BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Hierarchical Reasoning Model: A Brain Inspired AI Framework for De
 ep Reasoning - Prof Meng Lu\, Peking University &amp\; Sapient Inc
DTSTART:20250707T150000Z
DTEND:20250707T160000Z
UID:TALK234142@talks.cam.ac.uk
CONTACT:Karen Scrivener
DESCRIPTION:One of the key features of the human brain is its deep reasoni
 ng ability derived from its structural dynamics. In contrast\, current art
 ificial intelligence models are constrained by fixed depth\, with a predet
 ermined number of layers that do not adapt based on the complexity of the 
 task at hand. Additionally\, the monolithic nature of their latent space r
 estricts the representation of diverse\, task-specific features\, further 
 impairing their reasoning flexibility. To tackle the current limitations o
 f reasoning in large language models\, we present the Hierarchical Reasoni
 ng Model (HRM)\, a novel architecture that sets new benchmarks in artifici
 al general intelligence (AGI) by leveraging neurocognitive principles of m
 odularity\, recurrence\, and synchronised rhythms. \n\nHRM consists of two
  modules that operate hierarchically and at different frequencies: the Fou
 ndational Module (FM) receives input and performs rapid\, iterative local 
 processing through multiple recurrent cycles\, while the Metamodule (MM) o
 perates at a slower temporal scale\, integrating abstract representations 
 and providing top-down strategic modulation to the FM. Empirical evaluatio
 ns demonstrate that HRM\, with 0.027B parameters\, significantly outperfor
 ms state-of-the-art models\, including current LLMs\, on complex reasoning
  tasks such as the Abstraction and Reasoning Corpus (40.3% in ARC-1 and 5%
  in ARC-2) and Sudoku (55%)\, using around 1000 training samples. These re
 sults highlight HRM’s ability to generalise from minimal data and execut
 e multi-step reasoning effectively. Model analysis demonstrates that the F
 M refines local features through iterative self-recurrence cycles in a way
  similar to sensory cortices in the brain\, while the MM acts like the pre
 frontal cortex\, providing strategic\, higher-order guidance and ensuring 
 convergence through slower\, more deliberate updates. Furthermore\, HRM ex
 hibits key emergent properties\, including self-correction\, functional se
 gregation\, and hierarchical dimensionality\, which enable dynamic explora
 tion\, efficient global reasoning\, and adaptive computation. In summary\,
  HRM integrates key brain-inspired principles to propose a compelling fram
 ework for AGI\, offering significant advancements in both generalization a
 nd reasoning.
LOCATION:Lecture Theatre 2\, Department of Chemical Engineering and Biotec
 hnology\, West Cambridge Site
END:VEVENT
END:VCALENDAR