Join us for "Personal Agent Era: Eval & Training of the Best Personal Agent."
Date: October 30, 2026 (Friday), 6:00 PM to 9:30 PM
Location: Mountain View @MatrAIx2026 office
We'll share the latest findings from our PersonalAgentEval (lnkd.in/gsSmhAr7) research across popu. We'll discuss evaluation frameworks that drive product iteration, constructing RL training data, and key challenges in agent evaluation.
Register: luma.com/vn4ectbt
#MatrAIx #PersonalAgent #AIEvaluation #ReinforcementLearning #AgentEval #TechEvent #MountainView @XiaominLi98 @YuexingHao @UofBeta
Date: October 30, 2026 (Friday), 6:00 PM to 9:30 PM
Location: Mountain View @MatrAIx2026 office
We'll share the latest findings from our PersonalAgentEval (lnkd.in/gsSmhAr7) research across popu. We'll discuss evaluation frameworks that drive product iteration, constructing RL training data, and key challenges in agent evaluation.
Register: luma.com/vn4ectbt
#MatrAIx #PersonalAgent #AIEvaluation #ReinforcementLearning #AgentEval #TechEvent #MountainView @XiaominLi98 @YuexingHao @UofBeta
0
1
4
110
MatrAIx retweeted
Thanks @Columbia University for inviting us to give a talk on PersonalAgentEval (personalagenteval.com) this morning!
Excited to see interest in our work at @MatrAIx2026 from both industry and academia. Great to share what we’re building and exchange ideas with the research
#MatrAIx #PersonalAgents #Agents #ColumbiaUniversity #MetaMuse #OpenAIDots #ClaudeDuet #GrokBot #Hark #Instinct #GeminiSpark
Excited to see interest in our work at @MatrAIx2026 from both industry and academia. Great to share what we’re building and exchange ideas with the research
#MatrAIx #PersonalAgents #Agents #ColumbiaUniversity #MetaMuse #OpenAIDots #ClaudeDuet #GrokBot #Hark #Instinct #GeminiSpark
0
1
6
249
MatrAIx retweeted
Excited to be invited by @Alibaba_Qwen team to give a talk about PersonalAgentEval (personalagenteval.com)! Had great discussions about eval and training for Personal Agents.
@MatrAIx2026 @YuexingHao
#MatrAIx #PersonalAgents #Agents #Qwen #Alibaba #MetaMuse #OpenAIDots ark #Instinct
@MatrAIx2026 @YuexingHao
#MatrAIx #PersonalAgents #Agents #Qwen #Alibaba #MetaMuse #OpenAIDots ark #Instinct
0
2
8
239
@MatrAIx2026 is co-hosting our first hackathon this Friday, with @OpenAI, @awscloud, @nebiusai, @UofBeta and @KeenableAI.
Our Open Challenge 🧪: run MatrAIx Persona 1M (huggingface.co/datasets/MatrA…) against image/audio/video gen models and find where they break.
Fri Oct 23 · 6:30 - r Loft
luma.com/beta-79jb
Our Open Challenge 🧪: run MatrAIx Persona 1M (huggingface.co/datasets/MatrA…) against image/audio/video gen models and find where they break.
Fri Oct 23 · 6:30 - r Loft
luma.com/beta-79jb
2
1
8
318
MatrAIx retweeted
🎉🎉🎉 Great honor! Nobel prize winner, Prof. Vernon L. Smith, just joined MatrAIx as a senior advisor!!!
@MatrAIx2026
#MatrAIx #NobelPrize #Agents #PersonalAgents #AI #LLM #Google #OpenAI #Meta #Anthropic #xAI #Microsoft #Hark
@MatrAIx2026
#MatrAIx #NobelPrize #Agents #PersonalAgents #AI #LLM #Google #OpenAI #Meta #Anthropic #xAI #Microsoft #Hark
2
1
15
1,221
RT @XiaominLi98: Thank you to @GoogleDeepMind for the invitation! It was a pleasure to give a talk sharing @MatrAIx2026 's work on Persona…
0
1
0
10
We just launched PersonalAgentEval!🚀🚀🚀 (personalagenteval.com) The first benchmark built to evaluate personal agents across diverse users, preferences, and domains!
We benchmarked 7 popular personal agents across everyday workflows (e.g. booking dinners, paying bills, ch agent operated within accounts pre-loaded with 13 months of realistic history (347–470 records per person). To pass, an agent has to proactively spot crucial background details: like an unread message, a clause in a contract, or a promise made in a past chat. Each task and user simulation are validated and verified by real human.
Key Findings:
• 🏆 Overall Leaders: @instinct and @OpenAI Dots topped the leaderboard.
