Kanishka Misra

@kanishka

Assistant Professor of Linguistics at UT Austin. Works on computational understanding of language, concepts, and generalization. Aspiring wugologist! 🕸️👁️:

328 Following2,608 Followers377 Posts

Bluesky

Is this your profile?Claim it and kanishka.dirty.social is yours.Claim profile

Kanishka Misra@kanishka· 21h

Honored to be part of this really awesome lineup!

Core lecturers

Michael Biehl — University of Groningen
Caroline Rowland — Max Planck Institute for Psycholinguistics
Paul Smolensky — Microsoft Research & Johns Hopkins

Frontier lecturers

Afra Alishahi — Tilburg University
Marco Baroni — ICREA & Pompeu Fabra University
Francesco Cagnetta — Simplex
Alex Cristia — ENS & CNRS
Bruno Loureiro — ENS & CNRS
Kanishka Misra — UT Austin
Isabel Papadimitriou — University of British Columbia
Luca Saglietti — Bocconi University
Asad Sayeed — Gothenburg University
Larissa Samuelson — University of East Anglia
Kenny Smith — University of Edinburgh
Chris Summerfield — Oxford University

Kanishka Misra@kanishka· 21h

Tomorrow, I head to beautiful Gothenburg, Sweden, to lecture at Analytical Connectionism (www.analytical-connectionism.net/school/2026/). In a shocking turn of events, I will talk about controlled rearing -- a topic I have never before presented!

Title slide for the talk "The Role of Input in Shaping Linguistic Generalization: Insights from Controlled Rearing of Language Models"

Kanishka Misra@kanishka· Jul 15

Excited that our paper on minimal translation pairs for ASL was recognized with a best paper award at the Workshop on Generative AI for Sign Language @ CVPR 2026! Congrats to my co-authors @skarabuklu.bsky.social, Shester, Diane, Greg, and Karen! Paper: arxiv.org/abs/2604.27232

Certificate for best paper award from Generative AI for Sign Language workshop!

Kanishka Misra@kanishka· Jul 4

Doing my first ever conf-workshop keynote at the computational developmental linguistics workshop (comp-dev-ling.github.io)! Come one, come all!

Kanishka Misra@kanishka· Jun 29

I will be at #ACL2026 from July 2--7! I will be giving a keynote at CDL workshop on controlled rearing and hypothesis generation from language models! Tianyang Xu (first author) and I will present work on cross-modal generalization in VLMs on July 7! Paper: aclanthology.org/2026.acl-lon...

low effort slide on my activities during ACL.

Keynote at computational developmental linguistics workship (July 4, 3pm)

Poster w/ Tianyang Xu during Poster session G (July 7, 11-12)

Recruiting for my lab @ full conference!

Kanishka Misra@kanishka· Jun 23

#67!

Kanishka Misra@kanishka· Jun 16

No priming was observed today:

Reddit screenshot of a post saying “Trump is giving Iran $300 billion” referring to a tweet screenshot where the post says “Trump singing $300billion to Iran” like 10-12 times

Kanishka Misra@kanishka· May 11

New opinion piece on the interface between research on concepts and categories in minds vs. in neural network LMs! I take the position that there is much to be learned from this interface (e.g., learning about concepts from language alone) and outline some directions for future.

Title page of "Semantic Cognition for and from Language Models" followed by a figure showing tests that target conceptual structure and content vs. those that target function.

Kanishka Misra@kanishka· Apr 28

hmm

failure of the talkie model on the wug test and apparently also counting..

Kanishka Misra@kanishka· Apr 25

Seems from the CFP page that this (pre-selecting commit) is just about EMNLP, and not AACL, unless otherwise announced by AACL)

NEW ARR 2026 May submission policy for EMNLP: Given that the ARR 2026 May cycle is aligned with two conferences (EMNLP and AACL), authors will need to explicitly declare which conference they intend to commit to at submission time. This choice will be binding for EMNLP 2026: i.e., ARR 2026 May submissions that do not select EMNLP 2026 during submission will not be able to commit to EMNLP 2026. This policy does NOT apply to any papers submitted and reviewed in prior ARR cycles that wish to commit to EMNLP.

Kanishka Misra@kanishka· Apr 17

Perhaps the most important update to the manuscript is the strengthening of our initial hypotheses about how the harmonic alignment between discourse prominence and positional prominence affects cross-dative generalization:

Figure showing how “HAAP”, our graded measure of the extent to which a given exposure stimulus conforms to harmonic alignment effects and animacy prototypicality, ends up explaining LM generalization the best, resulting in the hypothesis that exposures conforming to HAAP will result in stronger cross-dative generalization in children.

