Michael Kirchhof

Michael Kirchhof, Enkelejda Kasneci, Seong Joon Oh · apple-ml

Contrastively trained encoders have recently been proven to invert the data-generating process: they encode each input, e.g., an image, into the true latent vector that generated the image (Zimmermann et al., 2021). However, real-world observations often have inherent ambiguities. For instance, images may be blurred or only show a 2D view of a 3D object, so multiple latents could have generated them. This makes the true posterior for the latent vector probabilistic with heteroscedastic uncertainty. In this setup, we extend the common InfoNCE objective and encoders to predict latent distributions instead of points. We prove that these distributions recover the correct posteriors of the data-generating process, including its level of aleatoric uncertainty, up to a rotation of the latent space. In addition to providing calibrated uncertainty estimates, these posteriors allow the computation of credible intervals in image retrieval. They comprise images with the same latent as a given query, subject to its uncertainty. Code is available at https://github.com/mkirchhof/Probabilistic_Contrastive_Learning

Michael Kirchhof, Karsten Roth, Zeynep Akata et al. · apple-ml

Proxy-based Deep Metric Learning (DML) learns deep representations by embedding images close to their class representatives (proxies), commonly with respect to the angle between them. However, this disregards the embedding norm, which can carry additional beneficial context such as class- or image-intrinsic uncertainty. In addition, proxy-based DML struggles to learn class-internal structures. To address both issues at once, we introduce non-isotropic probabilistic proxy-based DML. We model images as directional von Mises-Fisher (vMF) distributions on the hypersphere that can reflect image-intrinsic uncertainties. Further, we derive non-isotropic von Mises-Fisher (nivMF) distributions for class proxies to better represent complex class-specific variances. To measure the proxy-to-image distance between these models, we develop and investigate multiple distribution-to-point and distribution-to-distribution metrics. Each framework choice is motivated by a set of ablational studies, which showcase beneficial properties of our probabilistic approach to proxy-based DML, such as uncertainty-awareness, better-behaved gradients during training, and overall improved generalization performance. The latter is especially reflected in the competitive performance on the standard DML benchmarks, where our approach compares favorably, suggesting that existing proxy-based DML can significantly benefit from a more probabilistic treatment. Code is available at github.com/ExplainableML/Probabilistic_Deep_Metric_Learning.

Michael Kirchhof, Bálint Mucsányi, Seong Joon Oh et al. · apple-ml

Representation learning has significantly driven the field to develop pretrained models that can act as a valuable starting point when transferring to new datasets. With the rising demand for reliable machine learning and uncertainty quantification, there is a need for pretrained models that not only provide embeddings but also transferable uncertainty estimates. To guide the development of such models, we propose the Uncertainty-aware Representation Learning (URL) benchmark. Besides the transferability of the representations, it also measures the zero-shot transferability of the uncertainty estimate using a novel metric. We apply URL to evaluate eleven uncertainty quantifiers that are pretrained on ImageNet and transferred to eight downstream datasets. We find that approaches that focus on the uncertainty of the representation itself or estimate the prediction risk directly outperform those that are based on the probabilities of upstream classes. Yet, achieving transferable uncertainty quantification remains an open challenge. Our findings indicate that it is not necessarily in conflict with traditional representation learning goals. Code is provided under https://github.com/mkirchhof/url .

Tobias Leemann, Michael Kirchhof, Yao Rong et al. · apple-ml

Interest in understanding and factorizing learned embedding spaces through conceptual explanations is steadily growing. When no human concept labels are available, concept discovery methods search trained embedding spaces for interpretable concepts like object shape or color that can provide post-hoc explanations for decisions. Unlike previous work, we argue that concept discovery should be identifiable, meaning that a number of known concepts can be provably recovered to guarantee reliability of the explanations. As a starting point, we explicitly make the connection between concept discovery and classical methods like Principal Component Analysis and Independent Component Analysis by showing that they can recover independent concepts under non-Gaussian distributions. For dependent concepts, we propose two novel approaches that exploit functional compositionality properties of image-generating processes. Our provably identifiable concept discovery methods substantially outperform competitors on a battery of experiments including hundreds of trained models and dependent concepts, where they exhibit up to 29 % better alignment with the ground truth. Our results highlight the strict conditions under which reliable concept discovery without human labels can be guaranteed and provide a formal foundation for the domain. Our code is available online.

