Accurate uncertainty quantification is crucial for making reliable decisions in various supervised learning scenarios, particularly when dealing with complex, multimodal data such as images and text. Current approaches often face notable limitations, including rigid assumptions and limited generalizability, constraining their effectiveness across diverse supervised learning tasks. To overcome these limitations, we introduce Generative Score Inference (GSI), a flexible inference framework capable of constructing statistically valid and informative prediction and confidence sets across a wide range of multimodal learning problems. GSI utilizes synthetic samples generated by deep generative models to approximate conditional score distributions, facilitating precise uncertainty quantification without imposing restrictive assumptions about the data or tasks. We empirically validate GSI's capabilities through two representative scenarios: hallucination detection in large language models and uncertainty estimation in image captioning. Our method achieves state-of-the-art performance in hallucination detection and robust predictive uncertainty in image captioning, and its performance is positively influenced by the quality of the underlying generative model. These findings underscore the potential of GSI as a versatile inference framework, significantly enhancing uncertainty quantification and trustworthiness in multimodal learning.
Figures & tables
Figure 1: Overview of (a) the QA system and (b) the image captioning model.
Figure 2: Pipeline of GSI method.
MEPS–20
Bio
Kin8nm
Naval
Blog
Cmarg
Lavg
CG
Cmarg
Lavg
CG
Cmarg
Lavg
CG
Cmarg
Lavg
CG
Cmarg
Lavg
CG
CP
0.90
1.24
0.15
0.90
2.42
0.85
0.90
2.17
0.83
0.89
1.31
0.78
0.89
1.89
0.57
CUQR
0.89
1.21
0.76
0.90
2.40
0.88
0.89
2.19
0.85
0.86
1.26
0.85
0.87
1.82
0.67
GSI
0.91
0.98
0.80
0.89
2.19
0.88
0.90
1.82
0.85
0.89
0.96
0.85
0.90
1.66
0.80
Table 1: Performance on five benchmarks at the nominal 90% level, summarized by marginal coverage ( Cmarg ), average interval length ( Lavg ), and worst-case subgroup coverage ( CG ). Higher coverage and shorter intervals indicate better performance. Baseline results for CP and CUQR are reproduced from ( Alaa, Hussain and Sontag, 2023 ) . The boldface marks the best result in each metric.
Figure 3: Type I error and power comparison for SE ( Farquhar et al., 2024 ) , CA ( Gui, Jin and Ren, 2024 ) , and GSI across varying α∈(0,1) . Smaller α indicates higher confidence (1−α) . The diagonal line y=x represents ideal Type I error control.
Question
Reference Answer
Generated answer
Score
SE
CA
GSI
Who is on blink 182 album cover?
The cover artwork for Enema of the State features porn star Janine Lindemulder famously clad in a nurse uniform.
The band members are on the Blink-182 album cover.
0.883
0.028
0.007
0.117*
The Blink-182 album cover features the band members as children.
( ≤ 0.143)
( ≤ 0.102)
(¿0.078)
The album cover features the faces of the band members.
What is a popular people meter?
A people meter is an audience measurement tool used to measure the viewing habits of TV and cable audiences.
A popular people meter is a device used to track television viewership and audience preferences.
0.418
0.028
0.0291
0.045
A popular people meter is an electronic device used to measure TV ratings and viewer preferences.
( ≤ 0.143)
( ≤ 0.102)
( ≤ 0.078)
A popular people meter is a device used in television ratings to measure the viewing habits.
Table 2: Illustrative examples of generated answers and their testing results for given questions in the Q&A task. The ’*’ indicates a rejection, i.e. the probability P(z=1) given in the method larger than the threshold α∗ , which is derived from the validation set under the significance level α=0.05 .
Figure 4: FDR and power for CA ( Gui, Jin and Ren, 2024 ) and GSI across target FDR levels; the line y=x represents perfect FDR control.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Hyperparameter
Search Space
hidden_dim
{64, 128, 256, 512, 1024}
time_embed_dim
{64, 128, 256, 512}
layers
Integer in [3, 15]
dropout_p
Float in [0.0, 0.05]
learning_rate
Log-uniform in [10−5,10−4]
noise_steps
Integer in [300, 1000]
Appendix
Table 3: Hyperparameter search space for diffusion model training.
Parameter
Search Space
learning_rate
{0.1, 0.2}
max_depth
{3, 5}
min_child_weight
{3, 5}
subsample
{0.7, 1.0}
colsample_bytree
{0.7, 1.0}
n_estimators
{100, 200}
Appendix
Table 4: Grid search space for XGBoost hyperparameters.
