Ziling Cheng (@ziling_cheng) header
Ziling Cheng (@ziling_cheng) profile picture

Research MSc @Mila_Quebec @mcgill_nlp | Research Fellow @RBCBorealis | reasoning and hallucination x evaluation and interpretability

Canada
Joined March 2020
Photo by Ziling Cheng (@ziling_cheng)Photo by Ziling Cheng (@ziling_cheng)Photo by Ziling Cheng (@ziling_cheng)Photo by Ziling Cheng (@ziling_cheng)Photo by Ziling Cheng (@ziling_cheng)Photo by Ziling Cheng (@ziling_cheng)Photo by Ziling Cheng (@ziling_cheng)Photo by Ziling Cheng (@ziling_cheng)Photo by Ziling Cheng (@ziling_cheng)Photo by Ziling Cheng (@ziling_cheng)
1/N Why do LLMs fail at math word problems without CoT? 🔍Final-answer accuracy alone doesn’t tell the full story. 💥It’s not that they can’t understand the problem, the major bottleneck is doing the calculations correctly. 📌Read our EMNLP main paper: aclanthology.org/2025.emnlp-…
1
6
10
1,234
Ziling Cheng retweeted
Much of business analysis and social science is extreme drudgery: a demographic researcher might need to pull data from hundreds of sources to study population trends, before starting the actual analysis! We formalize this overlooked step as SODIUM, where the open web is treated as a latent database. Existing state-of-the-art AI agents perform poorly on SODIUM tasks: the best performing agent (AG2) achieves only 46% of accuracy. Our SODIUM-Agent achieves over 90% accuracy, outperforming state-of-the-art AI agents by 2x-73x. 1/8
2
18
2
27
5,039
Ziling Cheng retweeted
I'll be attending #ICLR2026 next week to present my work on In-Context Algebra! My poster will be on Fri, April 24 at 3:15-5:45PM at Pavilion 4 P4-#4011. If you're around, stop by and say hello! My DMs are open if you want to connect or meet up in Rio!
Can you solve this algebra puzzle? 🧩 cb=c, ac=b, ab=? A small transformer can learn to solve problems like this! And since the letters don't have inherent meaning, this lets us study how context alone imparts meaning. Here's what we found:🧵⬇️
2
1
15
635
Ziling Cheng retweeted
New paper: Convergent Evolution: How Different Language Models Learn Similar Number Representations. Language models, classical word embeddings, and even raw token frequencies all develop the same Fourier features for numbers. But only some develop the underlying structure. 🧵
2
22
1
107
46,939
Latent CoT is an alternative LLM reasoning scheme hypothesized to enable “superposition” allowing models to hold uncertainty over multiple concepts during reasoning 💭 We revisit superposition in 3 latent CoT approaches and find that it is largely an illusion 🔮! More in 🧵
10
32
4
170
17,261
Ziling Cheng retweeted
At #EACL2026 to present our work teasing apart exemplar- and abstraction-first accounts of learning in LMs by looking at behaviour during training! (w/ @chrmanning) Paper: arxiv.org/abs/2603.17475 👀 Abstraction drives language learning! 🕐 Pavillon de Rabat, Apr. 25, 4:30PM
1
10
40
7,049
Ziling Cheng retweeted
first paper of the phd 🥳 the Superficial Alignment Hypothesis (SAH) argues that pre-training adds most of the knowledge to a model, and post-training merely surfaces it. however, this hypothesis has lacked a precise definition. we fix this.
9
46
8
248
42,458
Ziling Cheng retweeted
🚨New paper Are visual tokens going into an LLM interpretable 🤔 Existing methods (e.g. logit lens) and assumptions would lead you to think “not much”... We propose LatentLens and show that most visual tokens are interpretable across *all* layers 💡 Details 🧵
3
59
11
279
70,245
Ziling Cheng retweeted
🚀 Safer & less over-refusal LLMs without retraining? Yes, it’s possible. We need a good context. Here’s a simple but powerful idea for LLM safety without further training your models 👀🧠 (arxiv.org/abs/2512.11986)
2
11
34
2,403
One of the first papers I contributed to was published at #BlackboxNLP 📃, what an honor to, 6 years later, give a keynote at this wonderful venue 🧡 Thanks for having me! #EMNLP2025
Our second keynote of the day has just started! @vernadankers is now presenting "Memorization: Myth or Mystery?"
3
9
71
10,972
Make sure to go to Verna‘s keynote on memorization if you are at #EMNLP2025 🚨 You will leave with a different perspective on the topic, I promise 💭
Ready for day 3 of #EMNLP2025 🎉🎉 I've been on the lookout for memorization, unlearning, interp, memory module papers & more, chat w me if these topics fascinate you too😻 Looking forward to more of Suzhou, the conf & my BlackboxNLP keynote Sunday 1.45PM! blackboxnlp.github.io/2025/
2
8
1,218
This is happening to at board #20! Come chat about reasoning x interpretability ;)
Ever wonder why math reasoning fails without CoT? The bottleneck isn’t “abstraction” but “calculation”. I’ll be presenting at COLM on Friday (11–12) — come chat about reasoning × interpretability! @Mila_Quebec @RBCBorealis 🔗arxiv.org/pdf/2505.23701 #COLM2025 #EMNLP2025
1
5
394
Ready for day 3 of #EMNLP2025 🎉🎉 I've been on the lookout for memorization, unlearning, interp, memory module papers & more, chat w me if these topics fascinate you too😻 Looking forward to more of Suzhou, the conf & my BlackboxNLP keynote Sunday 1.45PM! blackboxnlp.github.io/2025/
12
1
55
5,276
7/N Can we steer model behaviours by injecting internal abstraction representations? Using cross-problem patching, we show that abstraction representations are transferable. For example, injecting the “subtraction” logic from P2 can turn the answer of P1 from 5 + 3 → 5 − 3.
2
1
91
8/8 Huge thanks to my collaborators @Meng_0209,@yanshuaicao and Leila Pishdad, and my advisor Jackie Cheung for their guidance and support throughout this work, and to friends at @Mila_Quebec @McGill_NLP. @RBCBorealis @emnlpmeeting
1
119
4/N Why do models fail without CoT? 🤔 Surprisingly, they perform better at abstraction than computation, showing that computation is the major bottleneck for final-answer accuracy. ‼️This implies that final-answer accuracy alone may be misleading for model’s reasoning abilities.
1
60