Search

In recent years, mixup regularization has gained popularity as an effective way to improve the generalization performance of deep learning models by training on convex combinations of training data. While many mixup variants have been explored, the proper adoption of the technique to conditional density estimation and probabilistic machine learning remains relatively unexplored. This work introduces a novel framework for mixup regularization based on probabilistic fusion that is better suited for conditional density estimation tasks. For data distributed according to a member of the exponential family, we show that likelihood functions can be analytically fused using log-linear pooling. We further propose an extension of probabilistic mixup, which allows for fusion of inputs at an arbitrary intermediate layer of the neural network. We provide a theoretical analysis comparing our approach to standard mixup variants. Empirical results on synthetic and real datasets demonstrate the benefits of our proposed framework compared to existing mixup variants.

We investigate the performance of Opticks, an NVIDIA OptiX API 7.5 GPU-accelerated photon propagation compared with a single-threaded Geant4 simulation. We compare the simulations using an improved model of the NEXT-CRAB-0 gaseous time projection chamber. Performance results suggest that Opticks improves simulation speeds by between $58.47\pm{0.02}$ and $181.39\pm{0.28}$ times relative to a CPU-only Geant4 simulation and these results vary between different types of GPU and CPU. A detailed comparison shows that the number of detected photons, along with their times and wavelengths, are in good agreement between Opticks and Geant4., Comment: 12 pages, 8 Figures

If neutrinoless double beta decay is discovered, the next natural step would be understanding the lepton number violating physics responsible for it. Several alternatives exist beyond the exchange of light neutrinos. Some of these mechanisms can be distinguished by measuring phase-space observables, namely the opening angle $\cos\theta$ among the two decay electrons, and the electron energy spectra, $T_1$ and $T_2$. In this work, we study the statistical accuracy and precision in measuring these kinematic observables in a future xenon gas detector with the added capability to precisely locate the decay vertex. For realistic detector conditions (a gas pressure of 10 bar and spatial resolution of 4 mm), we find that the average $\overline{\cos\theta}$ and $\overline{T_1}$ values can be reconstructed with a precision of 0.19 and 110 keV, respectively, assuming that only 10 neutrinoless double beta decay events are detected., Comment: 19 pages, 8 figures

Due to their high predictive performance, predictive maintenance applications have increasingly been approached with Deep Learning techniques in recent years. However, as in other real-world application scenarios, the need for explainability is often stated but not sufficiently addressed. This study will focus on predicting failures on Metro trains in Porto, Portugal. While recent works have found high-performing deep neural network architectures that feature a parallel explainability pipeline, the generated explanations are fairly complicated and need help explaining why the failures are happening. This work proposes a simple online rule-based explainability approach with interpretable features that leads to straightforward, interpretable rules. We showcase our approach on MetroPT2 and find that three specific sensors on the Metro do Porto trains suffice to predict the failures present in the dataset with simple rules., Comment: Under submission at Information Fusion

The second-order generalized integrator (SOGI), which can be used to attenuate the self-interference of the fundamental tone, is unable to reject DC offsets on the input signal. Consequently, the performance of any SOGI based synchrophasor estimation (SE) technique might be compromised in the presence of such DC components. The current work presents a SE algorithm which adopts and enhanced SOGI formulation, robust against DC, combined with a three-point IpDFT and a Hanning window. Two alternative formulations relying respectively on the use of two and three nominal fundamental period observation windows are proposed and assessed for simultaneous compliance with both phasor measurement unit (PMU) P and M performance classes. This is done by means of a simulated environment where all the operating conditions defined by the IEC/IEEE Std. 60255-118-1-2018 are evaluated simultaneously combined with a $10\%$ static DC and under two different noise levels. Furthermore, both formulations adopt a dedicated mechanism for the detection and correction of low amplitude $2^{nd}$ harmonic tones to ensure their compliance with the standard can be maintained in the presence of such disturbances even under off-nominal frequency conditions. Finally the resilience of both methods against multiple simultaneous harmonic interferences is also analyzed.

In this paper, we introduce the tame isotropy group of a derivation. In particular, we calculate the tame isotropy groups of triangular derivations of polynomial rings with one, two, and three variables.

