Masahiko Haruno<\/a><\/p>\n\n\n\nAbstract:
As AI agents move into multi-turn, multi-agent settings, their performance depends on inductive biases and on modeling other agents rather than single-turn benchmark accuracy. I will present results from finitely repeated games showing stable social signatures and discuss methods that can shift equilibrium selection toward cooperative conventions and increase perceived humanness. I then situate these behaviors within limits of intuitive psychology revealed by multimodal social-inference tasks. Finally, I frame in-context learning as function approximation under explicit priors, making model biases measurable and steerable through targeted post-training. Taken together, these results motivate interactive evaluation that characterizes and shapes social inference in LLMs, with the goal of building agents that coordinate and collaborate rather than merely optimize their individual reward.<\/p>\n","protected":false},"featured_media":0,"template":"","acf":[],"_links":{"self":[{"href":"http:\/\/cinetjp-static3.nict.go.jp\/english\/wp-json\/wp\/v2\/event\/4801"}],"collection":[{"href":"http:\/\/cinetjp-static3.nict.go.jp\/english\/wp-json\/wp\/v2\/event"}],"about":[{"href":"http:\/\/cinetjp-static3.nict.go.jp\/english\/wp-json\/wp\/v2\/types\/event"}],"version-history":[{"count":7,"href":"http:\/\/cinetjp-static3.nict.go.jp\/english\/wp-json\/wp\/v2\/event\/4801\/revisions"}],"predecessor-version":[{"id":4864,"href":"http:\/\/cinetjp-static3.nict.go.jp\/english\/wp-json\/wp\/v2\/event\/4801\/revisions\/4864"}],"wp:attachment":[{"href":"http:\/\/cinetjp-static3.nict.go.jp\/english\/wp-json\/wp\/v2\/media?parent=4801"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}