[Colloquium] TODAY: 11/17 TTIC Colloquium: Kawin Ethayarajh, University of Chicago

[Mary Marre via Colloquium]

*When:*        Monday, November 17, 2025 at* 11:30** am CT *

*Where:       *Talk will be given *live, in-person* at

                   TTIC, 6045 S. Kenwood Avenue

                   5th Floor, Room 530


*Virtually:*  * Livestream only via Panopto
<https://uchicago.hosted.panopto.com/Panopto/Pages/Viewer.aspx?id=1cd9d913-2d6a-4479-9f2a-b392016b263b>*



*Who: *         Kawin Ethayarajh, University of Chicago



*Title:* We Post-Train Models Like They’re People; Should We?

*Abstract: *It is commonly thought that human sensibilities and preferences
are incorporated via the HF in RLHF. This is not entirely true—the RL is
also responsible. Drawing from Kahneman & Tversky’s Nobel-prize winning
work in behavioral economics, I will discuss how popular alignment methods:
(1) encode a model of utility similar to what humans have in prospect
theory; (2) even capture the subjective way in which humans perceive
probability. Moreover, by modifying the methods to more thoroughly capture
human inductive biases, we can post-train more efficiently, flexibly, and
performantly—even on verifiable tasks like mathematical reasoning where
human utility seems immaterial. Methods created this way, like KTO, have
had wide adoption in both academia and industry. There is no free lunch,
however: what might we be leaving on the table when even our methods mimic
people?

*Bio:* Kawin Ethayarajh is an Assistant Professor at UChicago Booth, where
he is the first core member of the Applied AI group. He works on
behavior-bound machine learning: understanding how AI is not only bound by
traditional factors like hardware and software, but also by the behavior of
real-world actors like workers, firms, and states. He is best known for
creating SHP, the first large-scale open-source dataset of human
preferences for post-training LLMs, and KTO, the industry standard for
aligning LLMs with class-imbalanced binary feedback. His work has been
recognized with an ICML 2022 Outstanding Paper award, a Facebook
Fellowship, and an NSERC PGS-D. Prior to UChicago, he graduated with a PhD
in CS from Stanford.

*Host: **Karen Livescu* <klivescu at ttic.edu>




Mary C. Marre
Faculty Administrative Support
*Toyota Technological Institute*
*6045 S. Kenwood Avenue, Rm 517*
*Chicago, IL  60637*
*773-834-1757*
*mmarre at ttic.edu <mmarre at ttic.edu>*


On Sun, Nov 16, 2025 at 1:55 PM Mary Marre <mmarre at ttic.edu> wrote:

