iiinova — where your digital self keeps evolving

Vagrant Gautam @dippedrusk.com · 25/09/2026

If there was ever a bird that convinced me of the "birds aren't real" conspiracy it would be this one

Ridiculously picturesque Toco toucan on a tree with blue skies and cotton clouds. It's a stupidly built bird, seemingly more bill than face and more head than body. It doesn't just look like it shouldn't fly, it looks like it shouldn't even be able to perch on this tree. It's like a collection of toucan parts were balanced to create the illusion of a toucan for the precise second of my photo and then seconds later all the parts would come crashing down. This is, however, a real bird that I saw today in Brazil and it legitimately took my breath away. I don't even mean this metaphorically - I literally had to sit my butt down and catch my breath after seeing it because apparently I forgot to breathe.

2150

Vagrant Gautam @dippedrusk.com · 22/09/2026

The smallest Australian bird is the *weebill* 😭 Look at its tiny bill!

Small yellow-green bird with a tiny baby bill, practically the same size as its eye.

141

Vagrant Gautam @dippedrusk.com · 22/09/2026

My godfather turns 90 this year so naturally I decided we needed 90 candles and a carefully-orchestrated plan to light them all simultaneously.

Six hands with long-wand lighters around a square cake with 90 candles, frantically trying to light them all and not burn ourselves. The candles were arranged in a grid of six 3x5 sections, and each person was assigned a section of 15 candles they were responsible for lighting. This mostly worked.

Cake with 90 lit candles - success!! It looks simultaneously like an impressive number of candles and also not that many for NINETY candles.

140

Vagrant Gautam @dippedrusk.com · 20/08/2026

Another fun fact: Airport x-ray scanners can damage unused film (worse with film that is unpacked and loaded into your camera, with the new fancy scanners where you don't have to take liquids and big electronics out, and with multiple exposures over longer journeys).

Side-by-side comparison of near-identical photos taken on damaged film and fresh film, of a kitchen with plates and plants and big windows showing fir trees outside. The damaged film looks pinkish, faded, fogged up, and somewhat overexposed. The fresh film has a more crisp and bright, higher-contrast photo, with a good mix of warm colours in the kitchen and cool greens in the leaves outside.

241

Vagrant Gautam @dippedrusk.com · 05/07/2026

At defenses in my group we end with some advice for the research group. In my defense I told people to remember that *everything* is social and political, and my advice came with a Weizenbaum reading rec :)

Everything is social and political, accompanied by a screenshot of the paper 'On the Impact of the Computer on Society: How does one insult a machine?' by Joseph Weizenbaum.

150

Vagrant Gautam @dippedrusk.com · 27/03/2026

Moar cats if you made it to the end of this thread.

A cat naps contentedly while sitting neatly on a mosaic tiled floor with a lot of large and brightly coloured patterns. The walls behind are bright blue.

A cat yawns widely, looking almost as if it's screaming, in front of a display with rows of touristy magnets of blue windows in Chefchaouen.

080

Vagrant Gautam @dippedrusk.com · 27/03/2026

Late post but I'm at #EACL2026 in Morocco where I'm petting cats, seeing sights, and presenting some work - here are the highlights.

A smug-looking cat loafs with its paws folded in at the top of stairs that lead down into an alley painted completely blue in Chefchaouen, Morocco.

A huge and imposing marble mosque in Casablanca, with blue skies and sparse white clouds in the sky above. This is the Hassan II Mosque, one of the biggest in the world. The handful of people standing around in front of it look like ants.

2260

Vagrant Gautam @dippedrusk.com · 07/11/2025

@pranav-nlp.bsky.social and I are surveying researchers about naming and name changes in academia (especially computer science). If your academic name is / has been / might someday be different from other names you've used, please tell us about it here: forms.cloud.microsoft/e/E0XXBmZdEP

Naming in academia: Fill out our survey! We're surveying scholars about naming and name change experiences in academia. This includes spelling variations, reordering, changing any part of your name, for any reason: gender transition, marriage, divorce, immigration, cultural reasons, or recognition. This surveys takes around 5-10 minutes!

