Raymond Chen

[Raymond Chen @RuishuoChen]

@Octmuseumpc @iclr_conf @jacobandreas @risteski_a @stellayu @guyvdb Haha I’d rather not assume bad intent 😳. But name collisions could be a real factor. Lots of Chinese names are identical in English, so maybe some profiles got mixed up? Hope it gets sorted out!

[Raymond Chen @RuishuoChen]

Has anyone else received an #ICLR2027 desk rejection related to profile or submission information that may be a misunderstanding? A co-author of ours got this notice, and all his submissions, including our joint paper, were desk rejected. We were notified on Oct 16 (~08:00 UTC).

[Raymond Chen @RuishuoChen]

We noticed a pattern: many affected authors have someone with the same name (but a different person) in their conflict-of-interest list on OpenReview. Since lots of Chinese names are identical in English, maybe that’s related? Just sharing in case it helps.

[Raymond Chen @RuishuoChen]

@Jerry_Life_Wu Thanks, but we don’t think so. Our co-author wasn’t involved at all, and at the time he even warned others not to take advantage of the leak. We’re still trying to figure out what was flagged.

[Raymond Chen @RuishuoChen]

He re-checked his OpenReview profile and couldn’t find anything inaccurate. He reached out to the PCs within two hours, and we understand they’re handling a huge workload. If you’ve had a similar experience, we’d be grateful to hear how it went 🙏 cc @iclr_conf

[Raymond Chen @RuishuoChen]

5/5 Routing improves with backbone capability. On SRA-Bench, even 0.6B GAVEL beats 32B progressive disclosure with a retrieved skill menu.
Building agents with growing skill libraries? We'd love to hear how you handle routing!
Paper: https://t.co/NCfwIpiImy

[Raymond Chen @RuishuoChen]

Can an LLM route skills without a skill catalog filling its context?
Introducing GAVEL ⚖️
The Router Within: Eliciting Native Skill Routing from a Frozen LLM
If you find it useful, an upvote on HF would mean a lot! 👍
https://t.co/uwMxOdobxh

[Raymond Chen @RuishuoChen]

4/5 In a minimal bash agent, Qwen3-32B + GAVEL loads the correct skill on 90.9% of Skill-Use tasks.
That exceeds the frontier-model Codex baselines reported in Skill-Use, whose best reaches 86.4%.

[Raymond Chen @RuishuoChen]

3/5 GAVEL leads on adjudicated routing accuracy across SkillRet, SRA-Bench and SkillRouter Eval-Core.
We also introduce SkillTraj: 372 simulated agent trajectories where skill needs arise amid noisy context. GAVEL leads across all four scenarios.

[Raymond Chen @RuishuoChen]

2/5 GAVEL uses two learned linear maps to read routing signals from a frozen LLM’s hidden states.
Glance shortlists skills. Verdict fuses the glance score with the same LLM’s likelihood and yes/no judgment as a product of experts.
Just 7.9M trained parameters on Qwen3-32B.

[Raymond Chen @RuishuoChen]

~50 posters at ICML 2026 were made with posterly.

Open-source skill, driven by any coding agent:
https://t.co/iyivO1AFGS

Site, blog, and a hosted service running my exact setup — no agent, don't trust yours, or don't want the hassle. Paid, receipt:
https://t.co/CyoPUzWJ6k

[Raymond Chen @RuishuoChen]

@CodexReleases When can I have any command like /rewind (claude code) in codex?

[Raymond Chen @RuishuoChen]

Is it art? Maybe debatable.

Is it usable? I think yes.

Open source here: https://t.co/iyivO1AFGS

Would love feedback, bug reports, and brutal academic-poster opinions.

#ICML2026 #AcademicPoster #LLM #ResearchTools

[Raymond Chen @RuishuoChen]

Conference prep season is here, and I was once again reminded that making academic posters is strangely painful.

Blank canvas? Hard.
Inherited template? Also hard.
AI-generated poster? Often messy: bad rendering, fuzzy figures, broken alignment.
So I made POSTERLY.

[Raymond Chen @RuishuoChen]

I tested it on two first-author ICML 2026 papers and one NeurIPS theory-ish project I contributed to.

Yes, the examples are all purple. NJU and Tsinghua both have purple-ish vibes, and I will not be taking questions on this aesthetic decision.