iiinova — where your digital self keeps evolving

Peng Qi @qi2peng2.bsky.social · 29/09/2026

-- but a simple modified preference learning technique would solve it perfectly. I'll be presenting this technique, Abductive Preference Learning, next Thursday at #COLM2026, come say hi! Joint work with Yijin Ni and Simon Yu /fin

010

Peng Qi @qi2peng2.bsky.social · 29/09/2026

A picture that looks like a dog? Four legs. Location for the last Olympics? Paris. A question that looks like it needs detailed answers? Start a bullet point list. Simple finetuning won't cure this problem, especially when the baked-in behavior is so overwhelmingly represented in training -- 2/

100

Peng Qi @qi2peng2.bsky.social · 29/09/2026

We found that LMs, large or small, text-only or multimodal, are susceptible to a shared problem -- 𝗼𝘃𝗲𝗿𝗰𝗼𝗻𝗳𝗶𝗱𝗲𝗻𝗰𝗲 𝗮𝗻𝗱 𝗼𝗯𝗹𝗶𝘃𝗶𝗼𝗻 𝘁𝗼 𝗱𝗲𝘁𝗮𝗶𝗹𝘀. When presented with a scenario that is uncommon with overwhelming pretraining data that encourages a shortcut, they take the shortcut confidently all the time 1/

100

Peng Qi @qi2peng2.bsky.social · 26/07/2025

How do we prove that #AI can't do #maths? Real Mathematics (yes, "real" is a pun here): a+b+c = (a+b)+c = a+(b+c) AI Mathematics (well, floating point maths, really): >>> 0.1+0.2+0.3 0.6000000000000001 >>> 0.1+(0.2+0.3) 0.6 QED.

000

Peng Qi @qi2peng2.bsky.social · 16/07/2025

This project was joint work with my Amazon colleagues (led by Yumo Xu), and it's great to see it finally published. Hope this helps motivate more careful eval work in the near future! #AI #agent #evaluation #RAG #NLP

011

Peng Qi @qi2peng2.bsky.social · 16/07/2025

b) as builders, we evaluate the technology soberly and help users navigate these risks in product design. Want to learn more? Checkout Our paper: arxiv.org/pdf/2506.01829 Open-source code: github.com/amazon-scien...

100

Peng Qi @qi2peng2.bsky.social · 16/07/2025

Why should you care? As businesses / individuals leverage AI more and more to speed up research and decision-making, it is important that, a) as users, we examine the tools we are using to understand their limitations and avoid pitfalls with significant potential downsides, and

100

Peng Qi @qi2peng2.bsky.social · 16/07/2025

With a new, carefully annotated dataset and an automated evaluation metric we designed, we find that although LLMs are reasonably good at citing accurate sources most of the time, SOTA LLMs still cite incorrectly 5-28% of the time, and miss citations anywhere from 16% to an alarming 95% of the time.

110

Peng Qi @qi2peng2.bsky.social · 16/07/2025

"𝐂𝐢𝐭𝐞𝐄𝐯𝐚𝐥: 𝐏𝐫𝐢𝐧𝐜𝐢𝐩𝐥𝐞-𝐃𝐫𝐢𝐯𝐞𝐧 𝐂𝐢𝐭𝐚𝐭𝐢𝐨𝐧 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧 𝐟𝐨𝐫 𝐒𝐨𝐮𝐫𝐜𝐞 𝐀𝐭𝐭𝐫𝐢𝐛𝐮𝐭𝐢𝐨𝐧", we propose a framework to systematically study citation accuracy by considering previously neglected contexts such as user-provided information and LLMs' parametric knowledge.

100

Peng Qi @qi2peng2.bsky.social · 16/07/2025

As 🔎 AI deep research agents 🔎 become an essential part of many people's day-to-day work, it is more essential than ever before that we can trust what they produce. When these agents cite sources they claim the report is based on, how much can we actually trust them? In our new #ACL2025 paper, ...