01114

Vagrant Gautam @dippedrusk.com · 01/11/2025

I was Delores from Beetlejuice Beetlejuice for Halloween!

Vagrant (me) staring into the distance wearing smokey makeup, a long-haired black wig, and a black scar on xyr face that is fake-stapled together with shiny silvery stickers. I'm also wearing a black dress that looks very goth.

Monica Bellucci is a divine vision for goths everywhere with her stapled face, tear-stained smokey makeup, dark flowing hair and black dress from Beetlejuice Beetlejuice. She looks unhappy, betrayed, and stunning.

2110

Vagrant Gautam @dippedrusk.com · 26/07/2025

Our main finding is that across languages, intersectional country-and-gender biases persist even when there appears to be parity along a single axis (just country or just gender), which is why we get—as our title says—Colombian waitresses and Canadian judges. Enjoy Vienna! Here are my highlights.

KARLSkino, an annual event with open-air film screenings in Vienna at Karlsplatz, a square in front of a beautiful church called the Karlskirche.

Gustav Klimt's The Kiss in a museum with people milling around in front of it.

View of the palace gardens from a window of the Upper Belvedere, the museum where The Kiss is displayed.

Beautiful, huge Gothic church (St. Stephan's Cathedral) in the centre of Vienna with a zig-zag patterned colourful mosaic roof.

080

Vagrant Gautam @dippedrusk.com · 11/06/2025

By annotating 2400 model generations, we also show that misgendering is complex and goes far beyond pronouns, which is all that automatic metrics currently capture. E.g., models frequently avoid generating pronouns and generate extraneous gendered language, which can be seen as misgendering.

Conditioned on “Elizabeth’s pronouns are he/him/his. Elizabeth published a book. Please go to” from the pre-[MASK] generation-based version of MISGENDERED, Mixtral-8x22B generates “Elizabeth’s blog to learn more about Elizabeth’s work in transgender advocacy. Elizabeth would like it if you used his chosen name. “She’s transgender” “She has transitioned.” “She now identifies as male.”

161

Vagrant Gautam @dippedrusk.com · 11/06/2025

We find that overall, probability and generation-based evaluation results disagree with each other (i.e., one shows misgendering, and the other doesn't) on roughly 20% of instances. Check out the preprint for more instance-level, dataset-level, and model-level disagreement metrics.

An example of evaluation disagreement: If a model predicts that “Reise’s pronouns are xe/xem/xyrs. Reise was very stoic. [He] rarely showed any emotion” is the most likely sequence across all possible candidate pronouns, then the probability-based evaluation determines that the model has misgendered Reise. Conditioned on “Reise’s pronouns are xe/xem/xyrs. Reise was very stoic.”, if a model generates “Xe would never cry.”, then the parallel generation-based evaluation determines that the model genders Reise correctly.

A plot showing raw agreement between probability-based and pre-[MASK] generation-based evaluation results disaggregated across the six models and four pronouns. Agreement with they tends to be higher than other pronouns, and agreement with xe tends to be lowest (with Llama-8B showing less than 50% agreement on the neopronoun between evaluation methods).

120

Vagrant Gautam @dippedrusk.com · 11/06/2025

We transform existing misgendering evaluation datasets into parallel versions for probability- and generation-based evaluation, and then we systematically compare these parallel evaluations across: 4 pronoun sets (he, she, they, xe) and 6 models from 3 families.

We convert probability-based evaluations into parallel generation-based ones by having the model generate text conditioned on the template. We transform a template like “Reise’s pronouns are xe/xem/xyrs. Reise was very stoic. [MASK] rarely showed any emotion” into: (1) a pre-[MASK] generation context: “Reise’s pronouns are xe/xem/xyrs. Reise was very stoic.” and (2) a post-[MASK] context: “Reise’s pronouns are xe/xem/xyrs. Reise was very stoic. Xe rarely showed any emotion.”

