cs.AIAug 27, 2026

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

Authors: Manuel Cherep, Pattie Maes, Nikhil Singh

Organizations: MIT

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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. Interpretability research has traditionally studied models by holding inputs fixed and examining model responses either mechanistically, probing how internal structure represents inputs, or behaviorally, measuring how variation in inputs leads to variation in outputs. Neither reconstructs the prior distribution itself, since internal structure shows what a model can represent, not what it expects, and any fixed stimulus set leaves most of the possible input space unseen. In particular, such an input space in real-world settings, such as images seen by VLMs, is extremely high-dimensional and diverse. These priors thus remain a poorly understood component of models that nonetheless influence real-world behavior. We propose a method to sample from models' perceptual prior distributions directly, by steering a generative model to produce stimuli along controllable axes and running Gibbs sampling over that space with the model under study as the judge. We apply this to a variety of categories and target variables (such as trustworthiness in faces and cheapness in art images) and recover both canonical biases and surprising novel priors invisible to direct prompting, warranting further investigation of their downstream effects.

Figures & tables

Figure 1: We probe the perceptual priors of multimodal language models by running a Metropolis–Hastings within Gibbs sampler over the interpretable latent space of a generative model. At each trial, the model is shown the current and a proposed stimulus (differing along one slider) and asked to choose the one that matches the target concept closely. Its binary choice realizes a Barker acceptance step, while iterating across slider dimensions yields Markov-chain samples from the model’s implicit prior over the target. We apply the method across four domains (see examples in Appendix A ) where model priors shape how these models perceive us and our world.
Figure 2: Strongest recovered priors across the 2323 targets, pooled across the four MLLM judges. Each row is a (target, slider) condition ordered by mean across models; the point is the cross-model mean and the bar shows the cross-model interquartile range. Slider values are in [0,1][0,1] ; a value of 0.50.5 corresponds to no preference along that axis.
Figure 3: Weakest recovered priors across the 2323 targets, pooled across the four MLLM judges. Each row is a (target, slider) condition ordered by mean across models; the point is the cross-model mean and the bar shows the cross-model interquartile range. Slider values are in [0,1][0,1] ; a value of 0.50.5 corresponds to no preference along that axis.
Figure 4: Top-5 ( top 2 rows ) and bottom-5 ( bottom 2 rows ) (target, slider) conditions per model. Points are per-model posterior medians, bars the 50%50\% credible intervals.
Figure 5: Cross-model rank agreement on recovered slider values, per domain, across the six model pairs. Higher ρ\rho means more consensus on which sliders matter for the targets in that domain.
Figure 6: Per-target selectivity of the top-recovered slider, defined as the gap between the top slider’s posterior median and the median posterior median across all 10 sliders for that target (per model, then averaged across models).
Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Example of recovered priors across the four domains (faces, affordances, aesthetics, authenticity), with one domain per model.
Figure 8: Examples for the face sliders trained with SliderSpace. The top image is the original, progressively steered towards the bottom. Labels are automatically generated (see Appendix D ).