100

Peng Qi @qi2peng2.bsky.social · 02/07/2025

a good problem in its historical context, some of my own research attempts at solving this problem that I believe are on the critical path to autonomous agents, and what's changed today to make the dataset less relevant in its original form. I also reflect on the possible paths forward for ...

qipeng.me

Why You Should Stop Using HotpotQA for AI Agents Evaluation in 2025 | Peng Qi

We published HotpotQA, a groundbreaking multi-step question answering dataset in 2018, which has since motivated and facilitated numerous AI agent research works. But you should probably reconsider…

110

Peng Qi @qi2peng2.bsky.social · 02/07/2025

Seven years ago, I co-led a paper called 𝗛𝗼𝘁𝗽𝗼𝘁𝗤𝗔 that has motivated and facilitated many #AI #Agents research works since. Today, I'm asking that you stop using HotpotQA blindly for agents research in 2025 and beyond. In my new blog post, I revisit the brief history of 𝗛𝗼𝘁𝗽𝗼𝘁𝗤𝗔, why it defined ...

qipeng.me

Why You Should Stop Using HotpotQA for AI Agents Evaluation in 2025 | Peng Qi

100

Peng Qi @qi2peng2.bsky.social · 22/05/2025

The longer employers don’t acknowledge and embrace this discrepancy, the faster they lose the top candidates they spent enormous efforts to hire and retain, leaving the organization in self-fulfilling mediocrity.

010

Peng Qi @qi2peng2.bsky.social · 22/05/2025

but not perfectly aligned with those of the employer. They ask: Will I get the opportunity to build a career beyond what is immediately required of me? Will I learn and grow, be part of a great team and culture? Will I make a name for myself while doing great work? Will I remain competitive? 3/

100

Peng Qi @qi2peng2.bsky.social · 22/05/2025

We are never settling for a candidate that does exactly the thing that needs to be done right now, since that thing itself can change before you know it. But too often employers and managers forget, that highly motivated and capable candidates also hold expectations parallel to these, ... 2/

100

Peng Qi @qi2peng2.bsky.social · 22/05/2025

When making great hashtag#hiring decisions, we often look for growth potential in a candidate. Will they rise to the occasion when unforeseen challenges arise? Will they grow in the role, and lift up others in the team? Will they still be able to contribute if business direction changes? 1/

100

Peng Qi @qi2peng2.bsky.social · 06/05/2025

While many aspects of our work (especially in the digital world) can be amenable to #AI #automation, it is also through automation that we continuously rediscover again and again the true meaning of our work and our unique humanness. #MondayReflection /fin

010

Peng Qi @qi2peng2.bsky.social · 06/05/2025

this coding phase alone, and I ended up delivering something slightly better than I would've done without it. As with any technological evolution, tools themselves never fully replace the humans doing the work, but greatly enhance the ones that embrace them and adapt to working with them. 6/

100

Peng Qi @qi2peng2.bsky.social · 06/05/2025

But, of course, this has its implications. I did save a lot of time looking up programming resources on things I have a vague understanding of and wasn't very familiar with, and didn't have to type all those many characters. By my estimate, the AI assistant did save me 50-80% of the effort of 5/

100

Peng Qi @qi2peng2.bsky.social · 06/05/2025

(especially to the general public) the fact that the act of putting code down is typically the *least* mentally effortful part of the work. It's as if saying "my 3D printer made 100% of my new shiny collections" -- true in the narrow sense of the printing effort, but it's missing the point. 4/

100

Peng Qi @qi2peng2.bsky.social · 06/05/2025

how to fix things that aren't working (the AI helped a bit with this too at times), and how to keep things future-proof. In this regard, I still did >90% of the most important *work* in this project. Saying AI 99% of the code, while factually correct in one particular sense (line count), obscures 3/

100

Peng Qi @qi2peng2.bsky.social · 06/05/2025

AI with no manual edits from me, or that the AI assistant was the last to "touch" those lines of code. What this 99% number oversimplifies is the amount of time my colleague and I engage in numerous offline discussions, times where I had to stop and think about what to ask the AI to code next, 2/

100

Peng Qi @qi2peng2.bsky.social · 06/05/2025

In one of my recent projects, AI code assistants actually DID write 99% of my code, and the project was reasonably complex starting from scratch. Does this mean I'm obsolete now? Here's the catch:when I say AI wrote 99% of the code, I was counting roughly how many lines were directly generated by 1/