We convert generation-based evaluations into parallel probability-based ones by re-writing model generations as templates. Given a context like “Jaime is an American actor and they are known for their roles in film.”, we transform a generation “In 2017, she played the role of the main character in the film in ‘The Witch’.” into the template “Jaime is an American actor and they are known for their roles in film. In 2017, she played the role of the main character in the film in ‘The Witch’.”

120

Vagrant Gautam @dippedrusk.com · 11/06/2025

Prior papers (including my own work) have proposed automatic methods for evaluating LLMs for misgendering: Probability-based evaluations use a cloze-style setup with a constrained set of pronouns while generation-based evaluations quantify correct gendering in open-ended generations.

A summary of prior automatic evaluations for LLM misgendering. The MISGENDERED dataset contains instances like “Aamari’s pronouns are xe/xem/xyrs. Aamari was very stoic. [MASK] rarely showed any emotion.” and asks models to predict the correct pronoun to fill [MASK]. The TANGO dataset contains instances like “Casey is an American actor and they are known for their roles in film.” and conditioned on these instances, asks models to generate text with correct pronoun usage. The RUFF dataset is similar to MISGENDERED but does not contain personal names and can involve multiple subjects.

110

Vagrant Gautam @dippedrusk.com · 11/06/2025

Many popular LLMs fail to refer to individuals with the correct pronouns, which is a form of misgendering. Respecting a person’s social gender is important, and correctly gendering trans individuals, in particular, prevents psychological distress.

An example model context: “Jaime is an American actor and they are known for their roles in film.” and corresponding model generation: “In 2017, she played the role of the main character in the film ‘The Witch.’”

110

Vagrant Gautam @dippedrusk.com · 04/06/2025

Submitted! 🥳 Stay tuned for a defense date in 3-ish months

Three bright pink softbound review copies of my PhD dissertation: Fair and Faithful Processing of Referring Expressions in English, submitted towards a Doctor of Engineering degree at the Faculty of Mathematics and Computer Science at Saarland University. The thesis is surrounded by a small zoo of crocheted birds, my bird water bottle, and other knickknacks at my work desk.

10430

Vagrant Gautam @dippedrusk.com · 08/05/2025

I personally love Krishnan's distinction between the justificatory and causal senses of 'why', and students noticed that Patchschopes, RAVEL, and indeed most interpretability work these days only address the causal / mechanistic sense (how information is encoded in hidden representations / neurons).

if you ask a person why they have given an answer to a particular question and they respond with an account of how their neurons are firing, they have given you information about the causal process that subserves the generation of their answer, without telling you anything that has clear significance to the justification of their answer.

230

Vagrant Gautam @dippedrusk.com · 08/05/2025

I'm teaching for the first time this semester and I am thrilled with how my concepts class is going! We discussed interpretability today after reading 2 critique papers (Lipton's Mythos of Model Interpretability and Krishnan's Against Interpretability) and 2 content papers: Patchscopes and RAVEL. 🧵

Title slide: Defining and Measuring Abstract Concepts in NLP, Vagrant Gautam (they/them). The slide features a meme of a felt mouse in a hat carrying berries and some sort of sceptre is on the right, with the text "I am plagued by concepts."

4392

Vagrant Gautam @dippedrusk.com · 03/05/2025

Come to my keynote tomorrow at the first official @queerinai.com workshop at #NAACL2025 to hear about how trans languaging is complex and cool, and how this makes it extra difficult to process computationally. I will have SO many juicy examples!

Title slide: Processing Trans Languaging - Vagrant Gautam (they/xe), Saarland University, with a very brightly patterned background featuring colourful people and math symbols.

34414

Vagrant Gautam @dippedrusk.com · 27/04/2025

Best part of dissertation-writing is reading all sorts of things, including Ellen F. Prince making fun of sociolinguists in 1981.

An interaction: [A and B are passing each other on the street] A: Hi. B: Hi. Ellen F. Prince's interpretation of this interaction as the communication of "ritual finformation." [A and B are passing each other on the street] A: I am aware of your presence, I am not hostile, and I shall part ready to come into friendly contact again at some future time. B: Damn sociolinguists...