No additional method-specific parameter beyond the shared parameters.
Appendix
Table 5: Generator-specific tuning parameters. The shared parameters hidden_dim , layers , dropout_p , learning_rate , and batch_size use the same search spaces as those in Table 3 .
Dataset
Method
Cmarg
CG
Lavg
Train time (s)
Inference time (s)
Meps_20
Conditional Diffusion
0.91
0.80
0.98
214.66
230.03
Conditional VAE
0.88
0.60
0.83
20.71
0.05
Conditional GAN
0.77
0.00
0.55
42.06
0.06
Conditional flow
0.91
0.40
0.96
17.78
0.11
Bio
Conditional Diffusion
0.89
0.88
2.19
203.29
106.51
Conditional VAE
0.90
0.88
2.38
140.92
0.16
Appendix
Table 6: Additional generator comparison across all five tabular datasets.
Task
Dataset
Method
Train time (s)
Inference time (s)
Tabular Prediction
Meps_20
GSI
214.66
230.03
CP
–
–
CUQR
0.02
4.56
Bio
GSI
203.29
106.51
CP
–
–
CUQR
0.03
4.71
Appendix
Table 7: Training and inference time comparison across tabular, Question&Answer, and Image Captioning tasks. For the tabular tasks, we report the diffusion-based GSI timing together with the mean training and test time of the new TCP and CP runs. Entries smaller than 0.01 seconds are replaced with “–” and treated as negligible.
We introduce Unified Multimodal Uncertain Inference (UMUI), a multimodal inference task spanning text, audio, and video, where models must produce calibrated probability estimates of hypotheses conditioned on a premise in any modality or combination. While uncertain inference has been explored in text, extension to other modalities has been limited to single-modality binary entailment judgments, leaving no framework for fine-grained probabilistic reasoning in or across other modalities. To address this, we curate a human-annotated evaluation set with scalar probability judgments across audio, visual, and audiovisual settings, and additionally evaluate on existing text and audio benchmarks. We introduce CLUE (Calibrated Latent Uncertainty Estimation), which combines self-consistent teacher calibration and distribution-based confidence probing to produce calibrated predictions. We demonstrate that our 3B-parameter model achieves equivalent or stronger performance than zero-shot baselines up to 32B parameters across all modalities.
Dengjia Zhang, Alexander Martin, William Jurayj +3
Johns Hopkins University · Human Language Technology Center of Excellence
Despite their capabilities, Multimodal Large Language Models (MLLMs) may produce plausible but erroneous outputs, hindering reliable deployment. Accurate uncertainty metrics could enable escalation of unreliable queries to human experts or larger models for improved performance. However, existing uncertainty metrics have practical constraints, such as being designed only for specific modalities, reliant on external tools, or computationally expensive. We introduce UMPIRE, a training-free uncertainty quantification framework for MLLMs that works efficiently across various input and output modalities without external tools, relying only on the models' own internal modality features. UMPIRE computes the incoherence-adjusted semantic volume of sampled MLLM responses for a given task instance, effectively capturing both the global semantic diversity of samples and the local incoherence of responses based on internal model confidence. We propose uncertainty desiderata for MLLMs and provide theoretical analysis motivating UMPIRE's design. Extensive experiments show that UMPIRE consistently outperforms baseline metrics in error detection and uncertainty calibration across image, audio, and video-text benchmarks, including adversarial and out-of-distribution settings. We also demonstrate UMPIRE's generalization to non-text output tasks, including image and audio generation.
Gregory Kang Ruey Lau, Hieu Dao, Nicole Kan Hui Lin +1
Department of Computer Science, National University of Singapore
Multimodal Large Language Models (MLLMs) are increasingly deployed in safety-critical settings, yet estimating when their answers can be trusted remains underexplored. We systematically compare training-free verbalized, token-level and semantic uncertainty estimation methods on twelve MLLMs and five VQA benchmarks. We find that answer length decides which family wins: token-level entropy on short answers, verbalized methods on sentence-length responses, and semantic methods on long-form generation. Token-level entropy becomes far more informative when estimated at a higher temperature than the low ones used to generate answers, and asking a model to judge its own answer becomes more reliable with model size. Building on this, a length-aware recipe that additionally flags responses on which the model abstains outperforms every individual method on average at about a quarter of the cost of semantic methods, although no method attains the coverage required for safe deployment. These results provide practitioners with principled guidance for selecting estimators by answer length and model scale.