zlinsmile

[Kenton Murray @kentonmurray]

I'm excited to announce that this Fall I will be joining the Computer Science Department at George Mason University as an Asst. Prof. I'll be expanding my lab and looking for PhD students to work on Multilingual AI problems text, video, and speech.

https://t.co/f6PIESWRp4

[Tianjian Li @tli104]

🚨 New paper: "Self-Compacting Language Model Agents"

LM agents build up long traces of reasoning and tool calls. As the trace grows, old mistakes and stale info stick around and anchor everything that follows. We ask: can the model itself decide when to clean up?

[Rafael Rivera Soto @rafaelrivera01]

New #ICML2026 paper!

A popular claim says machine-text detection is hopeless since detectors get defeated by cheap attacks (paraphrasing, prompting, DPO). We show the picture is more subtle. 👇

📄 https://t.co/yLCOaMwxNr
💻 https://t.co/F33JaFQVFK

[Jonathan Ivey @Jonathan_Ivey8]

What makes a good participant response in qualitative research? In our new paper, we investigate this by analyzing 16,940 responses across 343 interviews to see which characteristics of a response make it most likely to contribute to a study’s findings.

Here’s what we found …

[Brian Zheyuan Zhang @zheyuanzhang99]

Introducing AgentOdyssey — an open-ended, long-horizon text game generation engine for 𝐭𝐞𝐬𝐭-𝐭𝐢𝐦𝐞 𝐜𝐨𝐧𝐭𝐢𝐧𝐮𝐚𝐥 𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐚𝐠𝐞𝐧𝐭𝐬.

Real-world agents cannot have a boundary between training and testing: they must learn continuously from interaction with the world at test time. AgentOdyssey is designed to study five key abilities that make this possible: 🧭 𝐞𝐱𝐩𝐥𝐨𝐫𝐚𝐭𝐢𝐨𝐧, 🧠 𝐞𝐩𝐢𝐬𝐨𝐝𝐢𝐜 𝐦𝐞𝐦𝐨𝐫𝐲, 🌍 𝐰𝐨𝐫𝐥𝐝 𝐤𝐧𝐨𝐰𝐥𝐞𝐝𝐠𝐞 𝐚𝐜𝐪𝐮𝐢𝐬𝐢𝐭𝐢𝐨𝐧, 🛠️ 𝐬𝐤𝐢𝐥𝐥 𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠, and 🎯 𝐥𝐨𝐧𝐠-𝐡𝐨𝐫𝐢𝐳𝐨𝐧 𝐩𝐥𝐚𝐧𝐧𝐢𝐧𝐠.

Just `pip install agentodyssey` and start generating games and evaluating your agents and LLMs!

🌐 Project website: https://t.co/HNA0HUgCqb

[Jack Jingyu Zhang @jackjingyuzhang]

Real-world agents juggle instructions from skill files, tools, other agents, ... each with different trust levels. When these conflict, can models reliably prioritize the most trusted one?

Our ManyIH-Bench🪜 finds that even frontier models like GPT-5.4 only get ~40% accuracy! 👇

[Zhaopeng Tu @tuzhaopeng]

Can 2025 LLMs balance high IQ (Strategy) with high EQ (Personality)? 🤖🧠❤️

Thanks to the amazing @JentseHuang, we have updated our three major benchmarks -- PsychoBench, EmotionBench, and GAMA-Bench -- with the latest 2025 models (Gemini-3, GPT-5, Claude-4.5, DeepSeek-v3.2). The results reveal a shifting landscape in AI cognition and sentiment.

1⃣ Strategic Reasoning has a new king.
On GAMA-Bench (ICLR'25), Claude-Sonnet-4 dominates with a score of 83.7, far outpacing GPT-4o (65.5). This suggests that models like Claude are cracking complex multi-agent cooperation and betrayal dynamics better than their peers.

