Tessera
Folgen

Sathvik

@sathvik.bsky.social

computational psycholinguistics @ umd he/him

62 Posts285 Follower278 Folgt
Iris van Rooij 💭@irisvanrooij.bsky.social · 4mo ago

Now reading 📖 Cruz, Nicole (2026). Illusions of Understanding from Outsourcing Thinking to LLMs. Computational Brain & Behavior doi.org/10.1007/s421... 1/🧵

Illusions of Understanding from Outsourcing Thinking to LLMs - Computational Brain & Behavior

Computational Brain & Behavior - Some illusions of understanding are an inevitable part of the research process, while others can be avoided or overcome by careful critical thinking and...

doi.org

Timnit Gebru@timnitgebru.blacksky.app · last mo.

Now is a good reminder that machine learning is a rebranding of statistical learning. Reinforcement learning is something I first came across in control theory. Many ML formulations are from information theory, a much more theoretically grounded discipline that's been around for decades.

Casey Newton@caseynewton.bsky.social · last mo.

There's a reason why "these are just statistical predictions" hasn't caught on in the discourse, and it's not the tech press. It's that people see them reading and writing and reasoning and scheming and they realize that the stochastic-parrots framing is outdated and inadequate to the moment

Cory Shain@coryshain.bsky.social · 2mo ago

Word predictability effects are LINEAR! And logarithmic. In a new preprint led by lab alum Stephanie Cho (w/ Ryan Buggy & Adrian Staub), we find clear evidence that the two patterns coexist. 1/

Title and abstract for "Preactivation and probabilistic inference coexist during sentence comprehension"
Sathvik@sathvik.bsky.social · 2mo ago

LM surprisal outperforms surprisal over human cloze responses as a predictor of reading times. Why is this the case? In our ACL paper, we standardize how to treat cloze surprisal, manipulate LM probabilities to reflect biases in the cloze task, and discuss how LMs and cloze data can be combined.

Sam Gershman@gershbrain.bsky.social · last yr.

The sycophantic tone of ChatGPT always sounded familiar, and then I recognized where I'd heard it before: author response letters to reviewer comments. "You're exactly right, that's a great point!" "Thank you so much for this insight!" Also how it always agrees even when it contradicts itself.

Lindia Tjuatja@lindiatjuatja.bsky.social · last yr.

When it comes to text prediction, where does one LM outperform another? If you've ever worked on LM evals, you know this question is a lot more complex than it seems. In our new #acl2025 paper, we developed a method to find fine-grained differences between LMs: 🧵1/9

Bild
Aniello De Santo@anids.bsky.social · last yr.

A bit late but since I really like this paper, a bit of self-advertising! I am presenting at CMCL today work showing that metrics measuring how a Minimalist Grammar parser modulates memory usage can help us model Self-paced reading data for SRC/ORC contrasts: aclanthology.org/2025.cmcl-1.5/

Capturing Online SRC/ORC Effort with Memory Measures from a Minimalist Parser

Aniello De Santo. Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics. 2025.

aclanthology.org

Kanishka Misra@kanishka.bsky.social · last yr.

another day another minicons update (potentially a significant one for psycholinguists?) "Word" scoring is now a thing! You just have to supply your own splitting function! pip install -U minicons for merriment

from minicons import scorer from nltk.tokenize import TweetTokenizer lm = scorer.IncrementalLMScorer("gpt2") # your own tokenizer function that returns a list of words # given some sentence input word_tokenizer = TweetTokenizer().tokenize # word scoring lm.word_score_tokenized( ["I was a matron in France", "I was a mat in France"], bos_token=True, # needed for GPT-2/Pythia and NOT needed for others tokenize_function=word_tokenizer, bow_correction=True, # Oh and Schuler correction surprisal=True, base_two=True ) ''' First word = -log_2 P(word | <beginning of text>) [[('I', 6.1522440910339355), ('was', 4.033324718475342), ('a', 4.879510402679443), ('matron', 17.611848831176758), ('in', 2.5804288387298584), ('France', 9.036953926086426)], [('I', 6.1522440910339355), ('was', 4.033324718475342), ('a', 4.879510402679443), ('mat', 19.385351181030273), ('in', 6.76780366897583), ('France', 10.574726104736328)]] '''
Sathvik@sathvik.bsky.social · 2y ago

I’ll be presenting two posters on (psycho)linguistically motivated perspectives on LM generalization at #EMNLP2024! 1. Sensitivity to Argument Roles - Session 2 & #BlackBoxNLP 2. Learning & Filler-Gap Dependencies - #CoNLL Excited to chat with other folks interested in compling x cogsci! papers⬇️