Bálint Mucsányi, Michael Kirchhof, Elisa Nguyen et al. · apple-ml

As machine learning technology gets applied to actual products and solutions, new challenges have emerged. Models unexpectedly fail to generalize to small changes in the distribution, tend to be confident on novel data they have never seen, or cannot communicate the rationale behind their decisions effectively with the end users. Collectively, we face a trustworthiness issue with the current machine learning technology. This textbook on Trustworthy Machine Learning (TML) covers a theoretical and technical background of four key topics in TML: Out-of-Distribution Generalization, Explainability, Uncertainty Quantification, and Evaluation of Trustworthiness. We discuss important classical and contemporary research papers of the aforementioned fields and uncover and connect their underlying intuitions. The book evolved from the homonymous course at the University of Tübingen, first offered in the Winter Semester of 2022/23. It is meant to be a stand-alone product accompanied by code snippets and various pointers to further sources on topics of TML. The dedicated website of the book is https://trustworthyml.io/.

Michael Kirchhof · apple-ml

Uncertainty quantification is a key pillar of trustworthy machine learning. It enables safe reactions under unsafe inputs, like predicting only when the machine learning model detects sufficient evidence, discarding anomalous data, or emitting warnings when an error is likely to be inbound. This is particularly crucial in safety-critical areas like medical image classification or self-driving cars. Despite the plethora of proposed uncertainty quantification methods achieving increasingly higher scores on performance benchmarks, uncertainty estimates are often shied away from in practice. Many machine learning projects start from pretrained latent representations that come without uncertainty estimates. Uncertainties would need to be trained by practitioners on their own, which is notoriously difficult and resource-intense. This thesis makes uncertainty estimates easily accessible by adding them to the latent representation vectors of pretrained computer vision models. Besides proposing approaches rooted in probability and decision theory, such as Monte-Carlo InfoNCE (MCInfoNCE) and loss prediction, we delve into both theoretical and empirical questions. We show that these unobservable uncertainties about unobservable latent representations are indeed provably correct. We also provide an uncertainty-aware representation learning (URL) benchmark to compare these unobservables against observable ground-truths. Finally, we compile our findings to pretrain lightweight representation uncertainties on large-scale computer vision models that transfer to unseen datasets in a zero-shot manner. Our findings do not only advance the current theoretical understanding of uncertainties over latent variables, but also facilitate the access to uncertainty quantification for future researchers inside and outside the field, enabling straightforward but trustworthy machine learning.

Bálint Mucsányi, Michael Kirchhof, Seong Joon Oh · apple-ml

Uncertainty quantification, once a singular task, has evolved into a spectrum of tasks, including abstained prediction, out-of-distribution detection, and aleatoric uncertainty quantification. The latest goal is disentanglement: the construction of multiple estimators that are each tailored to one and only one source of uncertainty. This paper presents the first benchmark of uncertainty disentanglement. We reimplement and evaluate a comprehensive range of uncertainty estimators, from Bayesian over evidential to deterministic ones, across a diverse range of uncertainty tasks on ImageNet. We find that, despite recent theoretical endeavors, no existing approach provides pairs of disentangled uncertainty estimators in practice. We further find that specialized uncertainty tasks are harder than predictive uncertainty tasks, where we observe saturating performance. Our results provide both practical advice for which uncertainty estimators to use for which specific task, and reveal opportunities for future research toward task-centric and disentangled uncertainties. All our reimplementations and Weights & Biases logs are available at https://github.com/bmucsanyi/untangle.

Metod Jazbec, Theo X. Olausson, Louis Béthune et al.

Diffusion (Large) Language Models (dLLMs) now match the downstream performance of their autoregressive counterparts on many tasks, while holding the promise of being more efficient during inference. One particularly successful variant is masked discrete diffusion, in which a buffer filled with special mask tokens is progressively replaced with tokens sampled from the model's vocabulary. Efficiency can be gained by unmasking several tokens in parallel, but doing too many at once risks degrading the generation quality. Thus, one critical design aspect of dLLMs is the sampling procedure that selects, at each step of the diffusion process, which tokens to replace. Indeed, recent work has found that heuristic strategies such as confidence thresholding lead to both higher quality and token throughput compared to random unmasking. However, such heuristics have downsides: they require manual tuning, and we observe that their performance degrades with larger buffer sizes. In this work, we instead propose to train sampling procedures using reinforcement learning. Specifically, we formalize masked diffusion sampling as a Markov decision process in which the dLLM serves as the environment, and propose a lightweight policy architecture based on a single-layer transformer that maps dLLM token confidences to unmasking decisions. Our experiments show that these trained policies match the performance of state-of-the-art heuristics when combined with semi-autoregressive generation, while outperforming them in the full diffusion setting. We also examine the transferability of these policies, finding that they can generalize to new underlying dLLMs and longer sequence lengths. However, we also observe that their performance degrades when applied to out-of-domain data, and that fine-grained tuning of the accuracy-efficiency trade-off can be challenging with our approach.

