BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Memento EPFL//
BEGIN:VEVENT
SUMMARY:Reasoning via Flowing: Learning Computational Decompositions throu
 gh Reverse Corruption
DTSTART:20260825T150000
DTEND:20260825T170000
DTSTAMP:20261011T070641Z
UID:a2275d3ccae325b772dad60f0af07f059af1b45bca42796ddbbfa883
CATEGORIES:Conferences - Seminars
DESCRIPTION:Won Jun Kim\nEDIC candidacy exam\nExam president: Prof. Martin
  Schrimpf\nThesis advisor: Prof. Amir Zamir\nCo-examiner: Prof. Robert Wes
 t\n\nAbstract\nDifficult tasks often require multiple steps to reach the s
 olution. Mathematical proofs proceed through lemmas\, software is built fr
 om hierarchical abstractions\, proteins fold through transient conformatio
 ns\, and modern generative models synthesise images through sequences of i
 ntermediate representations. These intermediate states  progressively exp
 ose useful structure while reducing uncertainty about the solution. This t
 hesis investigates a strategy for systematic identification of useful inte
 rmediate representations for a given task. In particular\, we propose a le
 arning framework by asking the inverse question: how should information be
  progressively removed from a known solution until only the original probl
 em remains? We hypothesise that learning to reverse such corruption proces
 ses naturally induces trajectories whose intermediate states constitute ef
 fective computational decompositions. Flow and diffusion models provide th
 e mathematical machinery for investigating this hypothesis: these models l
 earn to reverse sequential corruptions that transform structured data into
  progressively higher-entropy noise\, thereby recovering the original samp
 le through a sequence of intermediate states. We extend this framework bey
 ond Euclidean data spaces by parametrising corruption processes over repre
 sentations that reflect informational\, rather than geometric\, progress t
 oward a solution. Under this view\, reverse trajectories become sequences 
 of progressively reconstructed information\, providing a principled framew
 ork for learning intermediate computations that resemble reasoning itself.
 \n\nSelected papers\n\n	Flow Matching for Generative Modeling (https://arx
 iv.org/abs/2210.02747)\n	Tree of Thoughts: Deliberate Problem Solving with
  Large Language Models (https://arxiv.org/abs/2305.10601)\n	Spatial Reason
 ing with Denoising Models (https://arxiv.org/abs/2502.21075).\n
LOCATION:BC 420
STATUS:CONFIRMED
END:VEVENT
END:VCALENDAR