FigureAsia 35 Under 35 · Science
Aditi Raghunathan
Age 31 · Reliable machine learning · India / United States
Researcher exposing why next-token prediction can memorise rather than plan—and how alternative objectives change that behaviour.
- Approximate age at the edition eligibility date
- 31
- Field
- Computer science
- Country or region
- India / United States
- FigureAsia U35 Assessment
- 91.4 / 100
Profile
Career and documented record
Aditi Raghunathan studies the gap between model performance and dependable reasoning. Her senior-authored paper “Roll the Dice & Look Before You Leap,” recognised as an ICML 2025 Outstanding Paper, used controlled tasks to isolate a weakness in standard next-token training: it can reward short-sighted memorisation even when a problem demands planning.
The experiments showed that predicting multiple future tokens, including through diffusion-style objectives, can improve originality and planning under the studied conditions. The contribution is diagnostic as much as architectural. It gives the field a cleaner way to ask whether a model learned a generative process or merely reproduced its surface statistics.
Raghunathan's wider programme on robustness and trustworthy learning gives the work continuity. She is not offering a universal cure for reasoning; she is replacing vague claims with controlled evidence about what training objectives encourage.
FigureAsia selection
Why Aditi Raghunathan is on the list
Raghunathan earns her place for turning a diffuse concern about language models into an experimentally tractable scientific question. The 2025 result is influential because it identifies mechanism, tests alternatives and leaves its claims inside a clearly defined setting.
Verified work
The 2025–26 record
ICML Outstanding Paper
Senior-authored an award-winning study of planning and memorisation under different prediction objectives.
Controlled causal evidence
Used designed tasks to separate next-token imitation from learning a generative process.
Reliability programme
Connected the result to a broader agenda in robust and trustworthy machine learning.
Field context
The work in its field
Machine-learning reliability depends on understanding not only whether a model succeeds, but which training signal produced the behaviour and whether it transfers beyond familiar patterns.
FigureAsia U35 Assessment
Assessment breakdown
91.4out of 100
Substantive 2025–2026 contribution
17.9 / 20
Senior-authored an award-winning study of planning and memorisation under different prediction objectives.
Verified scientific impact
13.4 / 15
ICML's Outstanding Paper award and the study's clear diagnosis gave the work rapid influence in a central AI debate.
Originality and distinction
9.2 / 10
The distinction lies in controlled experiments that make the myopia of next-token training measurable rather than anecdotal.
Field influence
9.1 / 10
For Raghunathan, field influence turns on whether this work changes the operating baseline in reliable machine learning; the record supports that judgement.
Individual agency
9.2 / 10
Raghunathan is the senior author and intellectual lead, with experimental work credited to the full team.
Durability and trajectory
4.7 / 5
A continuing programme at Carnegie Mellon University extends beyond this single result.
Asian significance and global relevance
4.7 / 5
Indian scientist educated at IIT Madras and now leading machine-learning research in the United States.
Evidential validity and reproducibility
7.3 / 8
The paper uses controlled settings and explicit alternatives; claims are not extended to every language model.
Advance in scientific knowledge
6.5 / 7
It clarifies how prediction objectives shape memorisation, originality and planning behaviour.
Translational or methodological utility
4.7 / 5
The findings give model builders a principled basis for evaluating multi-token and diffusion-style objectives.
Responsible research stewardship
4.7 / 5
The work resists benchmark theatre by foregrounding mechanism, failure modes and the limits of transfer.