> *When:*        Monday, November 17, 2025 at* 11:30** am CT *
>
> *Where:       *Talk will be given *live, in-person* at
>
>                    TTIC, 6045 S. Kenwood Avenue
>
>                    5th Floor, Room 530
>
>
> *Virtually:*  * Livestream only via Panopto
> <https://uchicago.hosted.panopto.com/Panopto/Pages/Viewer.aspx?id=1cd9d913-2d6a-4479-9f2a-b392016b263b>*
>
>
>
> *Who: *         Kawin Ethayarajh, University of Chicago
>
>
>
> *Title:* We Post-Train Models Like They’re People; Should We?
>
> *Abstract: *It is commonly thought that human sensibilities and
> preferences are incorporated via the HF in RLHF. This is not entirely
> true—the RL is also responsible. Drawing from Kahneman & Tversky’s
> Nobel-prize winning work in behavioral economics, I will discuss how
> popular alignment methods: (1) encode a model of utility similar to what
> humans have in prospect theory; (2) even capture the subjective way in
> which humans perceive probability. Moreover, by modifying the methods to
> more thoroughly capture human inductive biases, we can post-train more
> efficiently, flexibly, and performantly—even on verifiable tasks like
> mathematical reasoning where human utility seems immaterial. Methods
> created this way, like KTO, have had wide adoption in both academia and
> industry. There is no free lunch, however: what might we be leaving on the
> table when even our methods mimic people?
>
> *Bio:* Kawin Ethayarajh is an Assistant Professor at UChicago Booth,
> where he is the first core member of the Applied AI group. He works on
> behavior-bound machine learning: understanding how AI is not only bound by
> traditional factors like hardware and software, but also by the behavior of
> real-world actors like workers, firms, and states. He is best known for
> creating SHP, the first large-scale open-source dataset of human
> preferences for post-training LLMs, and KTO, the industry standard for
> aligning LLMs with class-imbalanced binary feedback. His work has been
> recognized with an ICML 2022 Outstanding Paper award, a Facebook
> Fellowship, and an NSERC PGS-D. Prior to UChicago, he graduated with a PhD
> in CS from Stanford.
>
> *Host: **Karen Livescu* <klivescu at ttic.edu>
>
>
>
>
>
> Mary C. Marre
> Faculty Administrative Support
> *Toyota Technological Institute*
> *6045 S. Kenwood Avenue, Rm 517*
> *Chicago, IL  60637*
> *773-834-1757*
> *mmarre at ttic.edu <mmarre at ttic.edu>*
>
>
> On Tue, Nov 11, 2025 at 5:19 PM Mary Marre <mmarre at ttic.edu> wrote:
>
>> *When:*        Monday, November 17, 2025 at* 11:30** am CT *
>>
>> *Where:       *Talk will be given *live, in-person* at
>>
>>                    TTIC, 6045 S. Kenwood Avenue
>>
>>                    5th Floor, Room 530
>>
>>
>> *Virtually:*  * tba*
>>
>>
>>
>> *Who: *         Kawin Ethayarajh, University of Chicago
>>
>>
>>
>> *Title:* We Post-Train Models Like They’re People; Should We?
>> *Abstract:* It is commonly thought that human sensibilities and
>> preferences are incorporated into generative models via the HF in RLHF.
>> This is not entirely true—the RL in RLHF is also responsible. Drawing
>> from Kahneman & Tversky’s Nobel-prize winning work in behavioral economics,
>> I will discuss how popular alignment methods: (1) encode a model of utility
>> similar to what humans have in prospect theory; (2) even capture the
>> subjective way in which humans perceive probability. Moreover, by modifying
>> the methods to more thoroughly capture human inductive biases, we can
>> post-train more efficiently, flexibly, and performantly—even on verifiable
>> tasks like mathematical reasoning where human utility seems immaterial.
>> Methods created this way, like KTO, have had wide adoption in both academia
>> and industry. There is no free lunch, however: what might we be leaving on
>> the table when even our methods mimc people?
>>
>> *Bio:* Kawin Ethayarajh is an Assistant Professor at UChicago Booth,
>> where he is the first core member of the Applied AI group. His works on *behavior-bound
>> machine learning*: understanding how AI is not only bound by traditional
>> factors like hardware and software, but also by the behavior of real-world
>> factors like workers, firms, and states. He is best known for creating SHP,
>> the first large-scale open-source dataset of human preferences for
>> post-training LLMs, and KTO, the industry standard for aligning LLMs with
>> class-imbalanced binary feedback. His work has been recognized with ICML
>> 2022 Outstanding Paper award, a Facebook Fellowship, and an NSERC PGS-D.
>> Prior to UChicago, he graduated with a PhD in CS from Stanford.
>> *Host: **Karen Livescu* <klivescu at ttic.edu>
>>
>>
>>
>> Mary C. Marre
>> Faculty Administrative Support
>> *Toyota Technological Institute*
>> *6045 S. Kenwood Avenue, Rm 517*
>> *Chicago, IL  60637*
>> *773-834-1757*
>> *mmarre at ttic.edu <mmarre at ttic.edu>*
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20251117/4c0a17d9/attachment-0001.html>