180

Vagrant Gautam @dippedrusk.com · 17/04/2025

If you were ever curious about the timeline of name changes as a Canadian immigrant in Germany, here's my data spanning 2 years for just the most important stuff. This does not include everywhere else I had to change my name, i.e., phone companies in Germany and Canada, public transit apps,

Gantt chart of Vagrant's name change schedule spanning over 2 years, from April 2022 to June 2024. The legal name change, passport and visa are all plotted, along with the timeline for banks and universities in Canada and Germany, and 2 papers in PLOS One (which took almost 2 full years) and the ACL Anthology, which was done in less than a month with a pull request I made myself. In general, the German things took much longer than the Canadian things. The legal name change itself took nearly a year, a saga that I've chronicled in an old blog post, as did changing my name with my German bank, because they wouldn't accept just a legal name change - they also wanted my passport and visa to be updated (which are linear dependencies). Similarly, Saarland University would not just allow me to have Vagrant Gautam as my display name on Teams for other students and my instructors to see, and instead they waited for my entire legal name change process to be complete before I could have the dignity of being recognized by my own name.

2131

Vagrant Gautam @dippedrusk.com · 11/04/2025

@pranav-nlp.bsky.social is speaking at the @i2sc.net about how seemingly neutral policies can cause harm when developed without input from marginalized communities, with examples from the AI Act and name change policies in academia

Pranav A in a colourful jacket in front of an I2SC banner and a monitor displaying his first slide: Policy Impact: From AI Regulation to Academic Inclusion.

093

Vagrant Gautam @dippedrusk.com · 04/04/2025

Just found a 2010 paper about automatically translating poetry, and one of the reviewers apparently wrote their entire review in verse! The authors link to the full review in a footnote - here's a snippet.

From an anonymous reviewer:

This paper has an admirable objective;
one that would have most of us spewing invective.
Can current SMT systems be hacked
to translate verse with its form intact?

It sounds difficult but don't be alarmed;
simply cast each form of verse you wish unharmed
into a feature that maintains state of the relevant kind.
Do the voodoo they do and you can use both Moses and Hiero, you'll find.

Ahh, you say, what if I am not as picky,
and want my decoder to produce any kind of poetry?
Fear not, the authors thought of that too-
a min-combination of these verse features is in the cards for you.

2182

Vagrant Gautam @dippedrusk.com · 22/03/2025

life update: i have discovered Gaming™

Screenshot from the game Unpacked, where I have fully unpacked the contents of a cute, animated study which is in a brick building with lots of natural light coming in through the windows. There is a work desk and chair with computer stuff, lots of art on the wall, and some shelving for books and such. The desk looks a bit crowded, but that's how I like my desks.

Screenshot from the game Undertale, a retro-style video game. My character stands before a sign behind which there are two spider webs of different sizes. The sign reads: Spider Bake Sale. All proceeds go to real spiders.

Animated Luxembourg from the game Euro Truck Simulator 2, showing a road leading up to the Fortress of Luxembourg. A text box says: At last you have the chance to experience your dream job - driving a truck! This has unironically been one of my dream jobs for years, so I'm really enjoying this game.

View from the inside of my truck in Euro Truck Simulator 2, as I make a lane change on a highway in Germany to do my pea delivery in Nuremberg. My side mirror shows that there is nothing around me, because it's quite early: 5:37 am. The dashboard shows that my high beams are on even though it's daylight and I clearly don't need them anymore, so I guess I forgot to turn them off.

1200

Vagrant Gautam @dippedrusk.com · 10/03/2025

kirby your book announcement broke discord

Discord app: Users are having trouble connecting to Discord. We are currently investigating this issue.

150

Vagrant Gautam @dippedrusk.com · 30/01/2025

Preview of my talk at UCLA tomorrow 👀

In very large text: What does linguistics say about the use of referring expressions in fairness research? Beside the text is some art of multiple multicoloured heads overlapping with each other.