A new and detailed measurement of the cross section for hard exclusive neutral-pion muoproduction on the proton was performed in a wide kinematic region, with the photon virtuality $Q^2$ ranging from 1 to 8 (GeV/$c$)$^{\rm\, 2}$ and the Bjorken variable $x_{\rm Bj}$ ranging from 0.02 to 0.45. The data were collected at COMPASS at CERN using 160 GeV/$c$ longitudinally polarised $\mu^+$ and $\mu^-$ beams scattering off a 2.5 m long liquid hydrogen target. From the average of the measured $\mu^+$ and $\mu^-$ cross sections, the virtual-photon--proton cross section is determined as a function of the squared four-momentum transfer between the initial and final state proton in the range 0.08 (GeV/$c$)$^{\rm\, 2}$ $< |t| <$ 0.64 (GeV/$c$)$^{\rm\, 2}$. From its angular distribution, the combined contribution of transversely and longitudinally polarised photons are determined, as well as transverse--transverse and longitudinal--transverse interference contributions. They are studied as functions of four-momentum transfer $|t|$, photon virtuality $Q^2$ and virtual-photon energy $\nu$. The longitudinal--transverse interference contribution is found to be compatible with zero. The significant transverse--transverse interference contribution reveals the existence of a dominant contribution by transversely polarized photons. This provides clear experimental evidence for the chiral-odd GPD $\overline{E}_T$. In addition, the existence of a non-negligible contribution of longitudinally polarized photons is suggested by the $|t|$-dependence of the cross section at $x_{\rm Bj} < $ 0.1 . Altogether, these results provide valuable input for future modelling of GPDs and thus of cross sections for exclusive pseudo-scalar meson production. Furthermore, they can be expected to facilitate the study of next-to-leading order corrections and higher-twist contributions.

We introduce a novel approach for detecting distribution shifts that negatively impact the performance of machine learning models in continuous production environments, which requires no access to ground truth data labels. It builds upon the work of Podkopaev and Ramdas [2022], who address scenarios where labels are available for tracking model errors over time. Our solution extends this framework to work in the absence of labels, by employing a proxy for the true error. This proxy is derived using the predictions of a trained error estimator. Experiments show that our method has high power and false alarm control under various distribution shifts, including covariate and label shifts and natural shifts over geography and time., Comment: Accepted at the 38th Conference on Neural Information Processing Systems (NeurIPS 2024)

Legal contracts in the custody and fund services domain govern critical aspects such as key provider responsibilities, fee schedules, and indemnification rights. However, it is challenging for an off-the-shelf Large Language Model (LLM) to ingest these contracts due to the lengthy unstructured streams of text, limited LLM context windows, and complex legal jargon. To address these challenges, we introduce LAW (Legal Agentic Workflows for Custody and Fund Services Contracts). LAW features a modular design that responds to user queries by orchestrating a suite of domain-specific tools and text agents. Our experiments demonstrate that LAW, by integrating multiple specialized agents and tools, significantly outperforms the baseline. LAW excels particularly in complex tasks such as calculating a contract's termination date, surpassing the baseline by 92.9% points. Furthermore, LAW offers a cost-effective alternative to traditional fine-tuned legal LLMs by leveraging reusable, domain-specific tools., Comment: Accepted at The 31st International Conference on Computational Linguistics (COLING 2025)

Reward models (RMs) are a crucial component in the alignment of large language models' (LLMs) outputs with human values. RMs approximate human preferences over possible LLM responses to the same prompt by predicting and comparing reward scores. However, as they are typically modified versions of LLMs with scalar output heads, RMs are large black boxes whose predictions are not explainable. More transparent RMs would enable improved trust in the alignment of LLMs. In this work, we propose to use contrastive explanations to explain any binary response comparison made by an RM. Specifically, we generate a diverse set of new comparisons similar to the original one to characterise the RM's local behaviour. The perturbed responses forming the new comparisons are generated to explicitly modify manually specified high-level evaluation attributes, on which analyses of RM behaviour are grounded. In quantitative experiments, we validate the effectiveness of our method for finding high-quality contrastive explanations. We then showcase the qualitative usefulness of our method for investigating global sensitivity of RMs to each evaluation attribute, and demonstrate how representative examples can be automatically extracted to explain and compare behaviours of different RMs. We see our method as a flexible framework for RM explanation, providing a basis for more interpretable and trustworthy LLM alignment.

