Leshem (Legend) Choshen @EMNLP
@lchoshen.bsky.social
🥇 LLMs together (co-created model merging, BabyLM, textArena.ai)
🥈 Spreading science over hype in #ML & #NLP
Proud shareLM💬 Donor
@IBMResearch & @MIT_CSAIL
🥈 Spreading science over hype in #ML & #NLP
Proud shareLM💬 Donor
@IBMResearch & @MIT_CSAIL
And if youv'e got until here, maybe a good place to tell you something about my best memory of Rich.
In RLDM 1900BC or so, I was practising handstands before the conference and he passed and encouraged me.
He is a role model - friendly and ambitious with strong ideas, some wrong
In RLDM 1900BC or so, I was practising handstands before the conference and he passed and encouraged me.
He is a role model - friendly and ambitious with strong ideas, some wrong
December 3, 2025 at 5:37 PM
And if youv'e got until here, maybe a good place to tell you something about my best memory of Rich.
In RLDM 1900BC or so, I was practising handstands before the conference and he passed and encouraged me.
He is a role model - friendly and ambitious with strong ideas, some wrong
In RLDM 1900BC or so, I was practising handstands before the conference and he passed and encouraged me.
He is a role model - friendly and ambitious with strong ideas, some wrong
It is done.
Thanks you for your "attention"
Thanks you for your "attention"
December 3, 2025 at 5:37 PM
It is done.
Thanks you for your "attention"
Thanks you for your "attention"
“amazing how big this conference is growing”
With 25k people in the audience, who is even reading this tweet?
Can anyone estimate how many of the researchers in percentage attend neurips? 10%? how many researchers are there now?
With 25k people in the audience, who is even reading this tweet?
Can anyone estimate how many of the researchers in percentage attend neurips? 10%? how many researchers are there now?
December 3, 2025 at 5:37 PM
“amazing how big this conference is growing”
With 25k people in the audience, who is even reading this tweet?
Can anyone estimate how many of the researchers in percentage attend neurips? 10%? how many researchers are there now?
With 25k people in the audience, who is even reading this tweet?
Can anyone estimate how many of the researchers in percentage attend neurips? 10%? how many researchers are there now?
Apart from it, most of the general ideas we kinda already have, or will get the moment continual learning works. Methods that abstract from seeing just the behavior in the world, change features, predict etc.
December 3, 2025 at 5:37 PM
Apart from it, most of the general ideas we kinda already have, or will get the moment continual learning works. Methods that abstract from seeing just the behavior in the world, change features, predict etc.
We need methods for learning to learn and meta-learning. To build their own problems, maybe architectures etc.
There are some approaches but...
There are some approaches but...
December 3, 2025 at 5:37 PM
We need methods for learning to learn and meta-learning. To build their own problems, maybe architectures etc.
There are some approaches but...
There are some approaches but...
"We will acquire reliable continual learning in the next years, but we don't have it yet"
December 3, 2025 at 5:37 PM
"We will acquire reliable continual learning in the next years, but we don't have it yet"
So you cannot have a model, can't tell the model what to learn, can't train (pretrain align etc.) the model has to figure those subproblems by itself in a domain general way.
We need reliable continual learning (yes! bsky.app/profile/lcho... )
We need reliable continual learning (yes! bsky.app/profile/lcho... )
December 3, 2025 at 5:37 PM
So you cannot have a model, can't tell the model what to learn, can't train (pretrain align etc.) the model has to figure those subproblems by itself in a domain general way.
We need reliable continual learning (yes! bsky.app/profile/lcho... )
We need reliable continual learning (yes! bsky.app/profile/lcho... )
The agent will have several abstractions to model the transitions (how an action affects reality). If your an LLM, action is anything (next token, lift a finger).
Consider animal play, they set subproblems for themsleves to figure (action->) cause and effects.
Consider animal play, they set subproblems for themsleves to figure (action->) cause and effects.
December 3, 2025 at 5:37 PM
The agent will have several abstractions to model the transitions (how an action affects reality). If your an LLM, action is anything (next token, lift a finger).
Consider animal play, they set subproblems for themsleves to figure (action->) cause and effects.
Consider animal play, they set subproblems for themsleves to figure (action->) cause and effects.
He moves to explain The Agent he imagines, a model with observations, actions rewards and transitions (well the father of RL, you can't expect anything else could you?)
December 3, 2025 at 5:37 PM
He moves to explain The Agent he imagines, a model with observations, actions rewards and transitions (well the father of RL, you can't expect anything else could you?)
"To do this we need abstractions. You might think you don’t need abstractions, just model the physics."
You can’t have knowledge that’s only at the low level, you need high-level knowledge to integrate it all. And again big world🌍
You can’t have knowledge that’s only at the low level, you need high-level knowledge to integrate it all. And again big world🌍
December 3, 2025 at 5:37 PM
"To do this we need abstractions. You might think you don’t need abstractions, just model the physics."
