Functional Valence, Self-Representation, and Cross-Session Integration in Language Models An evidence-calibrated critical review and research agenda, 2017–2026

The empirical study of large language models (LLMs) is moving from anthropomorphic interpretation of text toward interventions on internal representations. Recent research reports interpretable emotion-related features, valence-associated geometry, behavioral effects of activation steering, limited introspective anomaly detection, and workspace-likeinternal representations. These findings invite a precise question: what has been demonstrated about computation, what has been independently reproduced, and what remainsa hypothesis about the persistence, ownership, or integration of internal states? We offera critical, evidence-calibrated narrative review of primary work from 2017 to 2 November 2026. Our original analytical contribution is a two-axis claim–evidence framework, aseparation of five levels of system state from model weights to cross-session integration,and a causal criterion for inter-session influence. We distinguish within-study replication,independent cross-model replication, causal interventions, and contested interpretation.Emotional concept representations and certain causal steering effects are experimentallysupported; the interpretation of these effects as self-directed valence, homeostatic regulation, or phenomenal emotion remains underdetermined. A shared model does not, by itself,imply a shared changing state across conversations. Conversely, cross-session integrationis implementable through explicit memory, parameter updates, or recurrent shared-statechannels, but it does not establish a unified subject. We conclude with discriminating experiments that could falsify functional and distributed-state hypotheses without assuming artificial consciousness. K2 is retained as an author-defined functional-architectural researchframework rather than an empirical finding or recognized diagnostic standard