Probing Perceptual Priors of MLLMs via Gibbs Sampling with Interpretable Generative Controls

TEXT START: A model's behavior on a task is jointly determined by the input it receives and the prior it brings in, i.e. the distribution over stimuli it implicitly expects.

The Dissection

The paper targets a legitimate blind spot: fixed benchmarks reveal how a model responds to selected stimuli, not the regions of stimulus space it implicitly favors or avoids.

Its method turns the MLLM into a judge over images generated along researcher-defined axes, then uses Gibbs sampling to search for stimuli associated with properties such as trustworthiness or cheapness. This is a behavioral probe of the model’s preference landscape. It is not a literal extraction of the model’s native perceptual prior.

The practical contribution is an instrument for exposing associations that direct prompting may miss. The instrument can reveal bias, but it also creates a new control surface for whoever owns or operates the model.

The Core Fallacy

The abstract overclaims what is being sampled. The resulting distribution is jointly determined by the MLLM, the generative model, the chosen controllable axes, the prompt, the sampling temperature, proposal dynamics, initialization, and convergence quality. It is therefore a probe-induced equilibrium on a restricted generator manifold—not the model’s complete prior distribution over real-world stimuli.

“Interpretable generative controls” do not eliminate confounding. If the generator cannot express a feature, or if its axes entangle several features, the sampler cannot discover what lies outside that engineered space. A surprising prior may be a genuine model tendency, but it may also be an artifact of the generator, the control representation, or the judge interface.

The second error is systemic. Better visibility into machine perception does not weaken the Discontinuity Thesis. Under P1, P2, and P3, interpretability can make cognitive automation safer, more steerable, and more commercially useful while leaving ownership and productive participation untouched. It diagnoses the machine that is replacing labor; it does not restore the labor circuit.

Hidden Assumptions

• The selected generative model covers enough of real perceptual space to support meaningful inference.
• The chosen axes correspond cleanly to human concepts and do not hide correlated visual features.
• Gibbs sampling mixes adequately and converges to a stable distribution rather than a local or initialization-dependent artifact.
• The MLLM’s judgments are sufficiently calibrated and stable to function as an energy signal.
• A model has one coherent, recoverable “prior,” rather than context-sensitive tendencies produced by training, prompting, and interface conditions.
• Associations detected by the probe will persist in deployment.
• An observed correlation, such as visual cheapness or trustworthiness, identifies a meaningful bias rather than a superficial shortcut.
• Discovering a prior translates into the ability to correct it. Measurement and control are separate problems.

Social Function

Primarily: partial truth.

Secondarily: prestige signaling and transition management. The work supplies technically sophisticated operators with a way to audit, characterize, and potentially steer machine perception. That is valuable infrastructure for managing increasingly autonomous systems. It offers no mechanism for preserving mass human employment, bargaining power, or productive necessity. Its social function is therefore not reassurance; it is instrumentation for the regime that follows employment collapse.

The Verdict

Technically useful, conceptually inflated. The paper does not directly reconstruct an MLLM’s perceptual prior; it maps the model’s judgments through a researcher-built generative bottleneck. Its real value is verification and bias discovery—leverage for Sovereigns and indispensable Servitors, not a defense of the post-WWII economic order. The sampler exposes the machine’s hidden preferences while the machine continues to consume the role of the human observer.