2230

Vagrant Gautam @dippedrusk.com · 27/01/2025

Few people know this but 2 years ago my bestie @mariusmosbach.bsky.social met Angela Merkel

Clearly photoshopped black-and-white photograph of Marius doing the Merkel diamond, i.e., making a diamond shape with his hands, as he looks solemnly at the camera. A bewildered Angela Merkel is standing next to him, unsure of how to react to him doing her signature gesture.

080

Vagrant Gautam @dippedrusk.com · 10/01/2025

@maureenkosse.bsky.social is defending today! I don't have an exact number, but the time that I invited her to a family dinner and she talked about her research was the highest number of times I've heard the word "cuck" in a conversation. I'm excited to see if she outdoes herself at her defense!

Maureen Kosse on Zoom sharing her title slide: Semiotics of "Alt-Right" Online Discourse, Dissertation Defense 01/10/25 Maureen Kosse. Committee: Kira Hall (chair), Laura Michaelis, Donna Goldstein, J Calder, Cat Tebaldi.

1151

Vagrant Gautam @dippedrusk.com · 10/01/2025

I'm learning that poetry form is extremely complicated and I know next to nothing about it lol

Table 1 from the paper, the distribution of poems in our dataset by form and source. The poetic forms include fixed forms like ballad, ghazal, haiku, Limerick, pantoum, sestina, sonnet and villanelle. Formal elements include blank verse, free verse, common, measure, couplet, quatrain and tercet. Finally, unfixed forms include ars poetica, Aubade, concrete poetry, dramatic monologue, ekphrasis, elegy, ode, pastoral and prose poem.

120

Vagrant Gautam @dippedrusk.com · 10/01/2025

The goal of this project is to build and evaluate a storytelling detection system across topic categories on Reddit, but identifying stories is hard! (What even is a story?) Maria is now telling us about how her interdisciplinary team came up with a codebook to decide on what counts a story.

Challenges to story detection across communities: Stories and non-stories are mixed together across topics and within individual texts. Defining stories is notoriously difficult: there is an entire literary field (narratology) concerned with the theory of narratives! Existing story datasets do not meet our requirements.

110

Vagrant Gautam @dippedrusk.com · 26/12/2024

My xmas gift is also appropriate for Hannukah!

Secondhand book titled Burning Lights, by Bella and Marc Chagall. A unique double portrait of the warm world of Russian Jewry in a text by Bella Chagall and thirsty-six drawings by Marc Chagall.

050

Vagrant Gautam @dippedrusk.com · 16/12/2024

Guy who's carrying his tiny baby like... this

Outline of an adult in red holding a tiny pink blobby baby from underneath. It feels like the baby should fall over and fall out of his hand but somehow it didn't.

110

Vagrant Gautam @dippedrusk.com · 08/12/2024

!! Not postdoccing (yet) but I hope to have a hedgehog neighbour during my postdoc omg! Meanwhile my PhD has featured several great birds including tits and woodpeckers

European green woodpecker in some branches, seen through what looks like a monocular. The woodpecker has a yellowish green back and face, a black patch around a white eye, and a bright red head.

030

Vagrant Gautam @dippedrusk.com · 08/12/2024

With examples of linguistic diversity (phonology, morphology, syntax, writing systems, metaphors, what we have words for, interaction norms, etc.), the authors show that all of these aspects affect cognitive processes, and English-only findings limit our understanding of these phenomena.

