We introduce a framework for expanding residual computational graphs using jets, operators that generalize truncated Taylor series. Our method provides a systematic approach to disentangle contributions of different computational paths to model predictions. In contrast to existing techniques such as distillation, probing, or early decoding, our expansions rely solely on the model itself and requires no data, training, or sampling from the model. We demonstrate how our framework grounds and subsumes logit lens, reveals a (super-)exponential path structure in the recursive residual depth and opens up several applications. These include sketching a transformer large language model with n-gram statistics extracted from its computations, and indexing the models' levels of toxicity knowledge. Our approach enables data-free analysis of residual computation for model interpretability, development, and evaluation. The project website can be found here. Machine learning models, particularly large-scale foundation models, have become increasingly prevalent and impactful across a wide range of domains (Wei et al., 2021; Bommasani et al., 2023; Touvron et al., 2023b). While delivering strong results, their black-box nature has led to the development of techniques to assess their behavior and gain insights into their internal mechanisms. In this space, mechanistic interpretability (MI) (see e.g. Bereska & Gavves, 2024; Ferrando et al., 2024, for recent surverys) has emerged as an alternative to more classic local attribution methods such as SHAP (Lundberg, 2017) or integrated gradient (Sundararajan et al., 2017).

Retrieval-augmented generation (RAG) empowers large language models (LLMs) to utilize external knowledge sources. The increasing capacity of LLMs to process longer input sequences opens up avenues for providing more retrieved information, to potentially enhance the quality of generated outputs. It is plausible to assume that a larger retrieval set would contain more relevant information (higher recall), that might result in improved performance. However, our empirical findings demonstrate that for many long-context LLMs, the quality of generated output initially improves first, but then subsequently declines as the number of retrieved passages increases. This paper investigates this phenomenon, identifying the detrimental impact of retrieved "hard negatives" as a key contributor. To mitigate this and enhance the robustness of long-context LLM-based RAG, we propose both training-free and training-based approaches. We first showcase the effectiveness of retrieval reordering as a simple yet powerful training-free optimization. Furthermore, we explore training-based methods, specifically RAG-specific implicit LLM fine-tuning and RAG-oriented fine-tuning with intermediate reasoning, demonstrating their capacity for substantial performance gains. Finally, we conduct a systematic analysis of design choices for these training-based methods, including data distribution, retriever selection, and training context length.

Document understanding is a challenging task to process and comprehend large amounts of textual and visual information. Recent advances in Large Language Models (LLMs) have significantly improved the performance of this task. However, existing methods typically focus on either plain text or a limited number of document images, struggling to handle long PDF documents with interleaved text and images, especially in academic papers. In this paper, we introduce PDF-WuKong, a multimodal large language model (MLLM) which is designed to enhance multimodal question-answering (QA) for long PDF documents. PDF-WuKong incorporates a sparse sampler that operates on both text and image representations, significantly improving the efficiency and capability of the MLLM. The sparse sampler is integrated with the MLLM's image encoder and selects the paragraphs or diagrams most pertinent to user queries for processing by the language model. To effectively train and evaluate our model, we construct PaperPDF, a dataset consisting of a broad collection of academic papers sourced from arXiv, multiple strategies are proposed to generate automatically 1M QA pairs along with their corresponding evidence sources. Experimental results demonstrate the superiority and high efficiency of our approach over other models on the task of long multimodal PDF understanding, surpassing proprietary products by an average of 8.6% on F1. Our code and dataset will be released at https://github.com/yh-hust/PDF-Wukong.

A vast amount of textual data is added to the internet daily, making utilization and interpretation of such data difficult and cumbersome. As a result, automatic text summarization is crucial for extracting relevant information, saving precious reading time. Although many transformer-based models excel in summarization, they are constrained by their input size, preventing them from processing texts longer than their context size. This study introduces three novel algorithms that allow any LLM to efficiently overcome its input size limitation, effectively utilizing its full potential without any architectural modifications. We test our algorithms on texts with more than 70,000 words, and our experiments show a significant increase in BERTScore with competitive ROUGE scores.