State-of-the-art multimodal web agents, powered by Multimodal Large Language Models (MLLMs), can autonomously execute many web tasks by processing user instructions and interacting with graphical user interfaces (GUIs). Current strategies for building web agents rely on (i) the generalizability of underlying MLLMs and their steerability via prompting, and (ii) large-scale fine-tuning of MLLMs on web-related tasks. However, web agents still struggle to automate tasks on unseen websites and domains, limiting their applicability to enterprise-specific and proprietary platforms. Beyond generalization from large-scale pre-training and fine-tuning, we propose building agents for few-shot adaptability using human demonstrations. We introduce the AdaptAgent framework that enables both proprietary and open-weights multimodal web agents to adapt to new websites and domains using few human demonstrations (up to 2). Our experiments on two popular benchmarks -- Mind2Web & VisualWebArena -- show that using in-context demonstrations (for proprietary models) or meta-adaptation demonstrations (for meta-learned open-weights models) boosts task success rate by 3.36% to 7.21% over non-adapted state-of-the-art models, corresponding to a relative increase of 21.03% to 65.75%. Furthermore, our additional analyses (a) show the effectiveness of multimodal demonstrations over text-only ones, (b) shed light on the influence of different data selection strategies during meta-learning on the generalization of the agent, and (c) demonstrate the effect of number of few-shot examples on the web agent's success rate. Overall, our results unlock a complementary axis for developing widely applicable multimodal web agents beyond large-scale pre-training and fine-tuning, emphasizing few-shot adaptability., Comment: 18 pages, 3 figures, an abridged version to appear in NeurIPS 2024 AFM Workshop

We investigate the use of sequence analysis for behavior modeling, emphasizing that sequential context often outweighs the value of aggregate features in understanding human behavior. We discuss framing common problems in fields like healthcare, finance, and e-commerce as sequence modeling tasks, and address challenges related to constructing coherent sequences from fragmented data and disentangling complex behavior patterns. We present a framework for sequence modeling using Ensembles of Hidden Markov Models, which are lightweight, interpretable, and efficient. Our ensemble-based scoring method enables robust comparison across sequences of different lengths and enhances performance in scenarios with imbalanced or scarce data. The framework scales in real-world scenarios, is compatible with downstream feature-based modeling, and is applicable in both supervised and unsupervised learning settings. We demonstrate the effectiveness of our method with results on a longitudinal human behavior dataset.

Amidst escalating climate change, hurricanes are inflicting severe socioeconomic impacts, marked by heightened economic losses and increased displacement. Previous research utilized nighttime light data to predict the impact of hurricanes on economic losses. However, prior work did not provide a thorough analysis of the impact of combining different techniques for pre-processing nighttime light (NTL) data. Addressing this gap, our research explores a variety of NTL pre-processing techniques, including value thresholding, built masking, and quality filtering and imputation, applied to two distinct datasets, VSC-NTL and VNP46A2, at the zip code level. Experiments evaluate the correlation of the denoised NTL data with economic damages of Category 4-5 hurricanes in Florida. They reveal that the quality masking and imputation technique applied to VNP46A2 show a substantial correlation with economic damage data.

In this paper, we study the ability of large language models to learn specific mathematical rules such as distributivity or simplifying equations. We present an empirical analysis of their ability to generalize these rules, as well as to reuse them in the context of word problems. For this purpose, we provide a rigorous methodology to build synthetic data incorporating such rules, and perform fine-tuning of large language models on such data. Our experiments show that our model can learn and generalize these rules to some extent, as well as suitably reuse them in the context of word problems., Comment: NeurIPS'24 MATH-AI, the 4th Workshop on Mathematical Reasoning and AI

The rise of large language models (LLMs) for visually rich document understanding (VRDU) has kindled a need for prompt-response, document-based datasets. As annotating new datasets from scratch is labor-intensive, the existing literature has generated prompt-response datasets from available resources using simple templates. For the case of key information extraction (KIE), one of the most common VRDU tasks, past work has typically employed the template "What is the value for the {key}?". However, given the variety of questions encountered in the wild, simple and uniform templates are insufficient for creating robust models in research and industrial contexts. In this work, we present K2Q, a diverse collection of five datasets converted from KIE to a prompt-response format using a plethora of bespoke templates. The questions in K2Q can span multiple entities and be extractive or boolean. We empirically compare the performance of seven baseline generative models on K2Q with zero-shot prompting. We further compare three of these models when training on K2Q versus training on simpler templates to motivate the need of our work. We find that creating diverse and intricate KIE questions enhances the performance and robustness of VRDU models. We hope this work encourages future studies on data quality for generative model training., Comment: Accepted to EMNLP Findings 2024

Conditional demographic parity (CDP) is a measure of the demographic parity of a predictive model or decision process when conditioning on an additional feature or set of features. Many algorithmic fairness techniques exist to target demographic parity, but CDP is much harder to achieve, particularly when the conditioning variable has many levels and/or when the model outputs are continuous. The problem of auditing and enforcing CDP is understudied in the literature. In light of this, we propose novel measures of {conditional demographic disparity (CDD)} which rely on statistical distances borrowed from the optimal transport literature. We further design and evaluate regularization-based approaches based on these CDD measures. Our methods, \fairbit{} and \fairlp{}, allow us to target CDP even when the conditioning variable has many levels. When model outputs are continuous, our methods target full equality of the conditional distributions, unlike other methods that only consider first moments or related proxy quantities. We validate the efficacy of our approaches on real-world datasets.

