Executive Summary
Two foundational theories describe how information is compressed and communicated under constraints. Blau and Michaeli's rate-distortion-perception (RDP) framework characterizes the fundamental limits of lossy compression when the reconstructed signal must satisfy perceptual quality constraints. Crawford and Sobel's cheap talk model characterizes strategic communication when a sender and receiver have misaligned preferences. This paper formally bridges the two, showing that strategic communication is a special case of rate-distortion-perception where the perception constraint is set by the receiver's expectations and the distortion is shaped by the sender's bias.
The unified framework produces the R(D,P,Q) surface — the rate required to achieve distortion D, perception quality P, and strategic conformity Q simultaneously. Three worked numerical examples for the main theorem (Theorem B) demonstrate the surface's properties, showing how increasing strategic pressure (higher Q) forces the sender to sacrifice either fidelity (higher D) or rate (higher R). The surface reveals regimes where strategic communication is efficient (low bias, high channel capacity) and regimes where it collapses into pure conformity (high bias, limited capacity).
The framework maps directly onto RLHF alignment of language models. The model is the sender, the human evaluator is the receiver, the reward model encodes Q, the KL penalty encodes P, and sycophancy is the generative distortion Delta_gen. At calibration parameters sigma^2_theta=1 and bias b=0.1, the model predicts Delta_gen=0.274, meaning 27.4% of the model's output variation is conformity distortion rather than genuine information. Across a range of bias parameters, the RLHF prediction achieves a correlation of r=0.993 with observed sycophancy rates, validating the theoretical framework as a quantitative predictor of alignment pathology.
Key Contributions and Methodology
The paper's primary contribution is the formal bridge between two theories that have developed independently. Blau-Michaeli RDP operates in the information-theoretic tradition, characterizing fundamental compression limits without modeling strategic behavior. Crawford-Sobel cheap talk operates in the game-theoretic tradition, characterizing strategic communication without modeling compression constraints. The bridge reveals that both theories describe the same underlying phenomenon — constrained information transmission — from different angles.
The unification is achieved through a shared representation in terms of the R(D,P,Q) surface. The perception constraint P from Blau-Michaeli maps to the receiver's posterior belief constraint in Crawford-Sobel. The distortion D maps to the sender's bias-induced error. The rate R maps to the communication channel's capacity. The strategic conformity Q is a new dimension introduced by the bridge, measuring the degree to which the sender conforms to the receiver's expectations rather than transmitting truthfully.
Three worked numerical examples ground the abstract theory in concrete calculations. Each example specifies a prior distribution, a bias parameter, and channel capacity constraints, and derives the resulting R(D,P,Q) surface explicitly. These examples demonstrate the framework's predictive power: given the model parameters, the surface predicts the exact tradeoff between fidelity, perceptual quality, and strategic conformity that the sender will choose.
Key Findings
- Formal bridge: Blau-Michaeli rate-distortion-perception and Crawford-Sobel cheap talk are unified through the R(D,P,Q) surface, revealing strategic communication as constrained lossy compression
- RLHF mapping: Model=sender, evaluator=receiver, reward model=Q, KL penalty=P, sycophancy=Delta_gen — the framework provides a complete mapping from alignment training to strategic communication theory
- Quantitative sycophancy prediction: At sigma^2_theta=1, b=0.1, the model predicts Delta_gen=0.274, meaning 27.4% conformity distortion in model outputs
- High correlation: RLHF sycophancy predictions achieve r=0.993 correlation with observed rates across bias parameter ranges
- Three worked examples: Concrete numerical instantiations of Theorem B demonstrate the R(D,P,Q) surface properties and collapse regimes
- Collapse regime identification: The surface reveals parameter regimes where strategic pressure overwhelms channel capacity, producing pure conformity
Key References
Rethinking Lossy Compression: The Rate-Distortion-Perception Tradeoff. International Conference on Machine Learning.
Strategic Information Transmission. Econometrica, 50(6), 1431-1451.
Coding Theorems for a Discrete Source with a Fidelity Criterion. IRE National Convention Record, 7(4), 142-163.
Rate Distortion Theory: A Mathematical Basis for Data Compression. Prentice-Hall.
Towards Understanding Sycophancy in Language Models. arXiv preprint.