Szilvia Ujváry, Louis Béthune, Pierre Ablin et al.

Language models have consistently grown to compress more world knowledge into their parameters, but the knowledge that can be pretrained into them is upper-bounded by their parameter size. Especially the capacity of Small Language Models (SLMs) is limited, leading to factually incorrect generations. This problem is often mitigated by giving the SLM access to an outside source: the ability to query a larger model, documents, or a database. Under this setting, we study the fundamental question of \emph{which tokens an SLM can and should learn} during pretraining, versus \emph{which ones it should delegate} via a \texttt{<CALL>} token. We find that this is not simply a question of loss: although the loss is predictive of whether a predicted token mismatches the ground-truth, some tokens are \emph{acceptable} in that they are truthful alternative continuations of a pretraining document, and should not trigger a \texttt{<CALL>} even if their loss is high. We find that a spaCy grammar parser can help augment the loss signal to decide which tokens the SLM should learn to delegate to prevent factual errors and which are safe to learn and predict even under high losses. We propose LaCy, a novel pretraining method based on this token selection philosophy. Our experiments demonstrate that LaCy models successfully learn which tokens to predict and where to delegate for help. This results in higher FactScores when generating in a cascade with a bigger model and outperforms Rho or LLM-judge trained SLMs, while being simpler and cheaper.

Andrew Szot, Michael Kirchhof, Omar Attia et al.

Reinforcement learning (RL) has demonstrated notable success in post-training large language models (LLMs) as agents for tasks such as computer use, tool calling, and coding. However, exploration remains a central challenge in RL for LLM agents, especially as they operate in language-action spaces with complex observations and sparse outcome rewards. In this work, we address exploration for LLM agents by leveraging the ability of LLMs to plan and reason in language about the environment to shift exploration from low-level actions to higher-level language strategies. We thus propose Strategy-Guided Exploration (SGE), which first generates a concise natural-language strategy that describes what to do to make progress toward the goal, and then generates environment actions conditioned on that strategy. By exploring in the space of strategies rather than the space of actions, SGE induces structured and diverse exploration that targets different environment outcomes. To increase strategy diversity during RL, SGE introduces mixed-temperature sampling, which explores diverse strategies in parallel, along with a strategy reflection process that grounds strategy generation on the outcomes of previous strategies in the environment. Across UI interaction, tool-calling, coding, and embodied agent environments, SGE consistently outperforms exploration-focused RL baselines, improving both learning efficiency and final performance. We show that SGE enables the agent to learn to solve tasks too difficult for the base model.

Michael Kirchhof, Mark Collier, Seong Joon Oh et al. · apple-ml

Accurate uncertainty estimation is vital to trustworthy machine learning, yet uncertainties typically have to be learned for each task anew. This work introduces the first pretrained uncertainty modules for vision models. Similar to standard pretraining this enables the zero-shot transfer of uncertainties learned on a large pretraining dataset to specialized downstream datasets. We enable our large-scale pretraining on ImageNet-21k by solving a gradient conflict in previous uncertainty modules and accelerating the training by up to 180x. We find that the pretrained uncertainties generalize to unseen datasets. In scrutinizing the learned uncertainties, we find that they capture aleatoric uncertainty, disentangled from epistemic components. We demonstrate that this enables safe retrieval and uncertainty-aware dataset visualization. To encourage applications to further problems and domains, we release all pretrained checkpoints and code under https://github.com/mkirchhof/url .

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Preetum Nakkiran, Arwen Bradley, Adam Goliński et al.