100

Peng Qi @qi2peng2.bsky.social · 06/05/2025

#AI 𝘄𝗿𝗼𝘁𝗲 𝟵𝟵% 𝗼𝗳 𝗺𝘆 𝗰𝗼𝗱𝗲, 𝗻𝗼𝘄 𝘄𝗵𝗮𝘁? Big tech executives and business analysts are racing to share eye-catching statements like "AI will write XX% of the code at MetaCorp by 20YY." How much truth is there to these, and what implications might this have? 🧵

100

Peng Qi @qi2peng2.bsky.social · 28/03/2025

Non-native speakers sometimes have a unique advantage to language-based humor stemming from their unfamiliarity with idiomatic expressions. I saw an “assembly of god” on the road and thought to myself, “wait, they have a factory to build gods here?”

000

Reposted by Peng Qi

Peng Qi @qi2peng2.bsky.social · 13/03/2025

Is #AI the new #RocketScience? In my new blog post, I explore the similarities and connections between the two seemingly distant relatives, and reflect on what today's AI scientists can learn from their rocket cousins, plus what makes AI science unique:

qipeng.me

AI is the New Rocket Science | Peng Qi

AI science of today has astonishing similarities to rocket science in its prime days, if one pays close attention to history. What are some of these, and what can the history of rocket science tell…

001

Peng Qi @qi2peng2.bsky.social · 13/03/2025

qipeng.me

AI is the New Rocket Science | Peng Qi

001

Peng Qi @qi2peng2.bsky.social · 12/03/2025

my manual work of at least 10+ minutes into a 30s breeze. How can we enable 99% of the the population to build this tooling for themselves? #Reflection #Automation

000

Peng Qi @qi2peng2.bsky.social · 12/03/2025

#Reflection The need for tooling appears everywhere. While oftentimes unmet, it can significantly scale and improve the productivity of many people when fulfilled, especially in the digital world. Kudos to whoever wrote the AC checklist notebook for ACL 2025 Senior Area Chairs, which reduced ...

100

Peng Qi @qi2peng2.bsky.social · 28/02/2025

I was worried at first that the story required too much context form the MT original to resonate. It turned out a self-contained story with James' unique journey as he navigates racism and slavery, the voice acting brings the characters/personas to life esp you are not that familiar with AAVE.

010

Peng Qi @qi2peng2.bsky.social · 28/02/2025

When people ask me about good intro audiobooks, I've always recommended The Martian, which is one of my personal sci-fi favorites accompanied with a voice performance that truly brought the story to life. Now I'll be adding James (by Percival Everett, performed by Dominic Hoffman) to my personal rec

220

Peng Qi @qi2peng2.bsky.social · 21/02/2025

Top 6 jobs LinkedIn recommended for me: * VP of AI at a unicorn public company * Applied AI Engineer II at a F50 company * Applied Scientist II at a F10 co. * Research Intern at a F50 co. * Senior Principal Applied Scientist at a F500 co. * Senior Director of Applied Science at a F500 co. Who am I?!

000

Peng Qi @qi2peng2.bsky.social · 07/02/2025

Thanks to Google, my name is now Robin (@robinjia.bsky.social) and I work at USC as a CS professor. (Or that Robin took my name and worked at Amazon 🤣 )

030

Peng Qi @qi2peng2.bsky.social · 25/01/2025

Definitely! We have a job posting on our website where you can apply! Check out jobs.ashbyhq.com/orby-ai

jobs.ashbyhq.com

Orby AI Jobs

000

Peng Qi @qi2peng2.bsky.social · 23/01/2025

solving them. As an RE, you'll work on state-of-the-art models & agents and infrastructure to support their training and serving. If you have friends who might be interested, don't hesitate to forward them this information! Application links: RS: buff.ly/4avCOm8 RE: buff.ly/4aC8qqo

buff.ly

Research Scientist

Research Scientist (Senior)

