Current status of Samsung, which is loading the second deep chic shock ㄷㄷ

U+ 6:26 ‘;
4
Post
Google translation of the original English text
this is crazy
Samsung’s new Al model; DeepSeek and Gemini
10,000x smaller than 2.5 Pro, ARC-AGI 1 and
2 Yes, ahead
Samsung’s ultra-small recursive model (TRM) is a common
About 10,000 times smaller than LLM, but simply a text
More so because we think recursively instead of predicting.
It’s smart. First, draft your answer and then:
hidden for inference
Build a ‘scratchpad’ framework,
By repeatedly criticizing and improving the logical framework (up to 16 times),
At each cycle, we derive an improved answer:
This approach relies on architecture and inference loops (not just size).
It shows that (not) can boost intelligence:
Build a powerful and efficient model at low cost
Run, verify neural symbolic ideas, and much more
Apply the highest quality inference to a variety of application fields.
Let us help you:
Acceleration is everywhere
Post a reply

U+ 6:26 ‘;
0
22
Post
Less is More: Kecursive
‘Keasoning with
Networks
Alexia Jolicoeur-Martineau
Samsung SAIL Montreal
alexiajosamsung-com
Abstract
CrCr; ortrapy lo s
8
Hierarchical Reasoning Model (HRM) is
novel approach using twO small reflex net-
Fninren
Embcddirg
Works at different frequencies
This
8
biologically inspired method beats Large Lar-
Guage models (LLMs) on hard Puzzle tsks
such as Sudoku Maze, and ARC-AGI while
Add & Norm
trained with small models (27M parameters)
MLP
srmall data (~ I(N examples) HRM holds
great promise for solving hard problems with
Aid
Norm’
small nctworks
but it is not yet well
8
derstood and may be suboptimal
We Pro `
Sll wenvon
Pose
Rccursive Model
(IRM)
much
simpler recursive reasoning approach that
achieve significantly higher generalization
than HRM, while using
sinsle
network
Inpu (rain)
Prediction (yI
Lalenl (z
CJe SuD7I
IAnswerl
IRaasoninal
with only
layersWith only 7M Parameters;
TRM obtains 4596 test accuracy on ARC ACI
and 876 on ARC-AGI-2 higher than most
L.L.M.s
Decpseek Ri o3-mini Gemini 25
without than O01%6 of the parameters
Step
Urda’a
Given ^
mpyove (hd uloniri)
‘
1. Introduction
S unncve
‘Unapiedgwony?
While
powerful Large Language models (LLMs) can
struggle on hard question-answer
problemsg
Given
Anrloi Nsus
16 timof
miptorfy tnu Erudicton y)
that they Benerate their answer auto-regressively there
hichrsk of error sinice
siriele incorrect token can
Jackson Atkins
JacksonAtkinsx
19:00
My brain broke when
read this paper
tiny 7 million parameter model just beat
DeepSeek-R1, Gemini 2.5 pro, and 03-mini at
reasoning on both ARG-AGI1 and ARC-AGI 2
Less is More: Recursive
Reasoning with
Networks
Alexia Jolicoeur-Martineau
Samsung SAIL Montreal
alexia josamnsung comn
Abstract
‘Cross ontrcpy
8
Hierarchical Reasoning Model (HRM) is
novel approach using two small neural net-
FUvalgt
Emnboddirya
works recursively at different frequencies. This
8
biologically inspired meth(yd beats Large Lan)
gauge models (LLMs) on hard puzzle tasks
such ds Sudoku Maze; and ARC AGI while
Aid&NoT
trained with small models (27M parameters)
NLP
on small data (~ I0O exarnples) HRM holds
5
great promise for solving hard problems with
Aud
Nor
small networks,
but it i5 not yet well un-
S6l-Atonton
derstood and may be suboptimal
We Pro
7
U4
Tiv Recursive Mudel (TRM)
much
Post a reply
Tiny
Iiny
tny
(eB
Pro)
Itiig
Tiny

Key discovery: ultra-small Al, the amazing power of TRM
The core of this study is TRM (Tiny Recursive Model).
This is a new approach. TRM solves problems that are difficult for humans.
Like before, instead of giving the answer all at once, keep answering yourself.
‘Recursive’ that reviews and improves the reasoning process
Use method.

Overwhelming efficiency: TRM supports only 7 million parameters
Turtle is the latest giant with hundreds of billions of parameter names.
It is less than 0.0% of the size of the Al model:
Human-Level Reasoning Abilities: Sudogu’s Maze and Poetry
ARC-AGI, the ultimate test measuring pure reasoning ability
High-level performance similar to that of the benchmark; In the problem; TRM is the most
There is no performance that can overwhelm a large language model:
In the ARC-AGI-1 test, TRM achieved an accuracy of **44.69**
By recording the frame; Gemini 2.5 Pro, the world’s best model
(37.096) There is no going beyond the frame. 7.89 frame moon from ARC-AGI-2
Yours; This is also the Gemini 2.5 Pro (4.996) and the existing model
This figure surpasses HRM (5.09).