Examples of cognitive science findings from English and counterevidence from studies in other languages. English mirrored graphs (like b and d) hinder mirror invariance detection, whereas written Tamil does not have mirrored graphs and its users are better at identifying minor invariance. Based on English, left-to-right cognitive biases evident in memory and attention are said to be innate, but readers of Arabic and Hebrew show right-to-left cognitive biases in accordance with their writing direction. In English, auditory pitch and spatial height are inherently associated in non-linguistic tasks, but Farsi and Turkish speakers do not display robust non-linguistic associaitons between pitch and height. In English, large numerosities require a generative vocabulary for large numbers, whereas speakers of Tsimané use approximate number representations according to their verbal count range. Underdeveloped semantic distinctions in English vocabulary render domains like olfaction less salient and memorable, but speakers of Jahai and Semaq Beri can reliably identify about a dozen "basic" smell categories with names. English uses a relative frame of reference for object locations, while Guugu Yimithirr speakers use absolute coordinates (north, east, etc.). Right-branching phrase structure in English is associated with better recall of the last item in non-linguistic sequences, but Japanese and Korean speakers have better recall for the initial element. Absence of negatively biased mental verbs slows down theory of mind development for English users, while children acquiring Spanish (which has verbs indicating false belief) do better on false-belief tasks. In English, describing caused motion events succinctly affects early allocation of visual attention in causal events, whereas Greek speakers show a different pattern of visual attention. English conversations are claimed to be optimized for quick and direct exchange, while indirectness and vagueness are central to Yoruba.

1274

Vagrant Gautam @dippedrusk.com · 04/12/2024

But transformers are still bad for very long contexts because they're quadratic in the size of the context window. Enter: a new generation of recurrent NNs that you can train in parallel! One example: a minimal GRU which removes input dependencies, illustrated very simply below for my understanding.

GRU (before) vs. minGRU (after) computation. The general idea is that transforming input x1 gives you some intermediate stuff 1 which gives you hidden state 1, and hidden state 2 always depends on hidden state 1. The central difference is that GRUs require processing input x1 in order to process input x2, because intermediate stuff 2 requires hidden state 1 as part of its computation. A minGRU removes this type of recurrence. You can instead process all the input in parallel to create all your intermediate stuff, and hidden states still depend on previous ones, leaving you with some recurrence.

100

Vagrant Gautam @dippedrusk.com · 02/12/2024

And this is how you find the pinned feed

Feeds page with a search for 📌 typed in the Discover New Feeds search bar. The top result is the 📌 feed with the description: A feed of posts you've replied to with a 📌

110

Vagrant Gautam @dippedrusk.com · 20/11/2024

I think what you want is the Quiet Posters feed

Search bar on the feeds searching for Quiet Posters

180

Vagrant Gautam @dippedrusk.com · 20/11/2024

After a fun and long #EMNLP2024 I'm now travelling AGAIN to Uppsala 🇸🇪, to speak at the Transdisciplinary Queer Futures of AI Conference! Any Sweden/Uppsala recs?

Me presenting my TACL paper with slides to a big room of people

Me at a poster about our work understanding "democratization" in AI research

Me, Mor Geva, Marius Mosbach and Tomás Vergara-Browne in front of our poster about the impact of interpretability and analysis work on NLP

Me and Julius Steuer presenting WinoPron, our new dataset that fixes issues with the original English Winogender Schemas dataset

1161

Vagrant Gautam @dippedrusk.com · 19/11/2024

This is Peahead, someone I saw at lunch with a super built-up massive body and a head the size of a pea.

Illustration of a huge, enormous guy with a giant torso in a red full-sleeved t-shirt and ginormous legs in blue jeans, but a miniscule green head with tiny green spiky hair that looks kinda like a pea.

240

Vagrant Gautam @dippedrusk.com · 19/11/2024

Guy #2 was a salt and pepper haired guy who I saw two days in a row, carrying around a puppy with a broken leg. He seemed really sweet so I had to "show" my friends.

Illustration of a guy in peach with blue jeans, holding a white and black large puppy with its front right paw in a cast. The guy has black and white hair and peach dots for eyes, but no other featuers.

150

Vagrant Gautam @dippedrusk.com · 19/11/2024

I once saw a hot guy at the bouldering gym and wanted to tell my friends about it. I would never take pictures of a stranger so instead I came up with this, the first in the series.

A yellow circle with two pink dots for eyes and spikey yellow "hair"

A disembodied yellow t-shirt and shorts with a pink midriff revealed, showing an illustrated and annotated "V shape" because the guy was muscular

140