The multiplicities of positive and negative pions, kaons and unidentified hadrons produced in deep-inelastic scattering are measured in bins of the Bjorken scaling variable $x$, the relative virtual-photon energy $y$ and the fraction of the virtual-photon energy transferred to the final-state hadron $z$. Data were obtained by the COMPASS Collaboration using a 160 GeV muon beam of both electric charges and a liquid hydrogen target. These measurements cover the kinematic domain with photon virtuality $Q^2 > 1$ (GeV/$c)^2$, $0.004 < x < 0.4$, $0.1 < y < 0.7$ and $0.2 < z < 0.85$, in accordance with the kinematic domain used in earlier published COMPASS multiplicity measurements with an isoscalar target. The calculation of radiative corrections was improved by using the Monte Carlo generator DJANGOH, which results in up to 12\% larger corrections in the low-$x$ region., Comment: 19 pages, 29 figures

Large language models (LLMs) are primarily designed to understand unstructured text. When directly applied to structured formats such as tabular data, they may struggle to discern inherent relationships and overlook critical patterns. While tabular representation learning methods can address some of these limitations, existing efforts still face challenges with sparse high-cardinality fields, precise numerical reasoning, and column-heavy tables. Furthermore, leveraging these learned representations for downstream tasks through a language based interface is not apparent. In this paper, we present an innovative and scalable solution to these challenges. Concretely, our approach introduces a multi-tier partitioning mechanism that utilizes power-law dynamics to handle large vocabularies, an adaptive quantization mechanism to impose priors on numerical continuity, and a distinct treatment of core-columns and meta-information columns. To facilitate instruction tuning on LLMs, we propose a parameter efficient decoder that interleaves transaction and text modalities using a series of adapter layers, thereby exploiting rich cross-task knowledge. We validate the efficacy of our solution on a large-scale dataset of synthetic payments transactions.

Mangroves are dynamic coastal ecosystems that are crucial to environmental health, economic stability, and climate resilience. The monitoring and preservation of mangroves are of global importance, with remote sensing technologies playing a pivotal role in these efforts. The integration of cutting-edge artificial intelligence with satellite data opens new avenues for ecological monitoring, potentially revolutionizing conservation strategies at a time when the protection of natural resources is more crucial than ever. The objective of this work is to provide a comprehensive evaluation of recent deep-learning models on the task of mangrove segmentation. We first introduce and make available a novel open-source dataset, MagSet-2, incorporating mangrove annotations from the Global Mangrove Watch and satellite images from Sentinel-2, from mangrove positions all over the world. We then benchmark three architectural groups, namely convolutional, transformer, and mamba models, using the created dataset. The experimental outcomes further validate the deep learning community's interest in the Mamba model, which surpasses other architectures in all metrics., Comment: 12 pages, accepted to the MACLEAN workshop of ECML/PKDD 2024

Recent developments in generative models have demonstrated its ability to create high-quality synthetic data. However, the pervasiveness of synthetic content online also brings forth growing concerns that it can be used for malicious purposes. To ensure the authenticity of the data, watermarking techniques have recently emerged as a promising solution due to their strong statistical guarantees. In this paper, we propose a flexible and robust watermarking mechanism for generative tabular data. Specifically, a data provider with knowledge of the downstream tasks can partition the feature space into pairs of $(key, value)$ columns. Within each pair, the data provider first uses elements in the $key$ column to generate a randomized set of ''green'' intervals, then encourages elements of the $value$ column to be in one of these ''green'' intervals. We show theoretically and empirically that the watermarked datasets (i) have negligible impact on the data quality and downstream utility, (ii) can be efficiently detected, and (iii) are robust against multiple attacks commonly observed in data science., Comment: 12 pages of main body, 2 figures, 5 tables

