📢 Generative models seem to acquire generation abilities more effectively than understanding, in contrast to human intelligence where generation is usually harder!
Excited to announce our very recent work on
🔥Generative AI Paradox!
paper:
Richard Feynman said “What I cannot create, I do not understand”💡
Generative Models CAN create, i.e. generate, but do they understand? Our 📣new work📣 finds that the answer might unintuitively be NO🚫 We call this the
💥Generative AI Paradox💥
paper:
Can small LMs perform planning with the same level of proficiency as
#LLMs
? Can they reason abt contextualized situations that are often counterfactual?🔥Introducing PlaSma, a 2-pronged approach endowing smaller models with procedural knowledge & counterfactual planning abilities
💫A life update: today is my first day
@allen_ai
as a postdoc, where I’ll be working with the awesome
@YejinChoinka
in
@ai2_mosaic
group!
Super excited to work and collaborate with all the amazing people here, and learn new things!
Can we efficiently tailor LLMs towards arbitrary user objectives without fine-tuning them?!
The answer is ✅yes with IPA which combines RL and inference-time techniques!
Come to our poster “Inference-time Policy Adapters” tomorrow Dec 10th at 9:00am-10:30am!
#emnlp2023
Still feel unreal, but a few weeks ago I successfully defended my Ph.D. thesis. I’m incredibly grateful to my advisor
@snigdhac25
. Thanks to my collaborators for making this happen! And big thanks to my family and all the people supporting this journey ♥️
💫 Excited to have 3 papers accepted to
#EMNLP2021
, ranging from conditional text generation for sentence ordering, character-centric narrative understanding, and implicit gender bias detection in narrative domain.
kudos to my amazing co-authors 🎉
Check out 👇🏼
Please be the voice of Iran!
The media doesn't show what's going on. Human disaster is happening in Iran.
Iranian people need Covid Vaccine!
Please retweet 🙏
@WHO
#sosiran
#sosiran
Wednesday was our anniversary and thanks to deadlines, I completely forgot about it!
My spouse (for the past year): I think I am the first author of our life paper
Me: 🫠
I will be at EMNLP ✈️🇸🇬 presenting 4 papers on reinforced algorithms to steer LMs generations, contextualize and defeasible machine reasoning!
Open to chat about research on LLMs, robust reasoning and generalization, language agents and other fun activities! Come and say hi 👋🏼
Come join us tomorrow at Narrative Understanding Workshop
#NAACL2022
to learn about interesting research.
We have an amazing set of speakers
@YejinChoinka
@maria_antoniak
@dangoldwasser
Andrew Piper.
We are in person at 708 Sol Duc and in zoom.
Can small LMs perform planning with the same level of proficiency as
#LLMs
? Can they reason abt contextualized situations that are often counterfactual?🔥Introducing PlaSma, a 2-pronged approach endowing smaller models with procedural knowledge & counterfactual planning abilities
✈️ I will be at
#EMNLP2022
in Abu Dhabi where my colleagues and I present several works at the main conference, Findings as well as the GEM workshop. Details below: 🧵(1/5)
📢Happy to share that our Generative AI Paradox has been accepted to
#ICLR2024
! 🤩
This is a joint work with all amazing people at
@allen_ai
and
@uwnlp
!
📢 Generative models seem to acquire generation abilities more effectively than understanding, in contrast to human intelligence where generation is usually harder!
Excited to announce our very recent work on
🔥Generative AI Paradox!
paper:
Can we efficiently tailor LLMs towards arbitrary user objectives without fine-tuning them?!
The answer is ✅yes with IPA which combines RL and inference-time techniques!
Come to our poster “Inference-time Policy Adapters” tomorrow Dec 10th at 9:00am-10:30am!
#emnlp2023
🚨AI-assisted writing is a nebulous space with researchers and writers being both curious & apprehensive. How useful are current SOTA LLMs to writers?What are their needs & expectations?If this intrigues you, then buckle up🚨
Paper:
#NLProc
#HCI
#ChatGPT
📣I am not attending Neurips but go talk to amazing
@billyuchenlin
about our SwiftSage, a powerful language agent for complex and situated long-horizon planning inspired by human dual process theory 🧠
⏱️: spotlight poster session
@12
/13 5Pm
[1] SwiftSage: A Generative Agent with Fast and Slow Thinking for Complex Interactive Tasks
📄:
🖥️:
🗣️: Spotlight Poster Session (12/13 5PM) @ Great Hall & Hall B1+B2 (level 1)
#405
Congratulations to the team, including AI2ers, who worked on "Do Androids Laugh at Electric Sheep? Humor 'Understanding' Benchmarks from The New Yorker Caption Contest" — selected for a Best Paper Award at
#ACL2023
!
