iiinova — where your digital self keeps evolving

CoNLL 2026 @conll-conf.bsky.social · 05/07/2026

Wesley Scivetti adds that we should remain open to all theoretical strands in a big-picture sense. Claire Bonial also mentions that anyway can feel free to propose novel tracks and directions for the community to take.

110

CoNLL 2026 @conll-conf.bsky.social · 05/07/2026

Paul Van Eecke argues that through construction grammar and usage-based approaches, new aspects of processing and language that fit CoNLL's aims are introduced and can become part of a shared vocabulary for talking about language.

110

CoNLL 2026 @conll-conf.bsky.social · 05/07/2026

Construction grammar and constructivism in language acquisition were also very prominent through both keynotes.

120

CoNLL 2026 @conll-conf.bsky.social · 05/07/2026

Finally: our community discussion session. Program chair Claire Bonial introduces this year's new tracks, "Language and the Brain" and "Usage-based approaches (incl. Construction Grammars)". Both will receive a steady home at CoNLL and SIGNLL.

140

CoNLL 2026 @conll-conf.bsky.social · 05/07/2026

Next up: Thomas Hikaru Clark!

110

CoNLL 2026 @conll-conf.bsky.social · 05/07/2026

And now - last but not least - our outstanding and best paper session!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

We’re not done yet! Our afternoon session is starting in 10 minutes 🎉

000

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

(No live tweeting, as your favourite publicity chair will be busy.)

110

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Coming up: our third oral session!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

And that's a wrap for our morning keynote!

000

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Q: Humans know words like "spatula" in the real world, LLMs do not. Could that be an influencing factor? A: Models did seem to know that spatulas are used for flipping, and they also seem to know more facts than humans do (although they do not have the "feel" of a spatula).

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Q: How do different pressures interact? Gender bias, length, accessibility, etc. A: For example, the male-first bias seems to come from accessibility (because social bias views men as more powerful etc.).

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Q: What is the role of the human's social situation in this? Humans are not socially expected to make weird analogies, they want to conform. A: Humans were instructed to be creative! (Nobody is offended in this context).

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Q: Are models more creative, or do we train them to be (with reward, optimization, etc.)? A: Models are not more creative, but way are constrained by working memory, processing factors, constraints, etc.

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

A: There is more variation in humans, but it remains open how constrained models actually are.

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Q: If it is true that models are more creative, but within the model space they converge on the same creative expressions, is it then fair to compare humans and models? Should model constraints be studied in isolation?

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Models might get way more text, but we get more information (grounding, Tomasello's work on paying attention). @adelegoldberg.bsky.social is still stunned that models learn from just text; humans would never learn from, e.g. radio input alone!

010

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Q: A high-level question: Large LMs show cool effects, small LMs are not as good. How does this relate to the idea of preemption? How do LMs do this, when they do not have expectations of a context? A: It remains open how large models need to be to be similar to humans.

200

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Q: Models got instructions for creative examples, would this influence humans as well? A: Humans got the some instructions! And still they performed the way they performed.

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

A: People who are bilingual still have the same judgements although they are fluent in multiple languages. Also, it would be surprising if frequencies for paired nouns would differ drastically between languages.

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Q: LLMs are also multilingual. What is the role of all the languages in the LLM data, when prompting them in English? When we treat them like English speakers, although they are massively multilingual!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Q: Sometimes humans can learn from hearing something once (independent of preemption/entrenchment etc.) A: Even word-level frequencies already influence preferences.

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Q: In a bilingual situation, how do competing alternatives work? A: Context-dependence! Child decides based on context, for example (but preemption works less cleanly with second-language learners).

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Q: Why are LLMs not so creative in the long form (AI slop)? A: AI slop is not well-defined, what is it supposed to be? Longer AI text is pretty good compared to the average human. A (from the audience): Individual texts look more creative, but between texts (/models) there is stylistic convergence.

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Q: For the novel conjunction rating experiment, how does recall factor into the experimental setup where nothing needs to be retrieved? A: That leads to noise in the results, but still people have to imagine what they prefer. And the results are still strong enough!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Q: How does the preemption explanation go with creative language use (like "She spatula'ed him"). A: Preemption is about hearing alternatives in contexts where sth. else is expected. For spatula'ed, we do not have an alternative to be expected, so language users have to think through what it means!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

That concludes the talk. Time for Q&A!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

In other words: LLMs are less led astray by what is most accessible. We use good-enough production.

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

More experiments where humans diverge from models due to accessibility constraints:

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

And also for very small GPT-2 models, preference differs from humans. Explanations? Maybe cognitive retrieval/accessibility.

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

For conventional cases, people and models agree. For novel items, they do not. Not even the largest models (Claude, GPT-4). This is surprising in light of other findings.

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

What about LLMs? For Qwen and Gemma-3, this is much more varied. They do not seem to show the same preferences as humans do!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

This extends to many types of novel items!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

When people are tasked to write down couples they know, they first utter the name that they are closer to. Length has no effect, but gender has (male-first bias)!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Finally: good-enough production. In binomial order (see and hear, now and then), we have clear preferences. (before > after, important > less important, here > there, primed > unprimed; the more accessible thing comes out first!)

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Another experiment by @kanishka.bsky.social, who trained models that witnessed verbs only ever in specific contexts (double object, prepositional object). Models treat models witnessed in non-preemptive contexts differently!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

This also works, for example, for verb usage. People generally prefer to use familiar formulations. But do models do the same?

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

One example: the ordinalizer _th is very productive (4th, 5th, 299th, gazillionth...). Why not 1th? We systematically hear "first" in context where a rule would suggest "1th"!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

It is statistical preemption!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

It is not direct feedback, which is usually unavailable or ignored by the kids learning language!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

So now we know that people AND LLMs can be creative with constructions. But how are conventional constraints learned?

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Humans came up with more concrete interpretations, like "slap". LLMs were generally more creative and abstract (when rated by humans after generating interpretations)!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Another question: Do LLMs understand the meaning of novel verbs in argument-structure constructions? Nouns can be used as verbs in transitive sentences. When LLMs are tasked to find meaning for sentences like "Ella spatula'ed Tom.", they interpret it like "flipping the situation/humans".

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Prototypical examples are nicely recoverable from the embedding space. The model models both form AND function!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

And speakers' choices often correlate with model preferences!

100

CoNLL 2026 @conll-conf.bsky.social · 04/07/2026

Speakers make choices when they use constructions, and these are grounded in differences in meaning (e.g. the direct object vs. prepositional object cxns, courtesy of @kanishka.bsky.social).

120