If you’re at ICML today and want to think about how sequence models memorize arbitrary facts, come by our poster for a few puzzles!🧩
📍Hall A #1611 🕔5PM
We uncover geometric memory: facts organize more like a map than a lookup table, even when simple associations seem enough.
At #ICML2026 ? 👀
If you're into scientific discovery with LLMs, check out our two posters:
- 🧬DecAEvolve: Decompose, Adapt, and Evolve — Poster #2202 , Jul 8 • 10:30 AM • Hall A
- 🧠Diverse Scientific Hypothesis Search — Poster #1815 , Jul 8 • 10:30 AM • Hall A
We're exploring how test-time training/adaptation and diversity shape open-ended discovery.
Sadly I couldn't make it because of visa issues, but my advisor @chandankreddy will be there! Stop by if you'd like to chat!!
17/ It is time to stop evaluating in the dark.
Come test your models on the whole elephant!
w/ Xiaobin Shen*,@jeremycweiss, George H. Chen
Paper: https://t.co/JjQOnl83Gx
Code: https://t.co/BvNvYEAZxW
Data: https://t.co/ar5k0vikaH
ICLR Page: https://t.co/VfLx4NOAfO
#ICLR2026 #CausalInference #SurvivalAnalysis #MachineLearning
1/ Excited to present our #ICLR2026 paper today at the poster session!
Come say hi: Pavilion 3 P3-#101
(10:30 AM – 1:00 PM 🇧🇷).
“SURVHTE-BENCH: A Benchmark for Heterogeneous Treatment Effect Estimation in Survival Analysis”
Here is why we built it, and why it matters 🧵👇
16/ TL;DR:
• We unify survival HTE methods into 3 families.
• We benchmark 53 variants on 40 synthetic, 10 semi-synthetic, and 2 real datasets.
• We explicitly test violations of ignorability, positivity, and ignorable censoring.
• No method wins everywhere.