ICML 2026 Paper Introduces Trust Functions for Near-Lossless Weak-to-Strong Generalization

VIEWS20LIKES1REPLIES1

2/ #ICML2026 (Spotlight): Linguistic Nepotism: Shows that preferentially cite English sources even when better non-English evidence exists.

By @zoeykii, @MarineCarpuat, Paul McNamee, @EYangTW, Dawn Lawrie & @kevinduh. Paper: https://arxiv.org/abs/2509.13930

Original thread:

Dayeon (Zoey) Ki@zoeykii

1/ 🌍 Do #LLMs really treat all languages equally when citing evidence? 📑

In our new work, we uncover linguistic nepotism: models often trade off citation quality for language preference 👇

2h2010

Daniel Khashabi 🕊️@DanielKhashabi

4/ #ICML2026: Conformal Thinking reframes reasoning-budget selection as risk control w/ upper threshold stops when confident, a new lower threshold stops early on likely-unsolvable problems.

By @xidulu, @ZAlvin39105, @williamjurayj, @ben_vandurme, @MFarajtabar & @eric_nalisnick https://arxiv.org/abs/2602.03814

Xidulu@xidulu

1/ Our new #ICML paper targets two practical limitations of uncertainty-based early stopping for reasoning models: - Setting threshold values for uncertainty signals is hard. - "Stop when confident" principle fails to halt unsolvable problems

https://arxiv.org/abs/2602.03814

2h1400

Daniel Khashabi 🕊️@DanielKhashabi

3/ #ICML2026: We use early-exiting to cap how much corrupted/harmful in-context content can hurt LLM performance.

By @andrwynn, Metod Jazbec, Charith Peris, Rinat Khaziev, @anqi_liu33 & @eric_nalisnick. Paper: https://arxiv.org/abs/2510.02480

Original thread:

Andrea Wynn@andrwynn

Can we ensure that AI agents are safe, even with misleading & harmful inputs? -LLMs that use pre-trained knowledge when given task demonstrations are incorrect? -Agents that answer consistently even when given misleading or unhelpful information? 📰🧵: https://arxiv.org/abs/2510.02480 1/8

2h1300

Daniel Khashabi 🕊️@DanielKhashabi

5/ #ICML2026: VBVR - a massive suite (1M+ video clips, 200 tasks) for studying reasoning in video models

By Maijunxian Wang & @DengHokin (+50 more coauthors). Paper: https://arxiv.org/abs/2602.20159

Original thread:

Hokin Deng@DengHokin

#VideoReason We are open-sourcing the entire VBVR stack to speed-up the arrival of video reasoning as the next fundamental paradigm of intelligence

- 150+ synthetic generators - 1 million training clips - Cloud-scale data factory - Unified EvalKit - 100 rule-based evaluators - Strong baseline model

Checkout at http://video-reason.com

2h1300

Daniel Khashabi 🕊️@DanielKhashabi

6/ #ACL2026: arXiv2Table - a more realistic benchmark for LLM-based literature-review table generation.

By @MightyWeaver2, @JiefuOu, @yqsong & @ben_vandurme. Paper: https://arxiv.org/abs/2504.10284

2h12