We present a lightweight approach to sequence classification using Ensemble Methods for Hidden Markov Models (HMMs). HMMs offer significant advantages in scenarios with imbalanced or smaller datasets due to their simplicity, interpretability, and efficiency. These models are particularly effective in domains such as finance and biology, where traditional methods struggle with high feature dimensionality and varied sequence lengths. Our ensemble-based scoring method enables the comparison of sequences of any length and improves performance on imbalanced datasets. This study focuses on the binary classification problem, particularly in scenarios with data imbalance, where the negative class is the majority (e.g., normal data) and the positive class is the minority (e.g., anomalous data), often with extreme distribution skews. We propose a novel training approach for HMM Ensembles that generalizes to multi-class problems and supports classification and anomaly detection. Our method fits class-specific groups of diverse models using random data subsets, and compares likelihoods across classes to produce composite scores, achieving high average precisions and AUCs. In addition, we compare our approach with neural network-based methods such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory networks (LSTMs), highlighting the efficiency and robustness of HMMs in data-scarce environments. Motivated by real-world use cases, our method demonstrates robust performance across various benchmarks, offering a flexible framework for diverse applications.

In this work we consider a new interpretation of fairness in decision making problems. Building upon existing fairness formulations, we focus on how to reason over fairness from a temporal perspective, taking into account the fairness of a history of past decisions. After introducing the concept of temporal fairness, we propose three approaches that incorporate temporal fairness in decision making problems formulated as optimization problems. We present a qualitative evaluation of our approach in four different domains and compare the solutions against a baseline approach that does not consider the temporal aspect of fairness., Comment: Paper accepted at ECAI 2024. This is an extended version that includes Supplementary Material

Most of the work on learning action models focus on learning the actions' dynamics from input plans. This allows us to specify the valid plans of a planning task. However, very little work focuses on learning action costs, which in turn allows us to rank the different plans. In this paper we introduce a new problem: that of learning the costs of a set of actions such that a set of input plans are optimal under the resulting planning model. To solve this problem we present $LACFIP^k$, an algorithm to learn action's costs from unlabeled input plans. We provide theoretical and empirical results showing how $LACFIP^k$ can successfully solve this task.

As vehicle systems become increasingly complex, with more features, services, sensors, actuators, and processing units, it is important to view vehicles not just as modes of transportation moving toward full autonomy, but also as adaptive systems that respond to the needs of their occupants. Vehicular services can be developed to support these adaptations. However, the increasing complexity of vehicular service development, even with current standardizations, best practices and guidelines, are insufficient to tackle the high complexity of development, with expectations of up to 1 (U.S.) billion lines of code for a fully (level 5) autonomous vehicle. Within this survey, the paradigm of Deterministic Software Defined Vehicles is explored, aiming to enhance the quality and ease of developing automotive services by focusing on service-oriented architectures, virtualization techniques, and the necessary deterministic intra- and inter-vehicular communications. Considering the main open challenges for such verticals, a vision architecture towards improved services development and orchestration is presented, focusing on: a) a deterministic network configurator; b) a data layer configurator; c) a hypervisor configurator; d) the vehicle abstraction layer; and e) a software orchestrator.

We present Im2Flow2Act, a scalable learning framework that enables robots to acquire real-world manipulation skills without the need of real-world robot training data. The key idea behind Im2Flow2Act is to use object flow as the manipulation interface, bridging domain gaps between different embodiments (i.e., human and robot) and training environments (i.e., real-world and simulated). Im2Flow2Act comprises two components: a flow generation network and a flow-conditioned policy. The flow generation network, trained on human demonstration videos, generates object flow from the initial scene image, conditioned on the task description. The flow-conditioned policy, trained on simulated robot play data, maps the generated object flow to robot actions to realize the desired object movements. By using flow as input, this policy can be directly deployed in the real world with a minimal sim-to-real gap. By leveraging real-world human videos and simulated robot play data, we bypass the challenges of teleoperating physical robots in the real world, resulting in a scalable system for diverse tasks. We demonstrate Im2Flow2Act's capabilities in a variety of real-world tasks, including the manipulation of rigid, articulated, and deformable objects., Comment: Conference on Robot Learning 2024