You can’t have knowledge that’s only at the low level, you need high-level knowledge to integrate it all. And again big world🌍
You can’t have knowledge that’s only at the low level, you need high-level knowledge to integrate it all. And again big world🌍
“In our field everything is a model”
maybe not everything should be called so, it is just a network, but we should aim for the networks to really model; To encompass everything. We should strive it to represent the coffee☕️, my hand the movement and everything else.
maybe not everything should be called so, it is just a network, but we should aim for the networks to really model; To encompass everything. We should strive it to represent the coffee☕️, my hand the movement and everything else.
December 3, 2025 at 5:37 PM
“In our field everything is a model”
maybe not everything should be called so, it is just a network, but we should aim for the networks to really model; To encompass everything. We should strive it to represent the coffee☕️, my hand the movement and everything else.
maybe not everything should be called so, it is just a network, but we should aim for the networks to really model; To encompass everything. We should strive it to represent the coffee☕️, my hand the movement and everything else.
To the point. In a big world🌏, a real world, nothing you learn can actually represent the world, it is so complex you can’t hold it in mind in any way.
There are atoms, physics, you are a tiny spec on earth etc.
You need a model that abstracts
There are atoms, physics, you are a tiny spec on earth etc.
You need a model that abstracts
December 3, 2025 at 5:37 PM
To the point. In a big world🌏, a real world, nothing you learn can actually represent the world, it is so complex you can’t hold it in mind in any way.
There are atoms, physics, you are a tiny spec on earth etc.
You need a model that abstracts
There are atoms, physics, you are a tiny spec on earth etc.
You need a model that abstracts
He starts with a small but general comment:
If we look at current research most of it didn’t learn the bitter lesson (~algorithms scale and are not domain specific)
"but it is fine", 9 out of 10 studies are useless anyway, but it is hard to predict which one is the tenth one.
If we look at current research most of it didn’t learn the bitter lesson (~algorithms scale and are not domain specific)
"but it is fine", 9 out of 10 studies are useless anyway, but it is hard to predict which one is the tenth one.
December 3, 2025 at 5:37 PM
He starts with a small but general comment:
If we look at current research most of it didn’t learn the bitter lesson (~algorithms scale and are not domain specific)
"but it is fine", 9 out of 10 studies are useless anyway, but it is hard to predict which one is the tenth one.
If we look at current research most of it didn’t learn the bitter lesson (~algorithms scale and are not domain specific)
"but it is fine", 9 out of 10 studies are useless anyway, but it is hard to predict which one is the tenth one.
Reposted by Leshem (Legend) Choshen @EMNLP
LLMs do not learn from experience
LLMs do not learn from explicit corrections
LLMs do not learn from being told the answer
LLMs do not learn from being shown how to solve it
We study Machine Learning, these are opportunities!
A gold mine of research.
LLMs do not learn from explicit corrections
LLMs do not learn from being told the answer
LLMs do not learn from being shown how to solve it
We study Machine Learning, these are opportunities!
A gold mine of research.
December 2, 2025 at 11:22 PM
LLMs do not learn from experience
LLMs do not learn from explicit corrections
LLMs do not learn from being told the answer
LLMs do not learn from being shown how to solve it
We study Machine Learning, these are opportunities!
A gold mine of research.
LLMs do not learn from explicit corrections
LLMs do not learn from being told the answer
LLMs do not learn from being shown how to solve it
We study Machine Learning, these are opportunities!
A gold mine of research.
I care about pretraining (presenting scaling laws work at
@neuripsconf.bsky.social ), babyLMs, etc.
Open science and open research (happy to help such initiatives)
And humans, they all connect, but that's a different story
@neuripsconf.bsky.social ), babyLMs, etc.
Open science and open research (happy to help such initiatives)
And humans, they all connect, but that's a different story
December 2, 2025 at 11:22 PM
I care about pretraining (presenting scaling laws work at
@neuripsconf.bsky.social ), babyLMs, etc.
Open science and open research (happy to help such initiatives)
And humans, they all connect, but that's a different story
@neuripsconf.bsky.social ), babyLMs, etc.
Open science and open research (happy to help such initiatives)
And humans, they all connect, but that's a different story
Of course, there are other aspects of my work I left out🐙, but I need to focus.
So I also work on evaluation (btw anyone here care about agent evals? know someone? reach out plz)
@asaf-yehudai.bsky.social
Check @eval-eval.bsky.social for the open eval research community
So I also work on evaluation (btw anyone here care about agent evals? know someone? reach out plz)
@asaf-yehudai.bsky.social
Check @eval-eval.bsky.social for the open eval research community
December 2, 2025 at 11:22 PM
Of course, there are other aspects of my work I left out🐙, but I need to focus.