Kanishka Misra@kanishka· Apr 17

Announcing a new version of our 2024 paper on linguistic hypothesis generation from LMs! @najoung.bsky.social and I have systematized our hypothesis generation framework, added stringent criteria for model selection, 10x-ed our learning trials, and included an epigraph from Jeff Elman 🙏!

Title page for the paper “A systematic framework for generating novel experimental hypotheses from language models”, with an epigraph from Jeff Elman describing how Rumelhart and McClelland (1982) did hypothesis generation with their connectionist network, and a figure describing our pipeline.

Kanishka Misra@kanishka· Mar 31

Excited to visit @gronlp.bsky.social (April 1) and @amsterdamnlp.bsky.social (April 2) to present a talk on something I have been thinking together with @najoung.bsky.social for the past 3 years!

Talk, titled "Hypothesis generation as a bridge between human and machine cognitive science", with artwork by Dr. Judy Fan, stanford

Kanishka Misra@kanishka· Mar 10

If models were generalizing arbitrarily, then we shouldn’t see any differences in their performance across these settings (i.e., no matter what, crow == bird). However, we find that models seem to only generalize when the training data preserves category coherence! 9/

Results from counterfactual shuffling experiments. Models tend to generalize equally well when the coherence was preserved and not so well when it was disrupted, even in the absence of all hypernyms.

Kanishka Misra@kanishka· Mar 10

By coherence we mean the visual similarity between members of the same category, which we calculate using the DINOv2 embeddings used in our VLM training. Even in the original configuration, we found models to perform better on categories that were visually more coherent 8/

Macro F1 scores on unseen images vs. Visual coherence across the 53 hypernym categories for the Qwen3-1.7B backbone (at 100% ablation). r (Pearson’s correlation) = .43, indicating positive relation.

Kanishka Misra@kanishka· Mar 10

To test this, we created counterfactual data: 1) where category-label pairings were shuffled across categories (🪛= “robin”; 🎸= “crow”) and 2) where they were shuffled within categories (🦅=“robin”; 🦜=“crow”). These swaps also manipulate the categories’ visual coherence 7/

Examples of image-leaf mappings resulting from our counterfactual shuffles, in comparison with the original configuration (top). VC indicates the visual coherence of the category under the data configuration. VC for birds in the original set: .30; for within-category shuffles: .30; for across-category shuffle: .12.

Kanishka Misra@kanishka· Mar 10

Are LMs simply executing something like “If crow THEN bird?” regardless of what the image shows? E.g., if during supervision we label images of kayaks as “crow” would the model still generalize to birds or does the model expect categories to have some level of coherence? 6/

figure depicting two hypotheses that models might entertain – 1) arbitrary prediction of hypernyms regardless of what the input looks like during supervision; 2) sensitivity to the fact that the category (e.g., birds) is not visually coherent.

Kanishka Misra@kanishka· Mar 10

Having established these preconditions to our task, we then find that models are also able to generalize (non-trivially) to hypernyms without ever having “seen” them explicitly, suggesting that LM representations support cross-modal generalization! 5/

Main results (see fig 4 in the paper). Salient result: models tend to generalize to hypernyms without any evidence encountered during training, suggesting that they show cross-modal generalization.

Kanishka Misra@kanishka· Mar 10

We establish that this paradigm works in the first place with a vision encoder that has never been trained on language data (i.e., ❌ SigLIP ✅DINO), that the models learn the task on the lower-level categories themselves, and that the LMs indeed have taxonomic knowledge 4/

left: Plot showing that models using the DINOv2 Encoder, which has never seen text information tend to generalize similar to those using the SigLIP encoder, which has seen text information. Right: Table showing that both Qwen3 LMs to demonstrate non-trivial hypernymy knowledge.

Kanishka Misra@kanishka· Mar 10

Taxonomic knowledge is interesting because of number of hypotheses about the learnability of category knowledge from linguistic cues, for both computational models and humans. Evidence of cross-modal generalization would lend strong support for these hypotheses! 3/

3 papers on hypernym acquisition in models (Hearst, 1992; Geffet and Dagan, 2005) and humans (Wilson et al., 2023) - see paper for details.

Kanishka Misra@kanishka· Mar 10

We use a VLM-training paradigm (frozen vision encoder w/o language training mapped to frozen LM) where we partially supervise on lower level categories during training, and then test if the LM recovers hypernymy knowledge from what it has seen in language data. 2/

Figure depicting an instance of our experiments. During training, the projector is deprived of explicit supervision on high-level categories (hypernyms, e.g., animal) at various amounts, and is trained to detect the presence (and absence) of lower-level categories (e.g., koala), keeping the image encoder and the LM backbone frozen. After training, the VLM is tested for generalization to hypernym categories, given previously unseen images.

