Alan Ritter

@alan_ritter

Computing professor at Georgia Tech - NLP, ML, AI

Atlanta, GA

aritter.github.io

Replies Reposts Media

Alan Ritter
@alan_ritter
Jul 20
It’s great to see so much excitement around self-distillation! However, some recent work highlights an important failure mode: when the teacher is given the solution, it can suppress verification, backtracking, and exploration, which make reasoning models effective. We
Ethan Mendes
@EthanMendes3
Jul 5
Excited to be heading to ICML! I will be presenting our paper on rethinking how to train models with expert solutions using self-distillation. 🗓️ July 8, 2026, 5:00 PM – 6:45, Hall A #2502
3
Alan Ritter
@alan_ritter
Jun 22
Excited to attend #ICML2026 in Seoul 🇰🇷 together with my Ph.D. students: @EthanMendes3, @hyungjoochae, @CherylolGuo! We’ll be presenting work on: - Multilingual reasoning - Self-distillation from expert reasoning - Cloning synthetic web environments Poster info + paper links
ICML Conference and Georgia Tech Computing
2
Alan Ritter
@alan_ritter
Mar 25
📣 New paper by my Ph.D. student @hyungjoochae How can we build a safe, scalable learning environment for web agents? Paper: arxiv.org/abs/2603.10505 We introduce VERIENV, which clones real websites into executable synthetic environments using coding agents.
1
Alan Ritter
@alan_ritter
Feb 16
VLMs are surprisingly good at geolocating images, but do they respect people’s expectations about location privacy? Ray’s ICLR 2026 paper digs into this.
Ray Yang
@RuixinYang6
Feb 12
🚨 Excited to share that our paper "Do Vision-Language Models Respect Contextual Integrity in Location Disclosure?" is accepted to #ICLR 2026! Recent work shows that VLMs excel at geolocating images, posing serious privacy risks. However, privacy is not simply about refusing to
1
Alan Ritter
@alan_ritter
Feb 6
On the hardest problems—where we want models to learn the most -- RL gives zero signal, with no correct rollouts to learn from. Ethan’s recent work (@EthanMendes3) shows how models can still learn -- by leveraging expert solutions that are normally out-of-distribution for LRMs.
Ethan Mendes
@EthanMendes3
Feb 4
It's difficult to learn with RL on hard problems: (1) they can be non-verifiable, (2) models fail to generate correct completions that can be rewarded ❌ SFT on raw human solutions hurts performance 💡 We propose DAIL to effectively learn directly from high-quality solutions
1