Mrinmaya Sachan

Zhijing Jin, Sydney Levine, Fernando Gonzalez et al. · allen-ai, cmu

AI systems are becoming increasingly intertwined with human life. In order to effectively collaborate with humans and ensure safety, AI systems need to be able to understand, interpret and predict human moral judgments and decisions. Human moral judgments are often guided by rules, but not always. A central challenge for AI safety is capturing the flexibility of the human moral mind -- the ability to determine when a rule should be broken, especially in novel or unusual situations. In this paper, we present a novel challenge set consisting of rule-breaking question answering (RBQA) of cases that involve potentially permissible rule-breaking -- inspired by recent moral psychology studies. Using a state-of-the-art large language model (LLM) as a basis, we propose a novel moral chain of thought (MORALCOT) prompting strategy that combines the strengths of LLMs with theories of moral reasoning developed in cognitive science to predict human moral judgments. MORALCOT outperforms seven existing LLMs by 6.2% F1, suggesting that modeling human reasoning might be necessary to capture the flexibility of the human moral mind. We also conduct a detailed error analysis to suggest directions for future work to improve AI safety using RBQA. Our data is open-sourced at https://huggingface.co/datasets/feradauto/MoralExceptQA and code at https://github.com/feradauto/MoralCoT

Jingwei Ni, Zhijing Jin, Markus Freitag et al. · deepmind, eth-zurich

Human-translated text displays distinct features from naturally written text in the same language. This phenomena, known as translationese, has been argued to confound the machine translation (MT) evaluation. Yet, we find that existing work on translationese neglects some important factors and the conclusions are mostly correlational but not causal. In this work, we collect CausalMT, a dataset where the MT training data are also labeled with the human translation directions. We inspect two critical factors, the train-test direction match (whether the human translation directions in the training and test sets are aligned), and data-model direction match (whether the model learns in the same direction as the human translation direction in the dataset). We show that these two factors have a large causal effect on the MT performance, in addition to the test-model direction mismatch highlighted by existing work on the impact of translationese. In light of our findings, we provide a set of suggestions for MT training and evaluation. Our code and data are at https://github.com/EdisonNi-hku/CausalMT

Zhijing Jin, Jiarui Liu, Zhiheng Lyu et al.

Causal inference is one of the hallmarks of human intelligence. While the field of CausalNLP has attracted much interest in the recent years, existing causal inference datasets in NLP primarily rely on discovering causality from empirical knowledge (e.g., commonsense knowledge). In this work, we propose the first benchmark dataset to test the pure causal inference skills of large language models (LLMs). Specifically, we formulate a novel task Corr2Cause, which takes a set of correlational statements and determines the causal relationship between the variables. We curate a large-scale dataset of more than 200K samples, on which we evaluate seventeen existing LLMs. Through our experiments, we identify a key shortcoming of LLMs in terms of their causal inference skills, and show that these models achieve almost close to random performance on the task. This shortcoming is somewhat mitigated when we try to re-purpose LLMs for this skill via finetuning, but we find that these models still fail to generalize -- they can only perform causal inference in in-distribution settings when variable names and textual expressions used in the queries are similar to those in the training set, but fail in out-of-distribution settings generated by perturbing these queries. Corr2Cause is a challenging task for LLMs, and would be helpful in guiding future research on improving LLMs' pure reasoning skills and generalizability. Our data is at https://huggingface.co/datasets/causalnlp/corr2cause. Our code is at https://github.com/causalNLP/corr2cause.

Kumar Shridhar, Alessandro Stolfo, Mrinmaya Sachan

Step-by-step reasoning approaches like chain of thought (CoT) have proved to be very effective in inducing reasoning capabilities in large language models. However, the success of the CoT approach is fundamentally tied to the model size, and billion parameter-scale models are often needed to get CoT to work. In this paper, we propose a knowledge distillation approach that leverages the step-by-step CoT reasoning capabilities of larger models and distills these abilities into smaller models. In this work, we propose an alternative reasoning scheme, Socratic CoT, that learns a decomposition of the original problem into a sequence of subproblems and uses it to guide the intermediate reasoning steps. We use Socratic CoT to train a combination of two small distilled models: a problem decomposer and a subproblem solver. In practice, given a new problem, the two distilled models work in sync to decompose and solve complex problems. On multiple reasoning datasets (GSM8K, StrategyQA, and SVAMP), our proposed distillation strategies boosts the performance of smaller models over 70% compared to the baselines. Finally, we investigate when Socratic CoT is an effective alternative to CoT, demonstrating cases where a much smaller model (GPT-2 large) can outperform a 10X larger model (GPT-3 6B). Our code is available here: https://github.com/kumar-shridhar/Distiiling-LM

