Statistical inference on black-box generative models in the data kernel perspective space
By Hayden S. Helm, Aranyak Acharyya, Youngser Park, Brandon Duderstadt, Carey E. Priebe. Annual Meeting of the Association for Computational Linguistics, 2024
12 citations recorded by the source corpus.
Generative models are capable of producing human-expert level content across a variety of topics and domains. As the impact of generative models grows, it is necessary to develop statistical methods to understand collections of available models. These methods are particularly important in settings where the user may not have access to information related to a model's pre-training data, weights, or other relevant model-level covariates. In this paper we extend recent results on representations of black-box generative models to model-level statistical inference tasks. We demonstrate that the model-level representations are effective for multiple inference tasks.
Subfields: Statistics and Probability, Generative Models
View Paper: https://arxiv.org/abs/2410.01106
Citations in the corpus
- Jailbreak susceptibility prediction and mitigation via the behavioral geometry of models (2026, unknown): “Low-dimensional representations of language models have been constructed from internal activations [33, 22, 18, 10], model weights [2], and embeddings of model responses [16, 14, 1, 5].”
- Control Charts for Multi-agent Systems (2026, unknown): “The Data Kernel Perspective Space (DKPS) embeds models based on response similarities and supports statistical inference in the black-box setting [1, 11].”
- Black-box model classification under the discriminative factorization (2026, unknown): “To demonstrate the impact of query set on model representations and subsequent inference, we use a task inspired by Helm et al. [16]: using only black-box access and a collection of labeled models, can we identify whether a particular topic appeared in a model’s fine-tuning corpus?”
- Recovering manifold structure in LLM responses through a joint Euclidean mirror (2026, unknown): “The setting in Helm et al. (2025) and related literature, such as Acharyya et al. (2024, 2025), is similar to ours: they embed large language models by defining distances between their responses to a fixed set of queries and then applying raw-stress MDS to embed the resulting distance matrix into…”
- Detecting Perspective Shifts in Multi-agent Systems (2025, unknown): “When T = 1 , TDKPS reduces to the non-dynamic Data Kernel Perspective Space framework first introduced in (Helm et al., 2024) and analyzed in (Acharyya et al., 2025; Helm et al., 2025a).”
- A Model of the Language Process (2026, unknown): “We call this curve the "temporal control curve," and conjecture that the model vectors constituting this curve can be applied to tasks including model comparison (Duderstadt et al., 2024), model level inference (Helm et al., 2025), and model perspective tracking (Helm et al., 2024).”
- Data Kernel Perspective Space Performance Guarantees for Synthetic Data from Transformer Models (2026, unknown): “…means via µ , and where we use the machinery of classical Multidimensional Scaling (MDS) [40] to get the representation Ψ of our collection of n models (model f ( i ) represented as the i -th row of Ψ ) as points in a low-dimensional Euclidean space that is amenable to subsequent inference [38].”
- Testing for LLM response differences: the case of a composite null consisting of semantically irrelevant query perturbations (2025, unknown): “…accessibility assumptions – for example, it is possible to determine if a model was trained on a particular type of data when token-wise log-probabilities of a relevant set of tokens are available (Shi et al., 2023) as well as when users only have access to model responses (Helm et al., 2025).”
- Extracting information from fine-tuned weights (unknown): “More recent work has shown that DKPS has desirable statistical properties [1] and can be applied to model-level inference problems such as predicting properties of a model’s training data [12].”