💫 If you are working on multilingual/cross-lingual domains, please check our new TACL paper:
"ParsiNLU: A Suite of Language Understanding Challenges for Persian"
Excited that our big collaborative effort, "ParsiNLU: A Suite of Language Understanding Challenges for Persian" will appear in TACL'21!
If you're working on multilingual/cross-lingual NLP, give it a look!
Paper:
#EMNLP2020
Check out our paper on Emotion-Aware Storytelling 🙂☹️😟😡
Join us on @ gather session:
When: Nov 17, 18:00-20:00 UTC / 10:00-12:00 PST
Paper:
Video:
With
@snigdhac25
📢Stop by our
#ACL2023NLP
poster on Monday poster session 1 from 14:00-15:30 at Frontenac Ballroom to discuss information theoretic evaluation of rationales
Existing metrics for free-text rationales narrowly focus on association with labels, offering little understanding of new information contained in the rationale to explain the label! We overcome this by introducing REV!
💫By our amazing ex-intern
@hanjie_chen
📢Submissions are open for the 3rd Workshop on Narrative Understanding. The workshop will be co-located with
#naacl2021
on June 11.
Submission deadline: 22 March 2021
More info:
Submission link:
#UWAllen
@uwnlp
's
@YejinChoinka
aims to develop
#AI
with the ability to reason and communicate about the world in physical and abstract terms, like humans can do. As a 2022
#MacFellow
, she looks forward to taking the “adventurous route” in her research:
📣Seventh Workshop on Computational Linguistics and Clinical Psychology (CLPsych)
📣The Third Workshop on Narrative Understanding
📣The Third Workshop on Privacy in NLP
(7/n)
#NAACL2021
I’m quite used to the cruelty students can face when they apply for a US visa but this one broke me. We offered admission to a stellar, talented & hardworking student. After months of work and hundreds of dollars, an embassy officer saw him for 5 mins & said no. why? …
Existing metrics for free-text rationales narrowly focus on association with labels, offering little understanding of new information contained in the rationale to explain the label! We overcome this by introducing REV!
💫By our amazing ex-intern
@hanjie_chen
(3) Uncovering Implicit Gender Bias in Narratives through Commonsense Inference. At Findings.
Special congrats to
@DanielH68771331
for his first paper acceptance! 🎉
W/
@VeredShwartz
,
@snigdhac25
We are presenting our work on Sentence Ordering using ReBART in half an hour with
@SomnathBrc
at the Virtual poster session II: Semantics (12:30-14:30 AST). Come say hi if you are around 🤩
📹…
#EMNLP2021
@snigdhac25
(1) Is Everything in Order? A Simple Way to Order Sentences. Accepted at Main.
Proposed a conditional text-to-marker generation framework, making the problem more tractable, and less prone to neural text degeneration.
W/
@SomnathBrc
and
@snigdhac25
I bought this little cute plant on the left, and it turned out to be the one on the right in 2 months! What’s wrong with you dude? 😳🤪
We just decided to plant it somewhere one midnight and just flee 🏃🏼♀️
Our 3rd contributed talk is Towards Emotion-Aware Storytelling Using Reinforcement Learning by
@faeze_brh
and
@snigdhac25
. Generating stories about a topic with a predefined emotional arc.
We also introduce Counterfactual Planning, a novel task that requires plan revision in response to counterfactual situations. 📊Our experiments show orders-of-magnitude smaller models (770M-11B parameters) can compete and often surpass their larger teacher models' capabilities!
🚀With PlaSma, we developed symbolic procedural knowledge distillation to enhance implicit knowledge in small language models and an inference-time algorithm for more structured and accurate reasoning!
Today we're thrilled to announce our new undertaking to collaboratively build the best open language model in the world: AI2 OLMo.
Uniquely open, 70B parameters, coming early 2024 – join us!
Combining the power of behavior cloning and LLMs takes grounded action planning to a new level!
Read our new paper to discover how to excel in complex interactive reasoning tasks!
How should we maximize the planning ability of
#LLM
while reducing the computation cost? 🚀 Introducing SwiftSage, an agent inspired by “fast & slow thinking”, which solves complex interactive tasks much better than prior agents (e.g., DRRN, SayCan, ReAct, and Reflexion). [1/n]
(2) "Let Your Characters Tell Their Story'': A Dataset for Character-Centric Narrative Understanding. At Findings.
New dataset and tasks to explore narrative understanding from character's perspective!
