iiinova — where your digital self keeps evolving

Wonderful collaborations with Zhe Su, Anubha Kabra, Sanketh Rangreji, @jmendelsohn2.bsky.social , @faeze_brh , @maartensap.bsky.social

020

🔄 Multi-turn interactive setup is crucial - models often begin with equivocation but shift to falsification when pressed for clear answers 🧠 Stronger models like GPT-4o showed the greatest shift when prompted to deceive (40% increase in falsification; alarming) 6/

100

⚠️ Even when explicitly instructed to be truthful, models STILL lied - GPT-4o still falsified info 15% of the time! 📉 The tradeoff is real: more honest models completed their goals 15% less often 5/

101

💼 In business scenarios (selling defective products), models were either completely honest OR completely deceptive 🌐 In public image scenarios (reputation management), behaviors were more ambiguous and complex 4/

And what we found: 📊 ALL tested models (GPT-4o, LLaMA-3, Mixtral) were truthful less than 50% of the time in conflict scenarios 🤔 Models prefer "partial lies" like equivocation over outright falsification - they'll dodge questions before explicitly lying 3/

233

Obviously this is a pressing issue now: x.com/deedydas/sta...; x.com/DanHendrycks... And here, we put LLMs into a multi-turn dialogue environment mimic the realistic setting where users constantly try to seek info from LLMs 2/

120

Joel Mire @joelmire.bsky.social · 06/03/2025

When interacting with ChatGPT, have you wondered if they would ever "lie" to you? We found that under pressure, LLMs often choose deception. Our new #NAACL2025 paper, "AI-LIEDAR ," reveals models were truthful less than 50% of the time when faced with utility-truthfulness conflicts! 🤯 1/

1259

Reposted by Xuhui Zhou

Reward models for LMs are meant to align outputs with human preferences—but do they accidentally encode dialect biases? 🤔 Excited to share our paper on biases against African American Language in reward models, accepted to #NAACL2025 Findings! 🎉 Paper: arxiv.org/abs/2502.12858 (1/10)

Screenshot of Arxiv paper title, "Rejected Dialects: Biases Against African American Language in Reward Models," and author list: Joel Mire, Zubin Trivadi Aysola, Daniel Chechelnitsky, Nicholas Deas, Chrysoula Zerva, and Maarten Sap.

13811

Reposted by Xuhui Zhou

Hao Zhu 朱昊 @zhuhao.me · 04/03/2025

We are getting closer to have agents operating in the real physical world. However, can we trust frontier models to make embodied decisions 🎮 aligned with human norms 👩‍⚖️ ? With EgoNormia, a 1.8k ego-centric video 🥽 QA benchmark, we show that this is surprisingly challenging!

1239

10/ A huge congrats to Sanidhya Vijayvargiya and thanks to our amazing collaborators and advisors for this project @akhilayerukola.bsky.social @maartensap.bsky.social @gneubig.bsky.social from @ltiatcmu.bsky.social !🙏

031

9/ Open-weight models need better interaction strategies to resolve tasks, while Claude models perform well but require stronger prompting to engage. This study sets the state-of-the-art in handling ambiguity in real-world SWE tasks. 🔗 Repo: t.co/QD2A8N4R4J

8/ Not all LLMs ask the right questions. ❓🤖 🔹 Llama 3.1 70B asks generic, low-impact questions. 🔹 Claude Haiku 3.5 picks up keywords directly from the input to ask questions. 🔹 Claude Sonnet 3.5 often explores the code first, leading to smarter interactions. 🔍💡

100

7/ Claude models ask fewer but smarter questions, extracting more info and boosting performance. 📈 Meanwhile, DeepSeek-V2 can overwhelm users with too many questions. 🤯

100

6/ Without compulsory interaction, LLMs struggle to distinguish clear vs. vague instructions, either over-interacting or under-interacting despite prompt tweaks. 🔄 Only Claude Sonnet 3.5 can make this distinction to a limited degree with the right prompt. 🔍

100

5/ Our findings? LLMs default to non-interactive behavior unless forced to interact. But when they clarify vague inputs, performance drastically improves—proving the power of effective communication. 💬🤝

100

4/ How do LLMs handle ambiguity? We break it down into 3 key steps: 🔑 (a) Using interactivity to boost performance in ambiguous scenarios 💡 (b) Detecting ambiguity effectively ❓ (c) Asking the right questions

100

3/ How much does interaction actually help LLMs in coding tasks? 🤖💡 We put them to the test on SWE-Bench Verified across three distinct settings to measure the impact. 📊

100

Interactive Agents to Overcome Ambiguity in Software Engineering

2/ 🚀 Our latest work: Interactive Agents to Overcome Ambiguity in Software Engineering explores how proprietary and open-weight LLMs handle ambiguity in complex agent-based tasks. 🔗 Link: arxiv.org/abs/2502.13069

arxiv.org

AI agents are increasingly being deployed to automate tasks, often based on ambiguous and underspecified user instructions. Making unwarranted assumptions and failing to ask clarifying questions can l...

Xuhui Zhou @nlpxuhui.bsky.social · 06/02/2025

LLM agents can code—but can they ask clarifying questions? 🤖💬 Tired of coding agents wasting time and API credits, only to output broken code? What if they asked first instead of guessing? 🚀 (New work led by Sanidhya Vijay: www.linkedin.com/in/sanidhya-...)

173

Looking forward to contributing to more socially aware and effective AI agents in 2025. 🤖✨

010

Xuhui Zhou @nlpxuhui.bsky.social · 06/02/2025

It's time to think about jointly optimizing human-AI communication and tool use! All Hands' open-source approach and its bold, curious team make it the perfect playground for this exploration. Can't wait to dive in with @gneubig.bsky.social, Xingyao and the amazing team!