Jakub Macina, Nico Daheim, Lingzhi Wang et al. · cmu, eth-zurich

Designing dialog tutors has been challenging as it involves modeling the diverse and complex pedagogical strategies employed by human tutors. Although there have been significant recent advances in neural conversational systems using large language models (LLMs) and growth in available dialog corpora, dialog tutoring has largely remained unaffected by these advances. In this paper, we rigorously analyze various generative language models on two dialog tutoring datasets for language learning using automatic and human evaluations to understand the new opportunities brought by these advances as well as the challenges we must overcome to build models that would be usable in real educational settings. We find that although current approaches can model tutoring in constrained learning scenarios when the number of concepts to be taught and possible teacher strategies are small, they perform poorly in less constrained scenarios. Our human quality evaluation shows that both models and ground-truth annotations exhibit low performance in terms of equitable tutoring, which measures learning opportunities for students and how engaging the dialog is. To understand the behavior of our models in a real tutoring setting, we conduct a user study using expert annotators and find a significantly large number of model reasoning errors in 45% of conversations. Finally, we connect our findings to outline future work.

Wangchunshu Zhou, Yuchen Eleanor Jiang, Long Li et al.

Recent advances on large language models (LLMs) enable researchers and developers to build autonomous language agents that can automatically solve various tasks and interact with environments, humans, and other agents using natural language interfaces. We consider language agents as a promising direction towards artificial general intelligence and release Agents, an open-source library with the goal of opening up these advances to a wider non-specialist audience. Agents is carefully engineered to support important features including planning, memory, tool usage, multi-agent communication, and fine-grained symbolic control. Agents is user-friendly as it enables non-specialists to build, customize, test, tune, and deploy state-of-the-art autonomous language agents without much coding. The library is also research-friendly as its modularized design makes it easily extensible for researchers. Agents is available at https://github.com/aiwaves-cn/agents.

David F. Jenny, Yann Billeter, Mrinmaya Sachan et al. · eth-zurich

The rapid advancement of Large Language Models (LLMs) has sparked intense debate regarding the prevalence of bias in these models and its mitigation. Yet, as exemplified by both results on debiasing methods in the literature and reports of alignment-related defects from the wider community, bias remains a poorly understood topic despite its practical relevance. To enhance the understanding of the internal causes of bias, we analyse LLM bias through the lens of causal fairness analysis, which enables us to both comprehend the origins of bias and reason about its downstream consequences and mitigation. To operationalize this framework, we propose a prompt-based method for the extraction of confounding and mediating attributes which contribute to the LLM decision process. By applying Activity Dependency Networks (ADNs), we then analyse how these attributes influence an LLM's decision process. We apply our method to LLM ratings of argument quality in political debates. We find that the observed disparate treatment can at least in part be attributed to confounding and mitigating attributes and model misalignment, and discuss the consequences of our findings for human-AI alignment and bias mitigation. Our code and data are at https://github.com/david-jenny/LLM-Political-Study.

Kumar Shridhar, Nicholas Monath, Raghuveer Thirukovalluru et al.

Ontonotes has served as the most important benchmark for coreference resolution. However, for ease of annotation, several long documents in Ontonotes were split into smaller parts. In this work, we build a corpus of coreference-annotated documents of significantly longer length than what is currently available. We do so by providing an accurate, manually-curated, merging of annotations from documents that were split into multiple parts in the original Ontonotes annotation process. The resulting corpus, which we call LongtoNotes contains documents in multiple genres of the English language with varying lengths, the longest of which are up to 8x the length of documents in Ontonotes, and 2x those in Litbank. We evaluate state-of-the-art neural coreference systems on this new corpus, analyze the relationships between model architectures/hyperparameters and document length on performance and efficiency of the models, and demonstrate areas of improvement in long-document coreference modeling revealed by our new corpus. Our data and code is available at: https://github.com/kumar-shridhar/LongtoNotes.

Kumar Shridhar, Jakub Macina, Mennatallah El-Assady et al. · cmu, eth-zurich

Socratic questioning is an educational method that allows students to discover answers to complex problems by asking them a series of thoughtful questions. Generation of didactically sound questions is challenging, requiring understanding of the reasoning process involved in the problem. We hypothesize that such questioning strategy can not only enhance the human performance, but also assist the math word problem (MWP) solvers. In this work, we explore the ability of large language models (LMs) in generating sequential questions for guiding math word problem-solving. We propose various guided question generation schemes based on input conditioning and reinforcement learning. On both automatic and human quality evaluations, we find that LMs constrained with desirable question properties generate superior questions and improve the overall performance of a math word problem solver. We conduct a preliminary user study to examine the potential value of such question generation models in the education domain. Results suggest that the difficulty level of problems plays an important role in determining whether questioning improves or hinders human performance. We discuss the future of using such questioning strategies in education.

Zhijing Jin, Max Kleiman-Weiner, Giorgio Piatti et al.