Vast amounts of heterogeneous knowledge are becoming publicly available in the form of knowledge graphs, often linking multiple sources of data that have never been together before, and thereby enabling scholars to answer many new research questions. It is often not known beforehand, however, which questions the data might have the answers to, potentially leaving many interesting and novel insights to remain undiscovered. To support scholars during this scientific workflow, we introduce an anytime algorithm for the bottom-up discovery of generalized multimodal graph patterns in knowledge graphs. Each pattern is a conjunction of binary statements with (data-) type variables, constants, and/or value patterns. Upon discovery, the patterns are converted to SPARQL queries and presented in an interactive facet browser together with metadata and provenance information, enabling scholars to explore, analyse, and share queries. We evaluate our method from a user perspective, with the help of domain experts in the humanities.

In psychotherapy, therapeutic outcome assessment, or treatment outcome evaluation, is essential for enhancing mental health care by systematically evaluating therapeutic processes and outcomes. Existing large language model approaches often focus on therapist-centered, single-session evaluations, neglecting the client's subjective experience and longitudinal progress across multiple sessions. To address these limitations, we propose IPAEval, a client-Informed Psychological Assessment-based Evaluation framework that automates treatment outcome evaluations from the client's perspective using clinical interviews. IPAEval integrates cross-session client-contextual assessment and session-focused client-dynamics assessment to provide a comprehensive understanding of therapeutic progress. Experiments on our newly developed TheraPhase dataset demonstrate that IPAEval effectively tracks symptom severity and treatment outcomes over multiple sessions, outperforming previous single-session models and validating the benefits of items-aware reasoning mechanisms. In psychotherapy, therapeutic outcome assessment, a.k.a treatment outcome evaluation under clinical settings, refers to the systematic evaluation of therapeutic processes and outcomes (Groth-Marnat, 2009), focusing on factors such as therapist effectiveness (Johns et al., 2019) and treatment efficacy (Jensen-Doss et al., 2018) to improve mental health care delivery. It plays a significant role in enhancing the quality and effectiveness of mental health care by providing actionable insights that guide therapists in refining their treatment approaches (Wampold & Imel, 2015), ultimately leading to better client outcomes and improved therapeutic relationships in real-world clinical practice (Maruish & Leahy, 2000). Over the last couple of years, the emergence of large language models has demonstrated their effectiveness in automatic evaluations, showing a high degree of alignment with human judgment when provided with proper instruction and contextual guidance (Liu et al., 2023; Li et al., 2024b; Kim et al., 2024). This aligns with the "LLMs-as-a-judge" paradigm, where LLMs are employed to simulate human evaluators by providing assessments based on natural language input (Zheng et al., 2023; Wang et al., 2024b). This paradigm has been extended to therapeutic outcome assessment by harnessing LLMs' ability to model complex therapeutic procedures and interactions, offering a novel pathway for automating the assessment of therapeutic efficacy (Chiu et al., 2024; Lee et al., 2024; Li et al., 2024a).

Unlike regular models, these very simple models--often referred to as toy models--are not required to be linked to the real world through structural similarity or resemblance relations. They are not meant to be approximations of the target world system, and in some cases, they are not even required to be representational. In semantic terms, they do not accurately map onto their targets. Despite these limitations, they are still useful in understanding theoretical concepts and possible configurations of the target system. Paradigmatic examples of toy models include Boyle's law and the Ising model in physics, the Lotka-Volterra model in population ecology, and the Schelling model in the social sciences (Weisberg, 2013). In recent years, philosophers of science have become interested in toy models (Grüne-Yanoff, 2009; Luczak, 2017; Reutlinger et al., 2018; Frigg & Nguyen, 2017; Nguyen, 2020). The main purpose of this literature is to explore the nature of these models and examine how they perform their epistemic function. Despite lacking the regular descriptive and predictive features of full-scale scientific models, they often offer an elementary understanding of a phenomenon. Their definitions of "toy model" differ as well as their assessment of the importance of representation in modelling generally, but they all agree that toy models play an important epistemic role in scientific research, exploration, and pedagogy.