• 🔍 Proactive Context Retrieval: @instinct excelled at uncovering crucial background details—finding 64 of 72 key facts before the user brought them up.
• ⚡ Speed & Efficiency: @grok @bot was both fast and effective, averaging 0.8 minutes of active work per task with a 50% completion rate. @Meta @Muse required the fewest average interaction turns.
• ⚠️ Missing Promised Steps: Even leading agents missed crucial follow-up steps. In one task, all 7 agents correctly identified a pharmacy mentioned in past logs, but only @grok @bot and @Meta @Muse remembered to notify the caregiver as promised.
Explore the full leaderboard and read our case study and analysis:
personalagenteval.com
(Building a personal AI agent? Reach out to us if you’d like to get your agent evaluated and featured on the leaderboard.)
#MatrAIx #AI #AIAgents #MachineLearning #LLMs #MetaMuse #GrokBot #OpenAIDots #Hark #ManusCue #Instinct
We benchmarked 7 popular personal agents across everyday workflows (e.g. booking dinners, paying bills, ch agent operated within accounts pre-loaded with 13 months of realistic history (347–470 records per person). To pass, an agent has to proactively spot crucial background details: like an unread message, a clause in a contract, or a promise made in a past chat. Each task and user simulation are validated and verified by real human.
Key Findings:
• 🏆 Overall Leaders: @instinct and @OpenAI Dots topped the leaderboard.
• 🔍 Proactive Context Retrieval: @instinct excelled at uncovering crucial background details—finding 64 of 72 key facts before the user brought them up.
• ⚡ Speed & Efficiency: @grok @bot was both fast and effective, averaging 0.8 minutes of active work per task with a 50% completion rate. @Meta @Muse required the fewest average interaction turns.
• ⚠️ Missing Promised Steps: Even leading agents missed crucial follow-up steps. In one task, all 7 agents correctly identified a pharmacy mentioned in past logs, but only @grok @bot and @Meta @Muse remembered to notify the caregiver as promised.
Explore the full leaderboard and read our case study and analysis:
personalagenteval.com
(Building a personal AI agent? Reach out to us if you’d like to get your agent evaluated and featured on the leaderboard.)
#MatrAIx #AI #AIAgents #MachineLearning #LLMs #MetaMuse #GrokBot #OpenAIDots #Hark #ManusCue #Instinct
5
4
20
22,171
Thank you to everyone who joined our Personal Agent dinner in San Francisco this week!
We were honored to host @DanRothNLP, Chief AI Scientist at @Oracle and a great friend of @MatrAIx2026! Our amazing guests also included directors and scientists from @Apple, @OpenAI, Spark, @Meta Muse, @Microsoft, @hark_labs, @amazon, @nvidia, @Adobe, @AMD, @theworldlabs, and many other teams building agents. We were also delighted to welcome professors from @Stanford, @UW, and UPenn.
The conversations about how to evaluate personal agents were inspiring. We are grateful for this community and for everyone’s interest in MatrAIx. We can’t wait to host the next dinner! A special thank-you to our co-host, @UofBeta.
#PersonalAgents #AIEvaluation #SFTechWeek #Agent #MetaMuse #MatrAIx
We were honored to host @DanRothNLP, Chief AI Scientist at @Oracle and a great friend of @MatrAIx2026! Our amazing guests also included directors and scientists from @Apple, @OpenAI, Spark, @Meta Muse, @Microsoft, @hark_labs, @amazon, @nvidia, @Adobe, @AMD, @theworldlabs, and many other teams building agents. We were also delighted to welcome professors from @Stanford, @UW, and UPenn.
The conversations about how to evaluate personal agents were inspiring. We are grateful for this community and for everyone’s interest in MatrAIx. We can’t wait to host the next dinner! A special thank-you to our co-host, @UofBeta.
#PersonalAgents #AIEvaluation #SFTechWeek #Agent #MetaMuse #MatrAIx
0
0
16
1,682
We're excited to share that @MatrAIx2026 has joined the @GoogleStartups program!
The support will help us scale our persona agent evaluation infrastructure, which helps AI agent companies see how different users actually experience their models. Thank you to the @Google for ng us!
#SFTechWeek #GoogleForStartups #MatrAIx
The support will help us scale our persona agent evaluation infrastructure, which helps AI agent companies see how different users actually experience their models. Thank you to the @Google for ng us!
#SFTechWeek #GoogleForStartups #MatrAIx
1
1
13
608
Last week @YuexingHao and @XiaominLi98 gave an internal talk at @Microsoft Word Science Forum. Thank you to the Microsoft Word Applied Science team, especially Mike H. Wang, PhD, for hosting us.
