
LOR deadline day ... so many required fields! ... luckily a few of them have a high character limit

Professor of CS at Johns Hopkins University, ACL Fellow. My tweets speak only for me.

LOR deadline day ... so many required fields! ... luckily a few of them have a high character limit

Congratulations to all @JohnsHopkins researchers participating in #ICLR2025! Check out all @JohnsHopkins accepted papers, tutorials, and workshops at ai.jhu.edu/news/johns-hopkin….

Who wants to come to JHU and do a postdoc with me?? I'm always enthusiastic about new modeling / inference / algorithmic ideas in NLP/ML. Also selected applications.
We’re thrilled to announce the #HopkinsDSAI Postdoctoral Fellowship Program! We’re looking for candidates across all areas of data science and AI, including science, health, medicine, the humanities, engineering, policy, and ethics. Apply today! ai.jhu.edu/postdoctoral-fell…

New #ACL2024 paper: LLMs in the Imaginarium: Tool Learning through Simulated Trial and Error (@BoshiWang2's internship work at Microsoft Semantic Machines) I like this work because it takes home an important insight: synthetic data + post-training is critical for agents. Agents need perception-decision-execution capability and data, which is hard to get from pre-training because data on the Internet is mostly artifacts produced by such processes, not capturing the processes per se. I believe LLMs + synthetic data + environmental feedback will prove to be an immensively successful recipe for agents, and our work is just an early example of that. Nice work @BoshiWang2 @hfang90 @adveisner @ben_vandurme
Thanks @_akhaliq for sharing our work led by @BoshiWang2 from @osunlp, so let's chat about how LLMs should learn to use tools, a necessary capability of language agents. Tools are essential for LLMs to transcend the confines of their static parametric knowledge and text-in-text-out interface, empowering them to acquire up-to-date information, call upon external reasoners, and take consequential actions in external environments. But existing tool learning strategies for LLMs are not fully unleashing this potential––we find existing LLMs, including GPT-4 and ones specifically tuned for tools like ToolLLaMA, only achieve 30% to 60% correctness in tool use. Very far from the high level of accuracy needed for practical deployment in high-stakes scenarios. A closer look at existing methods reveals that people seem to be overly focused on rushing to add as many tools as possible or make it easy to add new tools, either through in-context learning or fine-tuning. Perhaps surprisingly, a critical aspect, how to make LLMs use existing tools as accurately as possible, is largely overlooked. How to Truly Master a Tool? We turn to successful precedents in the biological system such as humans, apes, and corvids. Learning to use a tool is a rather advanced cognitive function that depends on many other cognitive functions: > Trial and error is essential for tool learning. We do not master a tool solely by reading the ‘user manual’; rather, we explore different ways of using the tool, observe the outcome, and learn from both successes and failures. > Intelligent animals do not just do random trial and error—we proactively imagine or simulate plausible scenarios that are not currently available to perception for exploration > Finally, memory, both short-term and long-term, is instrumental for the progressive learning and recurrent use of tools So why should we expect an LLM to magically master a tool by just looking at the ‘user manual’ (API specification) or several examples? That's even hard for humans who are equipped with many cognitive substrates missing in LLMs! Simulated Trial and Error (STE) > STE is a biologically inspired method for tool-augmented LLMs that combines trial and error, imagination, and memory. > Given a tool, STE leverages an LLM to simulate, or ‘imagine’, plausible scenarios for using the tool. It then iteratively interacts with the API to fulfill the scenario by synthesizing, executing, and observing the feedback from API calls, and then reflects on the current trial. We devise memory mechanisms to improve the quality of the simulated instructions. A short-term memory facilitates deeper exploration in a single episode, while a long-term memory maintains progressive learning over a long horizon > STE allows an LLM to sufficiently explore and probe the capability boundry of a tool. With the tool use examples from STE, one can either use them for in-context learning or fine-tuning Result Highlights > GPT-4: 60.8 -> 76.3 📈 > Mistral-7B: 30.1 -> 76.8 📈📈 and outperforms GPT-4! > Llama-2-7B: 10.7 -> 73.3 📈📈📈 > Also substantially outperform existing tool-augmented LLMs like ToolLLaMa-v2 (37.3) But will the tool fine-tuning destroys the LLM's other capabilities? No worries, we got you covered. A simple experience replay strategy is enough to maintain an LLM's existing capabilities while learning new tools! 📌 Paper: arxiv.org/abs/2403.04746 📌 Code: github.com/microsoft/simulat… (fairly easy to use. try it out on your own tools!)

