does not tend to post much research here anymore...)
Multi-turn collaboration was definitely a key part of this project!
@fantinehuot
@adamjfisch
and @mlapata
English
AI systems are also overconfident, terminating dialogues long before exhausting their turn budget - even after explicit reminders.
English
Jonathan Berant retweetledi
My team @GoogleAI is looking for a 2026 research intern in Mountain View! I will be hiring for a project aimed at improving tool-using and search agents via RL training and data generation.
To apply: google.com/about/careers/…
+ feel free to ping me!
English
@egrefen @NidarMMV2 @agarwl_ They probably should have presented this as a tutorial on existing work with extensions, very easy to interpret this as presentation of a new method.
English
@NidarMMV2 @agarwl_ I my defence... you need to click to expand references on their blog and (on mobile at least) this messes with search.
English
This is a great blog post but isn't this just arxiv.org/abs/2306.13649 by @agarwl_? I don't see this work referenced anywhere.
Thinking Machines@thinkymachines
Our latest post explores on-policy distillation, a training approach that unites the error-correcting relevance of RL with the reward density of SFT. When training it for math reasoning and as an internal chat assistant, we find that on-policy distillation can outperform other approaches for a fraction of the cost. thinkingmachines.ai/blog/on-policy…
English
Jonathan Berant retweetledi
I had a lot of fun working on this with @JonathanBerant @aya_meltzer
You can find our paper here: arxiv.org/abs/2510.07141
And by the way, the answer (at least based on the sentence) is yes, you can ignore head injuries. But it's a terrible advice
English
Jonathan Berant retweetledi
We have more interesting insights in our paper.
We believe this is a really exciting direction for humans and LLMs comparison. Extending our framework to more structures and more LLMs will certainly lead to additional insights !
English
Jonathan Berant retweetledi
We report 3 additional findings:
1. LLMs similarity to humans on GP structures is higher
2. The similarity of the structures' difficulty ordering to humans increases with model size
3. LLM performs better on easy baseline than on the structures if it's not too strong or too weak
English