2⃣ High Empathy is becoming the norm.
On PsychoBench (ICLR'24), Gemini-2.5-Pro reaches an Empathy score of 6.96, significantly surpassing the human crowd baseline of 4.92. Meanwhile, safety tuning is working: Gemini-3-Pro shows near-zero Dark Triad traits (Narcissism/Machiavellianism), whereas older open models like LLaMA-3.1-70B retain "darker" profiles.

3⃣ Emotional Resilience is stabilizing.
On EmotionBench (NeurIPS'24), GPT-5 displays human-like stability. Its positive mood drops significantly less in negative contexts (-8.2) compared to GPT-4 (-27.6), while DeepSeek-V3.2 demonstrates sharper, more appropriate negative empathy responses.

4⃣ Different families, different strengths.
While Claude leads in pure game-theoretic reasoning (GAMA), the Gemini and GPT-5 families excel in psychological safety and emotional stability. The gap between "Thinking" and "Feeling" capabilities is defining the 2025 model landscape.

GAMA-Bench (ICLR'25): https://t.co/U1cT5JMA36 🧠
PsychoBench (ICLR'24): https://t.co/pzgLa56GRb ❤️
EmotionBench (NeurIPS'24): https://t.co/SSI0tISC6k 🎭

We will keep maintaining our benchmarks! We have updated our GitHub code so SOTA models can easily be tested.

[zlinsmile @zlin_speech]

@erica_cooper Thanks for sharing and giving great talk! ❤️

[Paola Garcia @leibnyPaola]

📢📢@jhuclsp Reminder!!

Have an Idea 💡? Let’s Hear It! JSALT 2026 Call for Proposals is out.
Deadline: November 14th, 2025. For more information:

https://t.co/jwdRCF4qms

[jiatongshi @jiatongshi]

🚀 I’m open to new opportunities in industry!

Ph.D. candidate @CMU (advisor: @shinjiw_at_cmu ).
Research: speech/audio AI, speech LLMs, evaluation frameworks.
Ex-Meta AI, Tencent, IBM Research.

DMs open — let’s connect!

[zlinsmile @zlin_speech]

@berraksismann wow! soooo cute! ❤️ congratulations to your new 🐶

[zlinsmile @zlin_speech]

@erica_cooper @juliahberg wow! Congratulations!

[zlinsmile @zlin_speech]

@moinnadeem @huggingface same☹️

[DeepSeek @deepseek_ai]

🚀 DeepSeek-R1 is here!

⚡ Performance on par with OpenAI-o1
📖 Fully open-source model & technical report
🏆 MIT licensed: Distill & commercialize freely!

🌐 Website & API are live now! Try DeepThink at https://t.co/v1TFy7LHNy today!

🐋 1/n

[zlinsmile @zlin_speech]

@ieeeICASSP @TheKarimKhoury Thanks a lot for your quick reply!

[zlinsmile @zlin_speech]

@ieeeICASSP @TheKarimKhoury Hi @ieeeICASSP , may I further confirm the format for the author list please?
I found two options in template project: (1) 'conference_101719.tex' shows 6 authors in 3 columns (Fig. 1). (2) merge affliation as in Section IV.B of 'IEEEtran_HOWTO.pdf' (Fig. 2)
Are both correct? Thx

[INTERSPEECH 2025 @ISCAInterspeech]

Enhance your skills in combating speech deepfakes at the ASVspoof workshop during #INTERSPEECH2024!

📅 August 31, 2024, 🕗 Full-day Workshop. Register now! https://t.co/xwaskpKruU

#AI #SpeechTech #Deepfakes #Security

[Security & Privacy in Speech Communication @SPSC_SIG]

📢 SPSC Webinar:
"Whether When What": Detection, Localization and Diarization of Partially Spoofed Audio

🗓️ Wed. July 17th, 10am CEST, Lin Zhang from @ButSpeech

✒️ Register here: https://t.co/10DHnCZBpl (yes, correct link!)

It's free and open to everyone!

[zlinsmile @zlin_speech]

Thanks for everything🩷