Explaining Artificial Intelligence (AI) decisions is a major challenge nowadays in AI, in particular when applied to sensitive scenarios like medicine and law. However, the need to explain the rationale behind decisions is a main issue also for human-based deliberation as it is important to justify \textit{why} a certain decision has been taken. Resident medical doctors for instance are required not only to provide a (possibly correct) diagnosis, but also to explain how they reached a certain conclusion. Developing new tools to aid residents to train their explanation skills is therefore a central objective of AI in education. In this paper, we follow this direction, and we present, to the best of our knowledge, the first multilingual dataset for Medical Question Answering where correct and incorrect diagnoses for a clinical case are enriched with a natural language explanation written by doctors. These explanations have been manually annotated with argument components (i.e., premise, claim) and argument relations (i.e., attack, support), resulting in the Multilingual CasiMedicos-Arg dataset which consists of 558 clinical cases in four languages (English, Spanish, French, Italian) with explanations, where we annotated 5021 claims, 2313 premises, 2431 support relations, and 1106 attack relations. We conclude by showing how competitive baselines perform over this challenging dataset for the argument mining task.

Preference Optimization (PO) has proven an effective step for aligning language models to human-desired behaviors. Current variants, following the offline Direct Preference Optimization objective, have focused on a strict setting where all tokens are contributing signals of KL divergence and rewards to the loss function. However, human preference is not affected by each word in a sequence equally but is often dependent on specific words or phrases, e.g. existence of toxic terms leads to non-preferred responses. Based on this observation, we argue that not all tokens should be weighted equally during PO and propose a flexible objective termed SparsePO, that aims to automatically learn to weight the KL divergence and reward corresponding to each token during PO training. We propose two different variants of weight-masks that can either be derived from the reference model itself or learned on the fly. Notably, our method induces sparsity in the learned masks, allowing the model to learn how to best weight reward and KL divergence contributions at the token level, learning an optimal level of mask sparsity. Extensive experiments on multiple domains, including sentiment control, dialogue, text summarization and text-to-code generation, illustrate that our approach assigns meaningful weights to tokens according to the target task, generates more responses with the desired preference and improves reasoning tasks by up to 2 percentage points compared to other token- and response-level PO methods.

Cognitive biases are systematic deviations in thinking that lead to irrational judgments and problematic decision-making, extensively studied across various fields. Recently, large language models (LLMs) have shown advanced understanding capabilities but may inherit human biases from their training data. While social biases in LLMs have been well-studied, cognitive biases have received less attention, with existing research focusing on specific scenarios. The broader impact of cognitive biases on LLMs in various decision-making contexts remains underexplored. We investigated whether LLMs are influenced by the threshold priming effect in relevance judgments, a core task and widely-discussed research topic in the Information Retrieval (IR) coummunity. The priming effect occurs when exposure to certain stimuli unconsciously affects subsequent behavior and decisions. Our experiment employed 10 topics from the TREC 2019 Deep Learning passage track collection, and tested AI judgments under different document relevance scores, batch lengths, and LLM models, including GPT-3.5, GPT-4, LLaMa2-13B and LLaMa2-70B. Results showed that LLMs tend to give lower scores to later documents if earlier ones have high relevance, and vice versa, regardless of the combination and model used. Our finding demonstrates that LLM%u2019s judgments, similar to human judgments, are also influenced by threshold priming biases, and suggests that researchers and system engineers should take into account potential human-like cognitive biases in designing, evaluating, and auditing LLMs in IR tasks and beyond.