Kanishka Misra@kanishka· Mar 10

What is the interplay between representations learned from (language) surface forms alone, and those learned from more grounded evidence (e.g.,vision)? Excited to share new work understanding “Cross-modal taxonomic generalization” in (V)LMs arxiv.org/abs/2603.07474 1/

title section of the paper: “Cross-Modal Taxonomic Generalization in (Vision) Language Models” by Tianyang Xu, Marcelo Sandoval-Castañeda, Karen Livescu, Greg Shakhnarovich, Kanishka Misra.

Kanishka Misra@kanishka· Feb 5

Nearly 2 years ago, @jessyjli.bsky.social, @janetlauyeung.bsky.social, @valentinapy.bsky.social, and I decided that it's time to bring discourse structure to the center of NLP teaching.

Title card of our paper: "Which course? Discourse! Teaching Discourse and Generation in the Era of LLMs" by Junyi Jessy Li, Yang Janet Liu, Valentina Pyatkin, and William Sheffield.

Kanishka Misra@kanishka· Jan 27

This is not due to surface form differences alone – PCA on the hidden-state representations of the three types of propositions alongside their (near) synonymous counterparts (e.g., all/every vs. certain/some / indef. article/generics) shows that they are distinguished based on inductive constraints!

left, examples of statements varying in the scope of proposition:

All bears have the T9 hormone [ALL]
Every bear has the T9 hormone [ALL]

Bears have the T9 hormone [GENERIC]
A bear has the T9 hormone [GENERIC]

Some bears have the T9 hormone [SOME]
Certain bears have the T9 hormone [SOME]

Right: plots showing PCA representations from 7 different layers in the Qwen3 VL 8B model, the 3 related propositions were clustered together, especially in layer 18.

Kanishka Misra@kanishka· Jan 27

Finally, we tested the two models that robustly satisfied both these preconditions on tests mimicking the original Gelman et al. (2002) study, and observed them to show qualitatively similar patterns as children and adults (all > generic > some).

Plots showing results across two models (Qwen 3 VL 4B and 8B) and across two modalities (Language and Vision+Language). Models generally showed all > generics > some.

Kanishka Misra@kanishka· Jan 27

For presupposition #2, we created a new developmentally inspired dataset to test if models can answer questions targeting "all" and "some" in contexts presented to them in multiple modalities (language/vision) - only Qwen3 VL 4B and 8B got high accuracies across conditions!

Left: table showing questions belonging to multiple conditions for all and some, e.g., a figure with 6 blue blocks with the question: "Are all blocks blue in color?";

Right: results showing once again that qwen models learn all vs. some quite well ~ nearly 100% accuracy across conditions.

Kanishka Misra@kanishka· Jan 27

For presupposition #1, we tested VLMs on their ability to robustly identify categories across multiple object-centric images under various negative sampling conditions:

Figure showing an instance from our category identification experiments, and results.

Left: a picture of an aardvark, with four types of questions: 1) "Is there an aardvark in this image?" (positive); 2) "Is there a cougar in this image?" (Similarity); 3) "Is there a panda in this image?" (Same Taxonomic Category); 4) "Is there a stained glass in this image?" (Random)

Right: Results, showing qwen 3 and 2.5 models being one of the best.

Kanishka Misra@kanishka· Jan 27

We replicate Gelman et al. (2002), where subjects were taught novel properties of animals under 3 diff conditions and then asked if a specific animal had that property: All (all Xs have…) Some (some Xs have…) Generic (Xs have…) Subjects showed a consistent pattern: all > generic > some

"Children's Use of Generics in Inductive Inferences" by Susan Gelman, Jon Star, and Jonathan Flukes + an arrow saying that this is one of my fav. experiments

Kanishka Misra@kanishka· Jan 27

“All bears have a property”, “Some bears have a property”, “Bears have a property” are different in terms of how the property is generalized to a specific bear – a great example of how language constrains thought! This holds for kids, adults, and according to our new work, (V)LMs! 🧵

Title page of our paper: "Bears, all bears, and some bears. Language Constraints on Language Models' Inductive Inferences"

Kanishka Misra@kanishka· Nov 5

I’ll be in Boston attending BUCLD this week — I won’t be presenting but I’ll be cheering on @najoung.bsky.social who will present at the prestigious SLD symposium about the awesome work by her group, including our work on LMs as hypotheses generators for language acquisition! 🤠👻