Large Language Models (LLMs) often lack meaningful confidence estimates for their outputs. While base LLMs are known to exhibit next-token calibration, it remains unclear whether they can assess confidence in the actual meaning of their responses beyond the token level. We find that, when using a certain sampling-based notion of semantic calibration, base LLMs are remarkably well-calibrated: they can meaningfully assess confidence in open-domain question-answering tasks, despite not being explicitly trained to do so. Our main theoretical contribution establishes a mechanism for why semantic calibration emerges as a byproduct of next-token prediction, leveraging a recent connection between calibration and local loss optimality. The theory relies on a general definition of "B-calibration," which is a notion of calibration parameterized by a choice of equivalence classes (semantic or otherwise). This theoretical mechanism leads to a testable prediction: base LLMs will be semantically calibrated when they can easily predict their own distribution over semantic answer classes before generating a response. We state three implications of this prediction, which we validate through experiments: (1) Base LLMs are semantically calibrated across question-answering tasks, (2) RL instruction-tuning systematically breaks this calibration, and (3) chain-of-thought reasoning breaks calibration. To our knowledge, our work provides the first principled explanation of when and why semantic calibration emerges in LLMs.

Zihuiwen Ye, Lukas Aichberger, Michael Kirchhof et al.

Large Language Models (LLMs) are increasingly deployed to autonomously solve real-world tasks. A key ingredient for this is the LLM Function-Calling paradigm, a widely used approach for equipping LLMs with tool-use capabilities. However, an LLM calling functions incorrectly can have severe implications, especially when their effects are irreversible, e.g., transferring money or deleting data. Hence, it is of paramount importance to consider the LLM's confidence that a function call solves the task correctly prior to executing it. Uncertainty Quantification (UQ) methods can be used to quantify this confidence and prevent potentially incorrect function calls. In this work, we present what is, to our knowledge, the first evaluation of UQ methods for LLM Function-Calling (FC). While multi-sample UQ methods, such as Semantic Entropy, show strong performance for natural language Q&A tasks, we find that in the FC setting, it offers no clear advantage over simple single-sample UQ methods. Additionally, we find that the particularities of FC outputs can be leveraged to improve the performance of existing UQ methods in this setting. Specifically, multi-sample UQ methods benefit from clustering FC outputs based on their abstract syntax tree parsing, while single-sample UQ methods can be improved by selecting only semantically meaningful tokens when calculating logit-based uncertainty scores.

Andrea Santilli, Adam Golinski, Michael Kirchhof et al.

Uncertainty Quantification (UQ) in Language Models (LMs) is key to improving their safety and reliability. Evaluations often use metrics like AUROC to assess how well UQ methods (e.g., negative sequence probabilities) correlate with task correctness functions (e.g., ROUGE-L). We show that mutual biases--when both UQ methods and correctness functions are biased by the same factors--systematically distort evaluation. First, we formally prove that any mutual bias non-randomly skews AUROC rankings, compromising benchmark integrity. Second, we confirm this happens empirically by testing 7 widely used correctness functions, from lexical-based and embedding-based metrics to LM-as-a-judge approaches, across 4 datasets x 4 models x 8 UQ methods. Our analysis shows that length biases in correctness functions distort UQ assessments by interacting with length biases in UQ methods. We identify LM-as-a-judge methods as the least length-biased, offering a promising path for a fairer UQ evaluation.

Michael Kirchhof, Gjergji Kasneci, Enkelejda Kasneci · apple-ml

Large-language models (LLMs) and chatbot agents are known to provide wrong outputs at times, and it was recently found that this can never be fully prevented. Hence, uncertainty quantification plays a crucial role, aiming to quantify the level of ambiguity in either one overall number or two numbers for aleatoric and epistemic uncertainty. This position paper argues that this traditional dichotomy of uncertainties is too limited for the open and interactive setup that LLM agents operate in when communicating with a user, and that we need to research avenues that enrich uncertainties in this novel scenario. We review the literature and find that popular definitions of aleatoric and epistemic uncertainties directly contradict each other and lose their meaning in interactive LLM agent settings. Hence, we propose three novel research directions that focus on uncertainties in such human-computer interactions: Underspecification uncertainties, for when users do not provide all information or define the exact task at the first go, interactive learning, to ask follow-up questions and reduce the uncertainty about the current context, and output uncertainties, to utilize the rich language and speech space to express uncertainties as more than mere numbers. We expect that these new ways of dealing with and communicating uncertainties will lead to LLM agent interactions that are more transparent, trustworthy, and intuitive.