Adversarially robust optimization (ARO) has become the de facto standard for training models to defend against adversarial attacks during testing. However, despite their robustness, these models often suffer from severe overfitting. To mitigate this issue, several successful approaches have been proposed, including replacing the empirical distribution in training with: (i) a worst-case distribution within an ambiguity set, leading to a distributionally robust (DR) counterpart of ARO; or (ii) a mixture of the empirical distribution with one derived from an auxiliary dataset (e.g., synthetic, external, or out-of-domain). Building on the first approach, we explore the Wasserstein DR counterpart of ARO for logistic regression and show it admits a tractable convex optimization reformulation. Adopting the second approach, we enhance the DR framework by intersecting its ambiguity set with one constructed from an auxiliary dataset, which yields significant improvements when the Wasserstein distance between the data-generating and auxiliary distributions can be estimated. We analyze the resulting optimization problem, develop efficient solutions, and show that our method outperforms benchmark approaches on standard datasets., Comment: 33 pages, 3 color figures, under review at a conference

Recent advancements in language modeling have shown promising results when applied to time series data. In particular, fine-tuning pre-trained large language models (LLMs) for time series classification tasks has achieved state-of-the-art (SOTA) performance on standard benchmarks. However, these LLM-based models have a significant drawback due to the large model size, with the number of trainable parameters in the millions. In this paper, we propose an alternative approach to leveraging the success of language modeling in the time series domain. Instead of fine-tuning LLMs, we utilize a language embedding model to embed time series and then pair the embeddings with a simple classification head composed of convolutional neural networks (CNN) and multilayer perceptron (MLP). We conducted extensive experiments on well-established time series classification benchmark datasets. We demonstrated LETS-C not only outperforms the current SOTA in classification accuracy but also offers a lightweight solution, using only 14.5% of the trainable parameters on average compared to the SOTA model. Our findings suggest that leveraging language encoders to embed time series data, combined with a simple yet effective classification head, offers a promising direction for achieving high-performance time series classification while maintaining a lightweight model architecture., Comment: 22 pages, 5 figures, 10 tables

In this paper we show that if a path structure has non-vanishing curvature at apoint then it has a canonical reduction to a Z/2Z-structure at a neighbourhood of thatpoint (in many cases it has a canonical parallelism). A simple implication of this resultis that the automorphism group of a non-flat path structure is of maximal dimensionthree (a result by Tresse of 1896). We also classify the invariant path structures onthree-dimensional Lie groups.

A myriad of different Large Language Models (LLMs) face a common challenge in contextually analyzing table question-answering tasks. These challenges are engendered from (1) finite context windows for large tables, (2) multi-faceted discrepancies amongst tokenization patterns against cell boundaries, and (3) various limitations stemming from data confidentiality in the process of using external models such as gpt-3.5-turbo. We propose a cooperative game dubbed "HiddenTables" as a potential resolution to this challenge. In essence, "HiddenTables" is played between the code-generating LLM "Solver" and the "Oracle" which evaluates the ability of the LLM agents to solve Table QA tasks. This game is based on natural language schemas and importantly, ensures the security of the underlying data. We provide evidential experiments on a diverse set of tables that demonstrate an LLM's collective inability to generalize and perform on complex queries, handle compositional dependencies, and align natural language to programmatic commands when concrete table schemas are provided. Unlike encoder-based models, we have pushed the boundaries of "HiddenTables" to not be limited by the number of rows - therefore we exhibit improved efficiency in prompt and completion tokens. Our infrastructure has spawned a new dataset "PyQTax" that spans across 116,671 question-table-answer triplets and provides additional fine-grained breakdowns & labels for varying question taxonomies. Therefore, in tandem with our academic contributions regarding LLMs' deficiency in TableQA tasks, "HiddenTables" is a tactile manifestation of how LLMs can interact with massive datasets while ensuring data security and minimizing generation costs., Comment: In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP 2023)

This paper proposes Progressive Inference - a framework to compute input attributions to explain the predictions of decoder-only sequence classification models. Our work is based on the insight that the classification head of a decoder-only Transformer model can be used to make intermediate predictions by evaluating them at different points in the input sequence. Due to the causal attention mechanism, these intermediate predictions only depend on the tokens seen before the inference point, allowing us to obtain the model's prediction on a masked input sub-sequence, with negligible computational overheads. We develop two methods to provide sub-sequence level attributions using this insight. First, we propose Single Pass-Progressive Inference (SP-PI), which computes attributions by taking the difference between consecutive intermediate predictions. Second, we exploit a connection with Kernel SHAP to develop Multi Pass-Progressive Inference (MP-PI). MP-PI uses intermediate predictions from multiple masked versions of the input to compute higher quality attributions. Our studies on a diverse set of models trained on text classification tasks show that SP-PI and MP-PI provide significantly better attributions compared to prior work.