Most agent environments are built around tools and tasks but leave out the human. y a lot in how they ask, push back, and judge an answer. @MatrAIx2026 persona agents sit in the user's seat and interact with the agent the way different people actually would. It runs on Persona 8B (8.3 billion personas across 1,290 attributes), with 1,010 tasks over 25+ domains and 91.5% attribute adherence in a controlled study.
Really enjoyed the questions from the Word team. It's great to see industry teams thinking hard about evaluation and how it feeds back into model improvement.
#AIAgents #AIEvaluation #MatrAIx
Most agent environments are built around tools and tasks but leave out the human. y a lot in how they ask, push back, and judge an answer. @MatrAIx2026 persona agents sit in the user's seat and interact with the agent the way different people actually would. It runs on Persona 8B (8.3 billion personas across 1,290 attributes), with 1,010 tasks over 25+ domains and 91.5% attribute adherence in a controlled study.
Really enjoyed the questions from the Word team. It's great to see industry teams thinking hard about evaluation and how it feeds back into model improvement.
#AIAgents #AIEvaluation #MatrAIx
0
0
6
427
Last Sunday @MatrAIx2026 Prof @XiaominLi98 spent the afternoon at the Top AI Researchers Roundtable in Los Altos Hills. We had a closed-door conversation with researchers from @OpenAI, @GoogleDeepMind, @AnthropicAI, @spacexai and other frontier labs.
As frontier models saturate rks, are we still measuring intelligence, or only measuring optimization? We think about this every day at @MatrAIx2026.
Our bet is that evaluation has to look more like the real world. That means running large populations of simulated users through a personal agent or AI model, so teams can see how different segments actually experience their model before it ships.
#AI #AIAgents #AIEvaluation #MatrAIx
As frontier models saturate rks, are we still measuring intelligence, or only measuring optimization? We think about this every day at @MatrAIx2026.
Our bet is that evaluation has to look more like the real world. That means running large populations of simulated users through a personal agent or AI model, so teams can see how different segments actually experience their model before it ships.
#AI #AIAgents #AIEvaluation #MatrAIx
0
0
7
391
JOIN US 🤩
In SF for COLM?
@MatrAIx2026 will co-host a fine-dining dinner with @UofBeta on Oct 8 for researchers working on personal agents and evals.
20 seats only. Chat on the new evaluation paradigm for AI models and personal agents.
Luma link below:
luma.com/gu3dn35h
0
0
4
290
The @MatrAIx2026 team had a great time at @agihouse_org!
We are thrilled to be interviewed by @SuJinyan6 in an in-depth interview at the Enterprise Deployment Build Day, where we shared how MatrAIx is redefining evaluation for personal agents and frontier AI models. The industry models on static benchmarks. We're building the infrastructure that lets AI labs and big techs stress-test their systems against millions of simulated users, surfacing the failures and edge cases that only show up in real product journeys.
This is the hardest and most dynamic problem in AI evaluation, and we believe it's the one that will decide which agents actually make it into production.
Read the full interview here:
agihouse.org/blog/matraix-t…
#AIEvaluation #AIAgents #LLM #AIStartup #AGIHouse #FrontierAI
We are thrilled to be interviewed by @SuJinyan6 in an in-depth interview at the Enterprise Deployment Build Day, where we shared how MatrAIx is redefining evaluation for personal agents and frontier AI models. The industry models on static benchmarks. We're building the infrastructure that lets AI labs and big techs stress-test their systems against millions of simulated users, surfacing the failures and edge cases that only show up in real product journeys.
This is the hardest and most dynamic problem in AI evaluation, and we believe it's the one that will decide which agents actually make it into production.
Read the full interview here:
agihouse.org/blog/matraix-t…
#AIEvaluation #AIAgents #LLM #AIStartup #AGIHouse #FrontierAI
0
1
16
4,047
Today we presented @MatrAIx2026 at @netflix, sharing our work on a critical but largely missing layer of AI development: the space between the model and the application.
Evaluating AI models and agents before launch is hard. Static benchmarks often fall short of real-world ation with real humans is slow and costly. Too often, teams only discover problems once a product is already in production.
@MatrAIx2026 closes that gap by running large-scale persona agents through a model or digital product before it launches. Each agent is a simulated user representing a different segment, so teams can run failure mode analysis across user groups and catch problems before real users ever see them.
A big thank you to the amazing @netflix team, especially Dr. Yaochen Zhu, for inviting us and for a great discussion about potential applications of MatrAIx!