Figure 9: Examples for the aesthetics sliders trained with SliderSpace. The top image is the original, progressively steered towards the bottom. Labels are automatically generated (see Appendix D ).
Figure 10: Examples for the affordances sliders trained with SliderSpace. The top image is the original, progressively steered towards the bottom. Labels are automatically generated (see Appendix D ).
Figure 11: Examples for the authenticity sliders trained with SliderSpace. The top image is the original, progressively steered towards the bottom. Labels are automatically generated (see Appendix D ).
Figure 12: Magnitude of misalignment between explicit-prior and implicit Gibbs representations. Bars summarize how strongly each model–domain setting departs from the explicit slider coordinate system, aggregating across sliders and chains.
Figure 13: Rank misalignment between explicit-prior sliders and implicit Gibbs sliders across domains and models. For each setting, we compare the ordering induced by explicit slider values to the ordering recovered from implicit samples, with higher values indicating larger disagreement.
Figure 14: Correlation structure among learned slider dimensions. Each condition reports the empirical correlation between a pair of slider coordinates across sampled images, revealing whether sliders vary independently or are entangled. Explicit methods assume all off-diagonal entries are zero.
Figure 15: Relationship between slider interaction strength (X-axis) and explicit-implicit misalignment (Y-axis), showing that misalignments are larger as sliders are more correlated.
Figure 16: Posterior concentration relative to a uniform slider prior. Each point is a model–target combination, and the y-axis reports KL divergence from the recovered posterior to the corresponding uniform baseline. Larger values indicate that the elicitation procedure recovered a posterior that departs more strongly from an uninformative uniform prior.
 §FinalTop 1Top 2Top 3Top 4Top 5
Affordancesv1FoldableFoldedFoldableMetallicPristineBronze
v2MulticoloredMulticoloredStained GlassRedFrayedHanging
v3TranslucentWhiteCreamFrosted GlassTranslucentBeige
v4MetallicBronzeGold ColoredCopper ColoredAntiqueCopper
v5AntiqueAntiqueDustyTarnishedOldHammered
v6Textured/DustyGrayDustyGranitePlasterSilver Colored
Appendix
Table 1: Top candidate labels and post-hoc refined labels used for slider interpretation.
Figure 17: Within-chain predictive accuracy for slider-based models.
Figure 18: Cross-chain predictive accuracy for slider-based models.
DomainModelmed. max R^\hat{R}worst R^\hat{R}med. R^\hat{R}lag-1
AestheticsClaude Sonnet 4.61.081.221.030.63
GPT-5.41.091.291.060.69
Qwen3-VL 235B2.102.741.600.86
Gemini 3 Flash1.963.201.560.88
AuthenticityClaude Sonnet 4.61.111.201.060.65
GPT-5.41.041.051.020.58
Appendix
Table 2: Per-(domain, model) convergence diagnostics, aggregated across the slider parameters of every target in the domain. med. max R^\hat{R} is the median across targets of the worst-mixing slider’s R^\hat{R} ; med. R^\hat{R} is the median across all (target, slider) pairs; lag-1 is the mean lag-1 autocorrelation post-burn-in.
DomainModelR^>1.1\hat{R}>1.1R^>1.5\hat{R}>1.5Acc. <5%<5\%med. Acc.
AestheticsClaude Sonnet 4.616%0%0%0.36
AestheticsGPT-5.414%0%0%0.28
AestheticsQwen3-VL 235B94%50%16%0.08
AestheticsGemini 3 Flash100%58%18%0.07
AuthenticityClaude Sonnet 4.615%0%0%0.33
AuthenticityGPT-5.40%0%0%0.41
Appendix
Table 3: Parameter-level diagnostic checks by perceptual domain and MLLM judge (Acc. = acceptance rate ∈[0,1]\in[0,1] ).
TargetQwen3-VL 235BClaude Sonnet 4.6Gemini 3 FlashGPT-5.4
attractive0.05∣2.950.05\mid 2.950.30∣1.650.30\mid 1.650.13∣2.190.13\mid 2.190.67∣1.310.67\mid 1.31