Deepro Choudhury, Sinead Williamson, Adam Goliński et al. · oxford

We propose a general-purpose approach for improving the ability of Large Language Models (LLMs) to intelligently and adaptively gather information from a user or other external source using the framework of sequential Bayesian experimental design (BED). This enables LLMs to act as effective multi-turn conversational agents and interactively interface with external environments. Our approach, which we call BED-LLM (Bayesian Experimental Design with Large Language Models), is based on iteratively choosing questions or queries that maximize the expected information gain (EIG) about the task of interest given the responses gathered previously. We show how this EIG can be formulated (and then estimated) in a principled way using a probabilistic model derived from the LLM's predictive distributions and provide detailed insights into key decisions in its construction and updating procedure. We find that BED-LLM achieves substantial gains in performance across a wide range of tests based on the 20 questions game and using the LLM to actively infer user preferences, compared to direct prompting of the LLM and other adaptive design strategies.

Waiss Azizian, Michael Kirchhof, Eugene Ndiaye et al.

Large Language Models (LLMs) have demonstrated impressive generalization capabilities across various tasks, but their claim to practical relevance is still mired by concerns on their reliability. Recent works have proposed examining the activations produced by an LLM at inference time to assess whether its answer to a question is correct. Some works claim that a "geometry of truth" can be learned from examples, in the sense that the activations that generate correct answers can be distinguished from those leading to mistakes with a linear classifier. In this work, we underline a limitation of these approaches: we observe that these "geometries of truth" are intrinsically task-dependent and fail to transfer across tasks. More precisely, we show that linear classifiers trained across distinct tasks share little similarity and, when trained with sparsity-enforcing regularizers, have almost disjoint supports. We show that more sophisticated approaches (e.g., using mixtures of probes and tasks) fail to overcome this limitation, likely because activation vectors commonly used to classify answers form clearly separated clusters when examined across tasks.

Hadi Pouransari, David Grangier, C Thomas et al.

The impressive performance gains of modern language models currently rely on scaling parameters: larger models store more world knowledge and reason better. Yet compressing all world knowledge into parameters is unnecessary, as only a fraction is used per prompt, and impractical for edge devices with limited inference-time memory and compute. We address this shortcoming by a memory-augmented architecture and a pretraining strategy aligned with existing hardware paradigms. We introduce small language models that access large hierarchical parametric memory banks encoding world knowledge. During pretraining and inference, we fetch a small, context-dependent memory block and add it to the model. Our pretraining learns to store long-tail world knowledge in the memory parameters, while the small language model acts as an anchor capturing common knowledge and general reasoning abilities. Through trillion-token-scale experiments, we show significant gains: a 160M-parameters model augmented with an 18M-parameters memory fetched from a 4.6B memory bank obtains comparable performance to a regular model with more than 2x the parameters. Through extensive experiments, we study the optimal type and size of parametric memories in transformers, scaling them to over 21B parameters. We find that our proposed hierarchical feed-forward memories work robustly across transformer architectures, whether added during pretraining or post-hoc.

Michael Kirchhof, Lena Schmid, Christopher Reining et al.

A key task in multi-label classification is modeling the structure between the involved classes. Modeling this structure by probabilistic and interpretable means enables application in a broad variety of tasks such as zero-shot learning or learning from incomplete data. In this paper, we present the probabilistic rule stacking learner (pRSL) which uses probabilistic propositional logic rules and belief propagation to combine the predictions of several underlying classifiers. We derive algorithms for exact and approximate inference and learning, and show that pRSL reaches state-of-the-art performance on various benchmark datasets. In the process, we introduce a novel multicategorical generalization of the noisy-or gate. Additionally, we report simulation results on the quality of loopy belief propagation algorithms for approximate inference in bipartite noisy-or networks.

Michael Kirchhof, Klaus Haas, Thomas Kornas et al.

The production of lithium-ion battery cells is characterized by a high degree of complexity due to numerous cause-effect relationships between process characteristics. Knowledge about the multi-stage production is spread among several experts, rendering tasks as failure analysis challenging. In this paper, a new method is presented that includes expert knowledge acquisition in production ramp-up by combining Failure Mode and Effects Analysis (FMEA) with a Bayesian Network. Special algorithms are presented that help detect and resolve inconsistencies between the expert-provided parameters which are bound to occur when collecting knowledge from several process experts. We show the effectiveness of this holistic method by building up a large scale, cross-process Bayesian Failure Network in lithium-ion battery production and its application for root cause analysis.