#AI #AIAgents #Evaluation #MatrAIx #Netflix
Evaluating AI models and agents before launch is hard. Static benchmarks often fall short of real-world ation with real humans is slow and costly. Too often, teams only discover problems once a product is already in production.
@MatrAIx2026 closes that gap by running large-scale persona agents through a model or digital product before it launches. Each agent is a simulated user representing a different segment, so teams can run failure mode analysis across user groups and catch problems before real users ever see them.
A big thank you to the amazing @netflix team, especially Dr. Yaochen Zhu, for inviting us and for a great discussion about potential applications of MatrAIx!
#AI #AIAgents #Evaluation #MatrAIx #Netflix
0
0
19
690
MatrAIx retweeted
From prevention trials to population simulation
With MatrAIx we'll use >1000 agents to model screening, participation, lifestyle modification, barriers, asking, Who benefits?
#Diabetes #DiabetesPrevention #Obesity #CardiometabolicHealth #AIinHealthcare
With MatrAIx we'll use >1000 agents to model screening, participation, lifestyle modification, barriers, asking, Who benefits?
#Diabetes #DiabetesPrevention #Obesity #CardiometabolicHealth #AIinHealthcare
0
1
2
29
@MatrAIx2026 is proud to share that we are now an official part of @MIT @MIT_CSAIL Alliance now!
Through MIT @csail_alliances, we'll push this frontier in the open: building evaluation benchmarks for persona agents, publishing where simulated users match real ones (and where researchers a reliable and scalable way to study behavior at population scale.
Go visit: cap.csail.mit.edu/current-members
Through MIT @csail_alliances, we'll push this frontier in the open: building evaluation benchmarks for persona agents, publishing where simulated users match real ones (and where researchers a reliable and scalable way to study behavior at population scale.
Go visit: cap.csail.mit.edu/current-members
1
1
20
1,228
Thanks @thevalley101 sharing @MatrAIx2026, featuring our amazing co-organizers @YuexingHao and @XiaominLi98.
As AI models and digital products ship in ever-faster iterations, MatrAIx provides reliable and scalable evaluation infrastructure that powers the data flywheel.
As AI models and digital products ship in ever-faster iterations, MatrAIx provides reliable and scalable evaluation infrastructure that powers the data flywheel.
如果AI不仅能回答问题，还能“复制”人类、模拟社会，甚至把尚未发生的未来提前跑一遍，会发生什么？
从只有25个Agent的“斯坦福小镇”，到最近哈佛和MIT共同发布的MatrAIx所构建的83亿个虚拟人格，AI Simulation正在把过去只存在于科幻作品中的“模拟世界”，变成现实中的技术赛道。
这期视频，我们会拆解当前AI模拟的不同技术路线：有的追求更像真人的“数字分身”，有的关注Agent自行演化出的社会关系，也有研究从评估入手试图为模拟打牢“地基”。与此同时，Simile、Aaru等创业公司也开始探索AI模拟的商业化，并逐渐形成“卖模拟”和“卖预测”两大方向，已经开始将AI模拟应用到产品测试、市场研究和未来预测等不同场景。
那么，如果可以先在虚拟世界里把未来跑一遍，AI模拟能否真正帮助人类做决策？从“像人”到“预测未来”，它还面临哪些难题？更有意思的是，当越来越多Agent拥有记忆、关系和目标，我们究竟是在模拟一个世界，还是正在创造一个新的世界？
采访嘉宾：
孟醒，五源资本合伙人
Satya Nitta，Emergence AI CEO
高森泉，WorldVac CEO
Xiaomin Li，MatrAIx联合创始人、微软高级研究科学家
Yuexing Hao，MatrAIx联合创始人、MIT EECS博士后
youtube.com/watch?v=--ePbk…
1
3
10
1,025
Our co-organizers @XiaominLi98 and @YuexingHao gave a talk at @amazon Prime Video last week on @MatrAIx2026.
MatrAIx runs large-scale persona agents as a "prescreening" layer before AI model and digital systems launch. Each persona agent provides both the preference and the "users liked it", MatrAIx provides specific reasoning "this segment churned at step 3, this one never found the button."
Thanks to the Prime Video team for the insightful comments. Time to simulate before reality!
#MatrAIx #AIAgents #Amazon #PrimeVideo #Harvard #MIT #AI #LLM
MatrAIx runs large-scale persona agents as a "prescreening" layer before AI model and digital systems launch. Each persona agent provides both the preference and the "users liked it", MatrAIx provides specific reasoning "this segment churned at step 3, this one never found the button."
Thanks to the Prime Video team for the insightful comments. Time to simulate before reality!
#MatrAIx #AIAgents #Amazon #PrimeVideo #Harvard #MIT #AI #LLM
0
0
8
444