Great last-minute summer opportunity on LLMs for social good / democracy! (If you're an #NLProc PhD student.) Please retweet. (Team includes @adveisner @DanielKhashabi @ZiangXiao @AndrewJPerrin + students. 8 weeks alongside 3 other teams: fun, meaningful, educational, social.)
WHAT: "AI-Curated Democratic Discourse," a JSALT hackathon team this summer (Jun 10-Aug 2) GOAL: Redesign the social media UI to raise the quality of reading and posting, with the help of LLMs🤯 WHO: Looking for 1 more funded, in-person NLP PhD student! DM me with yr skillz.

Project description: cs.jhu.edu/~jason/tmp/eisner… Event format: clsp.jhu.edu/2024-jelinek-su… I'll take questions here. If you know someone great for this, please have them email me, thx! (Someone tuned into political discourse and UI design who is great at coaxing behavior out of LLMs?)

WHAT: "AI-Curated Democratic Discourse," a JSALT hackathon team this summer (Jun 10-Aug 2) GOAL: Redesign the social media UI to raise the quality of reading and posting, with the help of LLMs🤯 WHO: Looking for 1 more funded, in-person NLP PhD student! DM me with yr skillz.

Today I had to refer to an unnbalanced / nonn-balanced binary tree ...

(OED's earliest citation, de Moivre's 1718 translation of his own 1711 book, was surely influenced by Huygens? De Moivre used "Expectation"/"Risk" for a pos/neg value. Laplace 1814 gave a modern def of expected value, but called it "espérance mathématique": mathematical hope.)

Huygens 1657 was the first published work on probability. As far as I can tell, this 1714 translation first introduced this sense of "expectation" into English. (The quote is asking: What x should I be willing to pay for the right to get either a or b with equal odds?)

Can we go back to 1657 to tell Huygens that instead of "expectation," he should call it "argmean"? "Don't ask, Chris. It'll make sense in a few hundred years."

I was on the ACL committee that worked on this issue over the past 4-5 months. There were many, many viewpoints and proposals to sort through / debate. No perfect solutions. But amazingly, we got to a consensus on the best🤞way forward. Check out our report, via the link! (2/2)

ACL is shifting its strategy for protecting anon review. Instead of asking authors to keep quiet about their work during a long "anonymity period," we'll try not to rely on reviewers who actually heard about the work. We hope this plays out better. (1/2)
ACL announcement: "The ACL Executive Committee has voted to significantly change ACL's approach to protecting anonymous peer review. The change is effective immediately. (1/4) #NLPRoc

Thrilled to (finally) be able to share our paper: 🐑 Do language models know when they're hallucinating? TL;DR answer: yes! 🔗: arxiv.org/abs/2312.17249, and a🧵 w/ Semantic Machines @MSFTResearch: @ben_vandurme & @adveisner & Chris Kedzie

If you can break your AI problem down, an LLM is the 3D printer that can quickly create modules for your subtasks. No module should bear too much load, as they're all made of cheap plastic. But a 3D printer has countless uses. (Our previous 3D printer was MTurk.)

Dear ACL community, ACL is considering multiple proposals to change its anonymity period policy. It seeks immediate feedback from the community about the proposed changes. Please add your voice until Friday, September 22nd (AOE): aclweb.org/portal/content/su… #NLProc