Sign in

Jane Li 🦖

@janeli.bsky.social
47 followers 40 following 21 posts

Postdoc at McGill Linguistics & Mila; prev JHU CogSci. Language production, morphology, & phonology!

PostsRepliesMedia
Reposted by Jane Li 🦖
Mal Shah @compositiomality.bsky.social · 30/09/2026
Accepted to Linguistic Inquiry! I argue on syntactic grounds “some of the apples” is really “some apples of the apples” with NP-ellipsis. This explains which quantifiers and which modifiers (adjectives, prepositional phrases, relative clauses) can’t appear. Preprint: ling.auf.net/lingbuzz/009...
ling.auf.net
44512
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
For similar work, we recommend Wang et al. (2026) arxiv.org/abs/2605.05197 and Kryvosheieva et al. (2025) arxiv.org/abs/2512.03676. Much exciting work is happening in this space and I’ve learned a lot working on this project. Yay! [10/10]
031
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
The Q of whether NLMs can have genuine distinctions of grammaticality is debated (Dentella et al., Hu et al. 2023). We believe the methods applied in this work acts as a complementary lens to the study of NLM gram. knowledge, and contribute a new perspective on these debates.
100
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
Check out the preprint here: arxiv.org/abs/2607.15175 Feedback most welcomed! [8/10]
arxiv.org
Linear representations of grammaticality in neural language models
Whether neural language models (NLMs) possess the ability to distinguish strings on the basis of their grammaticality remains a debated topic in the computational linguistics literature. Existing evid...
100
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
Studying under what data or architectural conditions we see detectability or generalizability in rep. space, and tying this to existing work on syntactic representations in NLMs is a clear next step. [7/10]
110
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
This is not to say that all models show these patterns (not what we find either). So far, we find that the number of parameters has a large effect, though of course this effect could be due to many other factors (of-interest) in model architecture that affect param count. [6/10]
100
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
Sentence representations can also encode far more than grammaticality, of course. We find by crossing string probability and grammaticality during probe fitting, that the representational separation is driven by grammaticality rather than probability. [5/10]
110
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
By fitting probes on one phenomenon (or a mixture of phenomena) and testing on others, we find that there is strong generalizability across phenomena. An analogous relationship can be found across different languages. [4/10]
100
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
We show that gram distinctions live in sentence representation space of a range of high-capacity models, detectable by low-complexity linear probes. They are: -Generalizable across phenomena and even langs -Deconfounded from string probs -Distinct from other sentential confounds
110
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
While string probs can serve as a marker for grammaticality distinctions (by consistent assignment of p(gram) > p(ungram)), probabilities encode far more than grammaticality! Even in min. pair settings, they differ on dimensions such as local prob, plausibility, etc. [2/10]
100
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
🦀New preprint! (w/ @najoung.bsky.social)🦞 Is grammaticality a major organizing principle of NLM representations? We show that many NLMs exhibit abstract rep. separation for grammaticality. We believe this work addresses debates about confounds in measuring model gram. knowledge. [1/10]
12010
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
For similar work, we recommend Wang et al. (2026) arxiv.org/abs/2605.05197 and Kryvosheieva et al. (2025) arxiv.org/abs/2512.03676. Much exciting work is happening in this space and I’ve learned a lot working on this project. Yay! [10/10]
000
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
The Q of whether NLMs can have genuine distinctions of grammaticality is debated (Dentella et al., Hu et al. 2023). We believe the methods applied in this work acts as a complementary lens to the study of NLM gram. knowledge, and contribute a new perspective on these debates.
100
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
Check out the preprint here: arxiv.org/abs/2607.15175 Feedback most welcomed! [8/10]
arxiv.org
Linear representations of grammaticality in neural language models
Whether neural language models (NLMs) possess the ability to distinguish strings on the basis of their grammaticality remains a debated topic in the computational linguistics literature. Existing evid...
100
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
Studying under what data or architectural conditions we see detectability or generalizability in rep. space, and tying this to existing work on syntactic representations in NLMs is a clear next step. [7/10]
100
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
This is not to say that all models show these patterns (not what we find either). So far, we find that the number of parameters has a large effect, though of course this effect could be due to many other factors (of-interest) in model architecture that affect param count. [6/10]
100
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
Sentence representations can also encode far more than grammaticality, of course. We find by crossing string probability and grammaticality during probe fitting, that the representational separation is driven by grammaticality rather than probability. [5/10]
100
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
By fitting probes on one phenomenon (or a mixture of phenomena) and testing on others, we find that there is strong generalizability across phenomena. An analogous relationship can be found across different languages. [4/10]
100
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
We show that gram distinctions live in sentence representation space of a range of high-capacity models, detectable by low-complexity linear probes. They are: -Generalizable across phenomena and even langs -Deconfounded from string probs -Distinct from other sentential confounds
100
Jane Li 🦖 @janeli.bsky.social · 17/09/2026
While string probs can serve as a marker for grammaticality distinctions (by consistent assignment of p(gram) > p(ungram)), probabilities encode far more than grammaticality! Even in min. pair settings, they differ on dimensions such as local prob, plausibility, etc. [2/10]
100
Reposted by Jane Li 🦖
grushaprasad.bsky.social @grushaprasad.bsky.social · 05/08/2026
Excited to share a preprint of work done in collaboration with @emilynguist.bsky.social ! osf.io/preprints/ps... TLDR: some effect sizes are larger in lab than on web-based platforms. So when evaluating quantitative predictions, considering the testing platform can be important! 1/n
2388
Reposted by Jane Li 🦖
Najoung Kim @najoung.bsky.social · 01/07/2026
🐸 New position paper on compositionality! 🐸 I synthesize my thoughts about the proper role/interpretation of behavior and mechanism in asking the question "Is this system (mind or machine) exhibiting compositionality?"
No Escape from Behavior in Evaluating Compositionality by Najoung Kim
1274
Reposted by Jane Li 🦖
Manasi Malik @manasimalik.bsky.social · 26/02/2026
Excited to share new work on how the brain makes social inferences from visual input! 🧠👯‍♂️ (With @lisik.bsky.social , @shariliu.bsky.social, @tianminshu.bsky.social , and Minjae Kim!) www.biorxiv.org/content/10.6...
15217
Jane Li 🦖 @janeli.bsky.social · 04/11/2025
Jenn is awesome!! Feel free to message me (or sign up for the mentoring program!) if you have any questions about our program/doing linguistics research at JHU CogSci from a trainee’s perspective🙂
020
Jane Li 🦖 @janeli.bsky.social · 26/12/2023
No snow this year in Port Coquitlam, BC but grateful for family and friends during tough times 💫 (photo taken last Christmas sledding 🛷)
A person in a flannel jacket looking out to the snow and forest
050