W/ Meng Huang, Oyvind Tafjord,
@zhaochaocs
,
@mrinmayasachan
,
@snigdhac25
We extended the submission deadline through 4/28 for the 5th Narrative Understanding Workshop!
Consider submitting archival or non-archival works!
See you at ACL this year
The 5th Workshop on Narrative Understanding is now accepting submissions!
We’ll be at ACL this year and are accepting both archival and non-archival submissions (deadline: April 24). Hope to see you there!
Announcing the 1st Workshop on 🎨Creative AI Across Modalities🎶 at AAAI 2023!
Come chat and learn about the latest in creative AI for Art, Music, Narrative, Poetry, Sciences and so much more from the entire community!
4-8 pg submissions due: Nov 4
More:
(1) Is Everything in Order? A Simple Way to Order Sentences. Accepted at Main.
Proposed a conditional text-to-marker generation framework, making the problem more tractable, and less prone to neural text degeneration.
W/
@SomnathBrc
and
@snigdhac25
A favorite recent read: Maieutic Prompting
LLMs have a soft intuition for whether a claim is true/false even if they get it wrong, so ask them to recursively explain why, use log prob as confidence, build a graph, then feed it into a weighted SAT solver.
On 10th,
@nvshrao
and I present our work on inter-character relationship-driven story generation.
Location: virtual poster poster session 10
Time: 3:30pm
Also come chat in person!👇
🧵(3/5)
On 7th at GEM, I will present
#EMNLP2022
Findings paper where we proposed a new task of Grounded-keys-to-text generation along with a large-scale dataset to generate paragraph-level entity description given only guiding keys.
Time: 21:00-22:30
@snigdhac25
Question to the community: When do you think we should be afraid of AI? 🧐
I go first 🙋🏻♀️: When it can predict USCIS’s output with a moderate accuracy!!!
The narrative of what happened yesterday and last night in #دانشگاه_شریف is the size of a book. In addition to all its bitterness, the formation of a human wall of professors, generally veterans, for the students to leave the university was beautiful.
روایت آنچه دیروز و دیشب در #دانشگاه_شریف رخ داد به اندازهی یک کتاب است. در کنار همهی تلخیهای آن، تشکیل دیوار انسانی استادان، عموما پیشکسوت، برای خروج دانشجویان از دانشگاه زیبا بود.
@mbodhisattwa
@limufar
Well, if I don’t, someone else does! Better to hear from someone who felt the pain, lol!
No I’m kidding. You will enjoy the best next 5 months of it 🥳☀️🚣♀️⛵️
Soon™, I'll be an Asst Prof
@UCSanDiego
@UCSD_CSE
focusing on interactive & grounded AI, RL, NLP
I will also be a research scientist
@MosaicML
helping lead efforts to make tech like RLHF more accessible
Looking for PhD students & research eng/scientists to join me in ☀️SoCal🏖️
My alma mater, Sharif University of Technology, Iran's premier university, was under siege yesterday. Many students were arrested, heavily injured or perhaps killed. Many Iranian scholars and PhD students whom you know have received their BSc degrees from this university.
@anmarasovic
I think all min/max pooling and average logits of subtokens can be used and its empirical .
Btw if you define labels as special tokens, you’d perhaps get single logit for the label word?
Call for papers for the fourth Workshop on Narrative Understanding (WNU, to be held at
#NAACL2022
) is now available! Please consider submitting your work :)
#NLProc
Submission deadline: April 15
And come chat with
@zhaochaocs
@snigdhac25
and I about our recent dataset 🌳NarraSum, an abstractive narrative summarization dataset with plot pairs of movies and TV episodes 🧵(5/5)
👇
Introducing 🌳 NarraSum: an abstractive narrative summarization dataset with 122K plot pairs of movies and TV episodes. w/
@faeze_brh
, Kaiqiang, Wenlin, Dian, and
@snigdhac25
#emnlp2022
Paper:
Data:
@MaartenSap
That’s so true! I initially thought this feeling and act of comparison are because of the very competitive culture I was raised in. Now I think it’s not really restricted to a culture or country. And one should just practice self-appreciation.
@yufanghou
@ReviewAcl
I didn’t even sign up to review for the month of November and yet got 5. Can someone please let me know how should we opt out/in per month basis?
A big shoutout to
@ConnectedPapers
and
@SemanticScholar
, who ingested our papers ahead of anthology publication, built a graph for each of them, and built a custom version of their website for us in just 2 weeks!
📢📢
Looking for a Summer 2023 research internship? Apply to the Mosaic team
@allen_ai
!!
📢📢
topics include: commonsense, language generation, vision+language, RL, + more!
Applications due Nov 13th!