@CaltechLCSSP
psychology-of-ai.github.io Representational Control over Self-Report & Behavior Coherence in LLM Risk-Taking Project page for our paper on the representational basis of self-report--behavior dissociation in LLMs. New paper: "Representational Control over Self-Report & Behavior Coherence in LLM Risk-Taking"
๐ฏ ๐๐ผ ๐๐๐ ๐ ๐ต๐ฎ๐๐ฒ ๐ฎ ๐ฟ๐ถ๐๐ธ-๐๐ฎ๐ธ๐ถ๐ป๐ด ๐ฑ๐ถ๐๐ฝ๐ผ๐๐ถ๐๐ถ๐ผ๐ป ๐๐ต๐ฎ๐ ๐ฐ๐ผ๐ต๐ฒ๐ฟ๐ฒ๐ป๐๐น๐ ๐ด๐๐ถ๐ฑ๐ฒ๐ ๐ฏ๐ผ๐๐ต ๐๐ต๐ฒ๐ถ๐ฟ ๐๐ฒ๐น๐ณ-๐ฟ๐ฒ๐ฝ๐ผ๐ฟ๐๐ ๐ฎ๐ป๐ฑ ๐๐ต๐ฒ๐ถ๐ฟ ๐ฟ๐ถ๐๐ธ๐ ๐ฑ๐ฒ๐ฐ๐ถ๐๐ถ๐ผ๐ป๐?
๐ง๐;๐๐ฅ
๐ช๐ต๐ ๐ถ๐ ๐บ๐ฎ๐๐๐ฒ๐ฟ๐: 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.
๐ช๐ต๐ฎ๐ ๐๐ฒ ๐๐ต๐ผ๐: 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.
๐ง๐ต๐ฒ ๐ณ๐ถ๐ : Steering internal representations of stated and enacted risk together makes what a model says match what it does, without retraining.
๐ง๐ต๐ฒ ๐ฑ๐ฎ๐ป๐ด๐ฒ๐ฟ: 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.
๐ Paper: arxiv.org/abs/2610.04125
๐ Website: psychology-of-ai.github.io/representationโฆ
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 ๐
#AgenticAI #AISafety #AIAlignment #Interpretability #ActivationSteeringView quoted post on X
www.youtube.com How to prepare for the midterms | Votebeat Event The midterm elections are already underway. Elections in the United... 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 โฌ๏ธView quoted post on X