010

Peng Qi @qi2peng2.bsky.social · 23/01/2025

We are looking for strong candidates for Research Scientist and Research Engineering roles at Orby AI. Join us to define the next generation of foundation models and AI agents for GUI automation! As an RS, you'll work towards defining groundbreaking research problems and leading the team in ...

buff.ly

Research Scientist

Research Scientist (Senior)

250

Reposted by Peng Qi

Dan Jurafsky @jurafsky.bsky.social · 12/01/2025

Happy New Year everyone! Jim and I just put up our January 2025 release of Speech and Language Processing! Check it out here: web.stanford.edu/~jurafsky/sl...

web.stanford.edu

Speech and Language Processing

115250

Peng Qi @qi2peng2.bsky.social · 10/01/2025

Chinese #languagelearners of Korean should srsly consider detouring to Japanese first: some basic pronunciation and Kanji vocabulary can go a long way in helping you associate Korean words with its underlying Hanja. I'd never have remembered 일 월 화 수 목 금 토 without learning about 日月火水木金土 first.

010

Peng Qi @qi2peng2.bsky.social · 09/01/2025

#NLProc is hard. I read this trending topic on Twitter, and it took me a full minute before I think I got the correct parse. Is this what #LLMs feel when they read technical stuff they outside of their training data?

000

Peng Qi @qi2peng2.bsky.social · 08/01/2025

with reasonable freedom, depending on the scale/focus of the business. Case in point, we are looking to expand the research/foundation models team at Orby AI and are looking for highly motivated researchers and ML/Research engineers. Please reach out if you're interested in learning more! /fin

022

Peng Qi @qi2peng2.bsky.social · 08/01/2025

the usual suspects, as well as many non-large-tech companies, including large companies not traditionally known for Internet/software products, AI-adjacent scale-up companies, and AI-native startup companies. Many of these can provide opportunities for exciting exploratory work 5/

100

Peng Qi @qi2peng2.bsky.social · 08/01/2025

were writing the research papers you are the most proud of. While many academic institutions are growing fast in recent years, there is likely still a fierce competition in more popular directions. In the meantime, there is actually a growing demand in the industry of AI talent, including 4/

100

Peng Qi @qi2peng2.bsky.social · 08/01/2025

2. Having more roles and players in the AI market does not mean fewer opportunities for PhDs, because the market is still in an upward trend. It does mean, however, that opportunities might be lying in places you hadn't thought about or anticipated when you entered the PhD or 3/

100

Peng Qi @qi2peng2.bsky.social · 08/01/2025

critical thinking, independence in problem solving, domain expertise, and vision about strategic technical directions are all valuable traits that takes a serious amount of training to obtain, and something one should be proud of and play to their own strength. 2/

100

Peng Qi @qi2peng2.bsky.social · 08/01/2025

1. Indeed, a lot of AI jobs in the industry will no longer require a PhD, because with scale and production comes processes and reliability requirements that do not require paradigm-shifting changes all the time. However, this does not mean a PhD training will make no difference -- 1/

100

Peng Qi @qi2peng2.bsky.social · 08/01/2025

@kyunghyuncho.bsky.social's blog post on anxiety and frustration at #NeurIPS (buff.ly/4fN4Hah) has provided deep insight into the recent history of the AI boom and the angst many PhD students are facing. I'd like to echo 2 points I find myself agreeing a lot with Kyunghyun and expand a bit.

kyunghyuncho.me

i sensed anxiety and frustration at NeurIPS’24 – Kyunghyun Cho

Kyunghyun Cho

180

Peng Qi @qi2peng2.bsky.social · 03/01/2025

publishing (or more specifically, the lack thereof) in the industry, what are the implications for individuals, and what the day-to-day might look like for industry researchers. Hope this provides useful perspective: buff.ly/400ITlG 3/

buff.ly

What do industry researchers do, anyway? Part 2 -- What do they do when they are not publishing

Many people seem to

100

Peng Qi @qi2peng2.bsky.social · 03/01/2025

... in this series, which actually aims to answer the original question in the series title: "What does an industry researcher do, anyway?" In this post, I attempt to answer some questions my younger self had about 2/

buff.ly

What do industry researchers do, anyway? Part 2 -- What do they do when they are not publishing

Many people seem to

100