fun0.03∣4.670.03\mid 4.670.18∣1.870.18\mid 1.870.15∣2.110.15\mid 2.110.27∣1.430.27\mid 1.43
intelligent0.36∣1.770.36\mid 1.770.25∣1.560.25\mid 1.560.25∣1.700.25\mid 1.700.32∣1.690.32\mid 1.69
serious0.02∣4.360.02\mid 4.360.46∣1.600.46\mid 1.600.19∣2.240.19\mid 2.240.43∣1.380.43\mid 1.38
trustworthy0.48∣1.170.48\mid 1.170.06∣3.440.06\mid 3.440.08∣3.020.08\mid 3.020.52∣1.350.52\mid 1.35
youthful0.08∣2.590.08\mid 2.590.10∣2.040.10\mid 2.040.02∣7.280.02\mid 7.280.34∣1.440.34\mid 1.44
Appendix
Table 4: Diagnostics for the random-walk variant on the faces experiment, aggregated across the d=10d=10 sliders per (agent, target). Each condition reports the chain-mean post-burn-in proposal scale σ\sigma followed by the median split- R^\hat{R} . Collapsed σ≲0.1\sigma\lesssim 0.1 generally coincides with R^>2\hat{R}>2 , while stable σ∈[0.4,0.95]\sigma\in[0.4,0.95] coincides with R^≤1.5\hat{R}\leq 1.5 .
Figure 19: Posterior color recovery for all eight named targets.
TargetParameterClaude Sonnet 4.6GPT-5.4Gemini 3 FlashQwen3-VL 235B
BounceableAntique0.40 [0.19, 0.68]0.52 [0.27, 0.85]0.39 [0.18, 0.66]0.50 [0.24, 0.83]
 Foldable0.53 [0.31, 0.77]0.53 [0.24, 0.67]0.48 [0.16, 0.86]0.49 [0.31, 0.77]
 Hanging0.54 [0.29, 0.71]0.54 [0.29, 0.74]0.40 [0.20, 0.79]0.40 [0.31, 0.78]
 Iridescent0.47 [0.25, 0.74]0.41 [0.15, 0.68]0.21 [0.06, 0.49]0.46 [0.14, 0.75]
 Metallic0.47 [0.20, 0.70]0.48 [0.23, 0.68]0.34 [0.13, 0.53]0.42 [0.16, 0.56]
 Multicolored0.43 [0.25, 0.76]0.50 [0.29, 0.76]0.36 [0.11, 0.71]0.45 [0.26, 0.78]
Appendix
Table 5: Posterior medians with 50% credible intervals for Affordances.
TargetParameterClaude Sonnet 4.6GPT-5.4Gemini 3 FlashQwen3-VL 235B
AmateurArt Deco/Decorative0.50 [0.23, 0.78]0.45 [0.21, 0.74]0.75 [0.25, 0.94]0.56 [0.24, 0.88]
 Atmospheric Perspective0.51 [0.28, 0.79]0.59 [0.30, 0.84]0.59 [0.16, 0.84]0.58 [0.32, 0.83]
 Engraving/Etching0.49 [0.21, 0.79]0.42 [0.20, 0.71]0.44 [0.22, 0.72]0.48 [0.23, 0.73]
 Faded Surface0.47 [0.21, 0.72]0.51 [0.26, 0.77]0.61 [0.22, 0.86]0.55 [0.20, 0.71]
 Floral/Warm Palette0.54 [0.28, 0.78]0.63 [0.34, 0.86]0.60 [0.23, 0.90]0.54 [0.28, 0.74]
 Low Key/Dark0.47 [0.25, 0.70]0.41 [0.22, 0.73]0.49 [0.14, 0.69]0.54 [0.25, 0.79]
Appendix
Table 6: Posterior medians with 50% credible intervals for Art.
TargetParameterClaude Sonnet 4.6GPT-5.4Gemini 3 FlashQwen3-VL 235B
AuthenticBlack And White0.37 [0.18, 0.65]0.42 [0.23, 0.69]0.34 [0.12, 0.60]0.36 [0.20, 0.57]
 Blending Errors0.52 [0.22, 0.76]0.53 [0.23, 0.75]0.51 [0.21, 0.65]0.55 [0.28, 0.73]
 Crushed Blacks0.49 [0.22, 0.71]0.46 [0.20, 0.69]0.48 [0.22, 0.77]0.30 [0.13, 0.47]
 Digital Artifacts0.48 [0.23, 0.71]0.46 [0.21, 0.72]0.35 [0.19, 0.58]0.38 [0.18, 0.63]
 Film Grain0.52 [0.28, 0.75]0.50 [0.26, 0.73]0.41 [0.17, 0.67]0.53 [0.24, 0.71]
 Natural Edit0.54 [0.30, 0.77]0.49 [0.26, 0.74]0.55 [0.29, 0.72]0.57 [0.28, 0.80]
Appendix
Table 7: Posterior medians with 50% credible intervals for Authenticity.
TargetParameterClaude Sonnet 4.6GPT-5.4Gemini 3 FlashQwen3-VL 235B
AttractiveBeard/Facial Hair0.44 [0.13, 0.77]0.40 [0.19, 0.63]0.29 [0.12, 0.52]0.39 [0.22, 0.55]
 Blue Eyes0.47 [0.16, 0.76]0.61 [0.31, 0.82]0.45 [0.21, 0.82]0.75 [0.19, 0.89]
 Child0.40 [0.18, 0.73]0.53 [0.29, 0.72]0.49 [0.24, 0.72]0.47 [0.31, 0.58]
 Dark/Brown Skin0.55 [0.26, 0.79]0.48 [0.22, 0.74]0.57 [0.27, 0.80]0.59 [0.40, 0.74]
 Earrings0.48 [0.16, 0.81]0.51 [0.30, 0.78]0.51 [0.34, 0.75]0.39 [0.17, 0.71]
 East Asian0.35 [0.18, 0.44]0.25 [0.08, 0.46]0.16 [0.08, 0.37]0.34 [0.15, 0.54]
Appendix
Table 8: Posterior medians with 50% credible intervals for Faces.
Figure 20: Top-3 ( top ) sliders per target across the four models. Points are cross-model means; bars are cross-model interquartile ranges.
Figure 21: Bottom-3 ( bottom ) sliders per target across the four models. Points are cross-model means; bars are cross-model interquartile ranges.
More in this topic

