Minxin Du ✈️ICML@Seoul

Research Assistant Professor@HK PolyU Research Interests: (Differential) Privacy, Applied Crypto, and AI Security

[Minxin Du ✈️ICML@Seoul @DuMinxin]

At #ICML2026, @XXiaoyu44761 and I will present "Unlearning Isn't Deletion," (https://t.co/WsEUNVoW8d) on reversibility risks in LLM unlearning.

Xiaoyu is also at #ACL2026, presenting BiForget: automated forget-set generation from seed queries.

Around Seoul/San Diego? Say hi!

[Minxin Du ✈️ICML@Seoul @DuMinxin]

@chien_eli Wow, see u in Seoul!

[Minxin Du ✈️ICML@Seoul @DuMinxin]

@yifaer_chen 发姐，TSLA😂

[Minxin Du ✈️ICML@Seoul @DuMinxin]

@xwang_lk @SimularAI Congrats!

[Minxin Du ✈️ICML@Seoul @DuMinxin]

@AndrewYNg @jyx_su Sounds like a "double-edged sword;"
it may also be "abused" by irresponsible reviewers

[Xiaoyu Xu @ACL ICML 2026 @XXiaoyu44761]

🚀 Excited to present our EMNLP 2025 work "OBLIVIATE: Robust and Practical Machine Unlearning for Large Language Models"!
I'll be at my poster session on Wednesday, Hall C3, 11:00-12:30 Session 2.
Come chat if you're interested in LLM unlearning, privacy! Check out our paper: https://t.co/yj1K4ogZGq

#EMNLP2025 #LLM #Unlearning #Privacy

[Minxin Du ✈️ICML@Seoul @DuMinxin]

@josephimperial_ Wow, welcome to Suzhou/China;
btw, where is p2 at?

[Minxin Du ✈️ICML@Seoul @DuMinxin]

@phli25 Wow, congrats! well-deserved!

[Minxin Du ✈️ICML@Seoul @DuMinxin]

@xwang_lk from Usenix Sec'25 PC chairs' message, "the extremes are concerning: six individuals appear as co-authors on 20 or more submissions, with two authors appearing on 36 and 39 submissions respectively. At such volume, it becomes difficult not to question the nature and depth.

[Minxin Du ✈️ICML@Seoul @DuMinxin]

@xwang_lk for big-4 security conferences, 6 for each cycle (typically 2 per yr);
for ICDE, 6 for both cycles...

[Minxin Du ✈️ICML@Seoul @DuMinxin]

@xwang_lk wow...# of submissions?/papers from US almost cuts a half...

[Xiang Yue @xiangyue96]

People are racing to push math reasoning performance in #LLMs—but have we really asked why? The common assumption is that improving math reasoning should transfer to broader capabilities in other domains. But is that actually true?

In our study (https://t.co/UhfB0ByLqi), we evaluated over 20 open-weight reasoning models and found that:
➡️Only models trained with RL exhibit broad transfer of math reasoning skills to other tasks.
➡️Models trained with SFT show limited or no transfer—especially to non-reasoning domains.

To quantify this, we introduce the Transferability Index (TI), which measures how much gain in math could transfer to others. A positive score indicates effective transfer; a negative one suggests loss of general capability.

We evaluate the models on three benchmark categories:
- Math reasoning: MATH-500, AIME24/25, Olympiad
- Other reasoning: GPQA-D (Science), LiveCodeBench2 (Code), ACPBench (Agent Planning), HeadQA (Medical)
- Non-reasoning: CoQA (Conversational QA), IFEval (Instruction Following), HalluEval (Hallucination), MC-TACO (Commonsense)

Our findings challenge the blind pursuit of leaderboard performance in math reasoning via SFT. Simply creating more math-like SFT data may inadvertently harm a model’s broader generalization. Instead, RL appears to be key for truly transferable reasoning development.

[Minxin Du ✈️ICML@Seoul @DuMinxin]

@niloofar_mire A simple Q:
It's a bit "weird" that the final step works on a weaker, much smaller LM; why not let a weaker LM output a "plan" as an input to a stronger LM?

[Minxin Du ✈️ICML@Seoul @DuMinxin]

RT @XXiaoyu44761: 🚀Unlearning Isn’t Deletion: Investigating Reversibility of LLM Unlearning 🚀

LLMs often inadvertently memorize sensitive…

[Minxin Du ✈️ICML@Seoul @DuMinxin]

@xwang_lk For my case, mostly correct, except some minor issues (e.g., highlighting not that "representative" work)

[Minxin Du ✈️ICML@Seoul @DuMinxin]

@chien_eli Congrats!

[Minxin Du ✈️ICML@Seoul @DuMinxin]

@QiangTang7 Glad to see you'll come to HK again :D

[Ning Zhang @n1ngzh]

Join the ACM CCS 2025 Artifact Evaluation Committee and play a vital role in advancing high-quality research artifacts and fostering reproducibility in security research. Nominate yourself or a colleague today using this form: https://t.co/9WHE3HJb9e.
@acm_ccs

[Nils Lukas @NilsLukas7]

🌟 ML Security & Privacy Researchers: I'm seeking exceptional Visiting Students, Postdoctoral Fellows, or Research Associates to work in our group at MBZUAI in Abu Dhabi.

Interested? You can find more information here: https://t.co/vV1TOqEc2i

[Minxin Du ✈️ICML@Seoul @DuMinxin]

@xuandongzhao @yuxiangw_cs @lileics thx a lot!! we'll check it