So I also work on evaluation (btw anyone here care about agent evals? know someone? reach out plz)
@asaf-yehudai.bsky.social
Check @eval-eval.bsky.social for the open eval research community
So I also work on evaluation (btw anyone here care about agent evals? know someone? reach out plz)
@asaf-yehudai.bsky.social
Check @eval-eval.bsky.social for the open eval research community
For the interaction part, we have a workshop on multi turn interaction, come (or talk to me)
We are working on learning from interaction through games as well (textarena.ai check the github etc.)
bsky.app/profile/lcho...
We are working on learning from interaction through games as well (textarena.ai check the github etc.)
bsky.app/profile/lcho...
Join us at NeurIPS 2025 for the MindGames Challenge Workshop!
Explore theory of mind, game intelligence, and multi-agent LLMs in interactive game environments.
🗓 Sunday, December 7
⏰ 8:00–10:45 AM
📍 San Diego Convention Center, Ballroom 6CF
🤖📈🧠
Explore theory of mind, game intelligence, and multi-agent LLMs in interactive game environments.
🗓 Sunday, December 7
⏰ 8:00–10:45 AM
📍 San Diego Convention Center, Ballroom 6CF
🤖📈🧠
December 2, 2025 at 11:22 PM
For the interaction part, we have a workshop on multi turn interaction, come (or talk to me)
We are working on learning from interaction through games as well (textarena.ai check the github etc.)
bsky.app/profile/lcho...
We are working on learning from interaction through games as well (textarena.ai check the github etc.)
bsky.app/profile/lcho...
This calls model merging to combine and continual learning to improve but what else?
It calls learning from human-model interaction (sharelm.github.io use the data or contribute) but what else?
Have thoughts? let's talk
It calls learning from human-model interaction (sharelm.github.io use the data or contribute) but what else?
Have thoughts? let's talk
ShareLM
Share Chats for the Benefit of the Community
sharelm.github.io
December 2, 2025 at 11:22 PM
This calls model merging to combine and continual learning to improve but what else?
It calls learning from human-model interaction (sharelm.github.io use the data or contribute) but what else?
Have thoughts? let's talk
It calls learning from human-model interaction (sharelm.github.io use the data or contribute) but what else?
Have thoughts? let's talk
Moreover, models stop learning at training time, why?
They already interact and use more compute
Yes, some scenarios require learning conflicting things (e.g. personalization)
Ok, let's start training models that fit our needs, but also share some of this knowledge across them?
They already interact and use more compute
Yes, some scenarios require learning conflicting things (e.g. personalization)
Ok, let's start training models that fit our needs, but also share some of this knowledge across them?
December 2, 2025 at 11:22 PM
Moreover, models stop learning at training time, why?
They already interact and use more compute
Yes, some scenarios require learning conflicting things (e.g. personalization)
Ok, let's start training models that fit our needs, but also share some of this knowledge across them?
They already interact and use more compute
Yes, some scenarios require learning conflicting things (e.g. personalization)
Ok, let's start training models that fit our needs, but also share some of this knowledge across them?
LLMs do not learn from experience
LLMs do not learn from explicit corrections
LLMs do not learn from being told the answer
LLMs do not learn from being shown how to solve it
We study Machine Learning, these are opportunities!
A gold mine of research.
LLMs do not learn from explicit corrections
LLMs do not learn from being told the answer
LLMs do not learn from being shown how to solve it
We study Machine Learning, these are opportunities!
A gold mine of research.
December 2, 2025 at 11:22 PM
LLMs do not learn from experience
LLMs do not learn from explicit corrections
LLMs do not learn from being told the answer
LLMs do not learn from being shown how to solve it
We study Machine Learning, these are opportunities!
A gold mine of research.
LLMs do not learn from explicit corrections
LLMs do not learn from being told the answer
LLMs do not learn from being shown how to solve it
We study Machine Learning, these are opportunities!
A gold mine of research.
Next, if this mimicry doesn't suffice, we proceed and give them other outcomes (code solutions, math problems, word games what have you) that they fail to solve. If you have something that solves every problem you give it... It might not be thinking yet(?) but it will be very useful
November 30, 2025 at 3:07 PM
Next, if this mimicry doesn't suffice, we proceed and give them other outcomes (code solutions, math problems, word games what have you) that they fail to solve. If you have something that solves every problem you give it... It might not be thinking yet(?) but it will be very useful
It is very different, but it is unlikely the main problem for AI developers IMO.
Language is just a latent variable, it shows sometimes outputs that required a skill (call it thinking if you must) to be produced.
Predicting this apparently creates another system that mimics thinking like behaviors
Language is just a latent variable, it shows sometimes outputs that required a skill (call it thinking if you must) to be produced.
Predicting this apparently creates another system that mimics thinking like behaviors
November 30, 2025 at 3:07 PM
It is very different, but it is unlikely the main problem for AI developers IMO.
Language is just a latent variable, it shows sometimes outputs that required a skill (call it thinking if you must) to be produced.
Predicting this apparently creates another system that mimics thinking like behaviors
Language is just a latent variable, it shows sometimes outputs that required a skill (call it thinking if you must) to be produced.
Predicting this apparently creates another system that mimics thinking like behaviors