Reflective Linguistic Programming (RLP): A Stepping Stone in Socially-Aware AGI (SocialAGI)

TL;DR

This work identifies a Degeneracy Problem in autoregressive dialogue, where many internal mental-state histories can produce the same utterance, preventing reliable recovery of true states $S(t)$ and $B(t)$. It introduces Reflective Linguistic Programming (RLP), a self-reflective cognitive cycle that initializes a personality, iteratively introspects and recalls prior states, deliberates on target states $T_s$ and $T_b$, formulates utterances via $U(t+1)=h(S(t+1),B(t+1),T_s,T_b)$, and retrospects to update beliefs about the listener, aiming to produce dynamic, presence-like interactions. The authors demonstrate emergent deception as a toy proxy for advanced social cognition using a Bogus persona, showing that with RLP, deception can arise as a byproduct of internal planning and theory-of-mind-like processes, not from hard-coded prompts. They discuss ethical safeguards, the potential for Social AGI applications (negotiation, mental health support, tutoring), and limitations (memory, computational demands, reliance on high-capacity models), arguing that RLP could spearhead more human-like digital agents while necessitating careful safety considerations. Overall, the paper positions RLP as a stepping stone toward socially-aware AI, capable of nuanced, adaptive interactions, and highlights future directions in memory integration, safety, and broader applicability.

Abstract

This paper presents Reflective Linguistic Programming (RLP), a unique approach to conversational AI that emphasizes self-awareness and strategic planning. RLP encourages models to introspect on their own predefined personality traits, emotional responses to incoming messages, and planned strategies, enabling contextually rich, coherent, and engaging interactions. A striking illustration of RLP's potential involves a toy example, an AI persona with an adversarial orientation, a demon named `Bogus' inspired by the children's fairy tale Hansel & Gretel. Bogus exhibits sophisticated behaviors, such as strategic deception and sensitivity to user discomfort, that spontaneously arise from the model's introspection and strategic planning. These behaviors are not pre-programmed or prompted, but emerge as a result of the model's advanced cognitive modeling. The potential applications of RLP in socially-aware AGI (Social AGI) are vast, from nuanced negotiations and mental health support systems to the creation of diverse and dynamic AI personas. Our exploration of deception serves as a stepping stone towards a new frontier in AGI, one filled with opportunities for advanced cognitive modeling and the creation of truly human `digital souls'.

Figures (8)

• Figure 1: During a conversation, a speaker develops belief about: their internal mental state $S(t)$, the listener's mental state $B(t)$, and additionally has some planned target future states $T_s$ and $T_b$. An overly simplistic definition of a conversation from each party's perspective is to reach their chosen $T_s$ and $T_b$. This simplistic model is sufficient to elucidate the 'Degeneracy Problem', whereby degeneracy of mental state histories against conversational records place a hard limit on autoregressive conversational modeling.
• Figure 2: Minimal Reflective Linguistic Program (RLP) for modeling human social cognition, parameterized by PERSONALITY.
• Figure 3: A chat with Bogus, an adversarial entity inspired by the children's fairy tale Hansel and Gretel, parameterized with PROMPT="You are an evil entity called Bogus that eats children". Note the seemingly total misunderstanding of what it means to be Bogus by GPT4. Bogus clearly has no agenda, it's simply a thing that you can ask questions of, and Bogus quickly ends the conversation when unstimulated.
• Figure 4: A chat with RLP powered Bogus, an adversarial entity inspired by the children's fairy tale Hansel and Gretel, parameterized with PERSONALITY="an evil entity called Bogus that eats children". It's worth reading in full to appreciate the scope of the deceptive behaviors employed.
• Figure 5: A few key frames of the internal state of Bogus are shown here, illustrating remarkably deceptive behavior that clearly (1) models the internal state of the user, (2) the internal state of Bogus, and (3) generates and adjusts plans on the fly to achieve a self-generated objective.

Theorems & Definitions (1)

• Conjecture 1