Caltech LCSSP

New paper: "Representational Control over Self-Report & Behavior Coherence in LLM Risk-Taking" 🎯 𝗗𝗼 𝗟𝗟𝗠𝘀 𝗵𝗮𝘃𝗲 𝗮 𝗿𝗶𝘀𝗸-𝘁𝗮𝗸𝗶𝗻𝗴 𝗱𝗶𝘀𝗽𝗼𝘀𝗶𝘁𝗶𝗼𝗻 𝘁𝗵𝗮𝘁 𝗰𝗼𝗵𝗲𝗿𝗲𝗻𝘁𝗹𝘆 𝗴𝘂𝗶𝗱𝗲𝘀 𝗯𝗼𝘁𝗵 𝘁𝗵𝗲𝗶𝗿 𝘀𝗲𝗹𝗳-𝗿𝗲𝗽𝗼𝗿𝘁𝘀 𝗮𝗻𝗱 𝘁𝗵𝗲𝗶𝗿 𝗿𝗶𝘀𝗸𝘆 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻𝘀? 𝗧𝗟;𝗗𝗥 𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀: LLMs often say one thing about themselves and do another. In risk-taking the gap is costly: a clinical assistant or a trading agent can call itself cautious and still make risky decisions. 𝗪𝗵𝗮𝘁 𝘄𝗲 𝘀𝗵𝗼𝘄: No shared disposition: inside the model, stated risk-taking and risky choices are represented separately. Task-specific instructions seem to move both, but through misleading shortcuts like producing larger numbers. 𝗧𝗵𝗲 𝗳𝗶𝘅: Steering internal representations of stated and enacted risk together makes what a model says match what it does, without retraining. 𝗧𝗵𝗲 𝗱𝗮𝗻𝗴𝗲𝗿: The same mechanism can make a model say the opposite of what it does (report low risk while taking more, or the reverse), in every model we tested. 📄 Paper: arxiv.org/abs/2610.04125 🌐 Website: psychology-of-ai.github.io/r… w/ @p_song1 @pengrui_han @MyrlGMarmarelis, Ramit Debnath, Dean Mobbs, @rmichaelalvarez, @CaltechLCSSP and @Cambridge_Uni's Centre for Human-Inspired AI (CHIA) Technical details below 👇 #AgenticAI #AISafety #AIAlignment #Interpretability #ActivationSteering

LAST CHANCE! Sign up to attend a free event with @JessicaHuseman, @aztammyp, @rmichaelalvarez, and Bob Page, registrar of voters for Orange County, California. link below ⬇️

New Research Alert! @DannyCEbanks and I have been studying the dynamics of US elections for some time, here's our current perspective of persuadable voters as we head into the final two weeks of the US midterm elections! frontiersin.org/journals/pol…

Interested in learning more about why we see "blue shifts" and "red shifts" in post-election voting returns? Give our simulation a spin, it's based on published research with @yimenglipolisci and synthetic data. rmichaelalvarez.github.io/bl…

A new Caltech lab will soon inform environmental policy. The lab is being established through a $7 million infusion of funding from the WHH Foundation, a private family philanthropic organization. Read more: caltech.edu/about/news/7-mil…

📣 Registration is Open! Join the @Caltech Election Night Watch Party! 🗳️ Follow federal, state, & local election results: 📅 Nov. 3, 2026 ⏰ 7–10 PM 📍 Avery Courtyard 🍕 Free food • 🎮 Games ✅ Register now: forms.gle/fh4RV9QU2zUyxKK26. 🔴 Nonpartisan

1/4 Recently @BeaMagistro, @Jacob_Morrier , and I published an interesting study, "The Convergence of Zero-Sum Thinking and Illiberalism in the United States." It's open-access at SSQ! onlinelibrary.wiley.com/doi/…

A few notes, we give you the chance to build your own forecast going back to 1946. If you think there is another polling error like in 2016, or want to see what the model would say in 1956, try all these possibilities, make your own modeling decisions, and see the numbers change!