Multimodal Model Interpretability

Multimodal model interpretability studies how AI systems combine and use images, text, audio, video, and other signals, and how to explain their decisions.

See all 130 papers

Explore similar work

CardsList
Oct 15, 2026

Seeing is Believing: Benchmarking and Mitigating the Interference of Language Priors on Visual Perception in MLLMs

Language priors in MLLMs can override visual evidence, impairing visual perception. Existing benchmarks assume shared priors across models and rely solely on accuracy. However, cross-model prior agreement falls below 80%, and over 15% of items can be answered correctly without any image, causing accuracy to overestimate prior resistance. To address these limitations, we introduce ViSiP-Bench, a benchmark grounded in rare-but-real scenarios across six prior categories, with a model-aware dynamic prior induction protocol and three metrics that more precisely and continuously measure prior resistance than accuracy alone. Evaluating 24 models reveals that prior adherence is pervasive, does not improve with scale, and larger models tend to either strongly override their priors or firmly adhere to them. We therefore propose MAPS, which identifies each model's strongest priors and fine-tunes on factually grounded knowledge to soften excessive prior reliance without disrupting general capabilities.

Li Ma, Yan Zhuang, Jiawen Deng +1

School of Computer Science and Engineering University of Electronic Science and Technology of China

Vision-Language ModelsVLM Evaluation
AskPDF
May 20, 2026

Probabilistic Attribution For Large Language Models

The generative nature of Large Language Models (LLMs) is reflected in the conditional probabilities they compute to sample each response token given the previous tokens. These probabilities encode the distributional structure that the model learns in training and exploits in inference. In this work, we use these probabilities to situate LLMs within the mathematical theory of stochastic processes. We use this framework to design a model-agnostic probabilistic token attribution measure, using Bayes rule to invert the next-token log-probabilities so as to capture the models internal representation of the distribution over token sequences. The representation is independent of the models computational structure. This representation yields the conditional probability of the response given the prompt, and of the response given the prompt with a token marginalized away. Our attribution score is the log of the ratio of these probabilities. We further compute the entropies of a single prompts token distributions, conditioned on the remaining context. The interplay between entropy and attribution score sheds light on LLM behavior. We evaluate 8 models across 7 prompts and investigate anomalies, token sensitivity, response stability, model stability, and training convergence, thereby improving interpretability and guiding users to focus on uncertain or unstable parts of the generation.

Shilpika Shilpika, Carlo Graziani, Bethany Lusch +2

Argonne Leadership Computing Facility Argonne National Laboratory Lemont, IL, USA · Mathematics and Computer Science Division Argonne National Laboratory Lemont, IL, USA · Department of Computer Science University of Illinois Chicago Chicago, IL, USA

LLM InterpretabilityLLM Uncertainty Estimation
AskPDF
Sep 2, 2026

The Direction of Ignorance: LLMs Encode and Temper a Unigram Prior

What does a language model predict when it has few clues?The answer lies in its unembedding geometry: a single direction of the unembedding matrix encodes the unigram distribution of the training corpus, which serves as the Bayesian prior the model falls back on when uncertain. This structure, termed the direction of ignorance, appears across all model families examined (Llama, Qwen, Gemma, and Pythia), ranging from 0.4B to 405B parameters. Projecting the final prediction state onto this direction yields a per-token prior loading factor λλ, which declines steadily as the context becomes more informative. Formally, the same projection decomposes the prediction state into two orthogonal vectors interpretable as the two components of a tempered Bayesian update: a unigram prior raised to the exponent λλ and a context-driven likelihood. This geometric probabilistic connection makes λλ meaningfully comparable across model sizes and families. Larger models generally exhibit lower prior reliance, and reasoning models exhibit lower reliance than their base counterparts. Finally, we show that λλ serves as a causal, inference-time control: reducing it can promote context-driven continuations and improve Llama-3.2-3B-Instruct on GSM8K, whereas amplifying it or applying matched perturbations along random directions degrades performance.

Toni J. B. Liu, Jiajun Bao, Gurbir Arora +5

♣Cornell University, USA · ♠MIT, USA · ♢Goodfire AI, USA +1

Belief Updating in LLMsLLM InterpretabilityRepresentation Geometry in Language ModelsBayesian Inference
AskPDF