We evaluate the moral alignment of LLMs with human preferences in multilingual trolley problems. Building on the Moral Machine experiment, which captures over 40 million human judgments across 200+ countries, we develop a cross-lingual corpus of moral dilemma vignettes in over 100 languages called MultiTP. This dataset enables the assessment of LLMs' decision-making processes in diverse linguistic contexts. Our analysis explores the alignment of 19 different LLMs with human judgments, capturing preferences across six moral dimensions: species, gender, fitness, status, age, and the number of lives involved. By correlating these preferences with the demographic distribution of language speakers and examining the consistency of LLM responses to various prompt paraphrasings, our findings provide insights into cross-lingual and ethical biases of LLMs and their intersection. We discover significant variance in alignment across languages, challenging the assumption of uniform moral reasoning in AI systems and highlighting the importance of incorporating diverse perspectives in AI ethics. The results underscore the need for further research on the integration of multilingual dimensions in responsible AI research to ensure fair and equitable AI interactions worldwide. Our code and data are at https://github.com/causalNLP/moralmachine

Kumar Shridhar, Koustuv Sinha, Andrew Cohen et al. · berkeley, meta-ai

In recent years, Large Language Models (LLMs) have demonstrated remarkable generative abilities, but can they judge the quality of their own generations? A popular concept, referred to as self-refinement, postulates that LLMs can detect and correct the errors in their generations when asked to do so. However, recent empirical evidence points in the opposite direction, suggesting that LLMs often struggle to accurately identify errors when reasoning is involved. To address this, we propose a reasoning with refinement objective called ART: Ask, Refine, and Trust, which asks necessary questions to decide when an LLM should refine its output, and either affirm or withhold trust in its refinement by ranking the refinement and the initial prediction. On two multistep reasoning tasks of mathematical word problems (GSM8K) and question answering (StrategyQA), ART achieves a performance gain of +5 points over self-refinement baselines, while using a much smaller model as the decision maker. We also demonstrate the benefit of using smaller models to make refinement decisions as a cost-effective alternative to fine-tuning a larger model.

Wangchunshu Zhou, Yuchen Eleanor Jiang, Ethan Wilcox et al. · harvard

Large language models generate fluent texts and can follow natural language instructions to solve a wide range of tasks without task-specific training. Nevertheless, it is notoriously difficult to control their generation to satisfy the various constraints required by different applications. In this work, we present InstructCTG, a controlled text generation framework that incorporates different constraints by conditioning on natural language descriptions and demonstrations of the constraints. In particular, we first extract the underlying constraints of natural texts through a combination of off-the-shelf NLP tools and simple heuristics. We then verbalize the constraints into natural language instructions to form weakly supervised training data. By prepending natural language descriptions of the constraints and a few demonstrations, we fine-tune a pre-trained language model to incorporate various types of constraints. Compared to existing search-based or score-based methods, InstructCTG is more flexible to different constraint types and has a much smaller impact on the generation quality and speed because it does not modify the decoding procedure. Additionally, InstructCTG allows the model to adapt to new constraints without re-training through the use of few-shot task generalization and in-context learning abilities of instruction-tuned language models.

Vilém Zouhar, Shehzaad Dhuliawala, Wangchunshu Zhou et al. · eth-zurich, microsoft-research

Machine translation quality estimation (QE) predicts human judgements of a translation hypothesis without seeing the reference. State-of-the-art QE systems based on pretrained language models have been achieving remarkable correlations with human judgements yet they are computationally heavy and require human annotations, which are slow and expensive to create. To address these limitations, we define the problem of metric estimation (ME) where one predicts the automated metric scores also without the reference. We show that even without access to the reference, our model can estimate automated metrics ($ρ$=60% for BLEU, $ρ$=51% for other metrics) at the sentence-level. Because automated metrics correlate with human judgements, we can leverage the ME task for pre-training a QE model. For the QE task, we find that pre-training on TER is better ($ρ$=23%) than training for scratch ($ρ$=20%).

Yifan Hou, Jiaoda Li, Yu Fei et al. · eth-zurich

Recent work has shown that language models (LMs) have strong multi-step (i.e., procedural) reasoning capabilities. However, it is unclear whether LMs perform these tasks by cheating with answers memorized from pretraining corpus, or, via a multi-step reasoning mechanism. In this paper, we try to answer this question by exploring a mechanistic interpretation of LMs for multi-step reasoning tasks. Concretely, we hypothesize that the LM implicitly embeds a reasoning tree resembling the correct reasoning process within it. We test this hypothesis by introducing a new probing approach (called MechanisticProbe) that recovers the reasoning tree from the model's attention patterns. We use our probe to analyze two LMs: GPT-2 on a synthetic task (k-th smallest element), and LLaMA on two simple language-based reasoning tasks (ProofWriter & AI2 Reasoning Challenge). We show that MechanisticProbe is able to detect the information of the reasoning tree from the model's attentions for most examples, suggesting that the LM indeed is going through a process of multi-step reasoning within its architecture in many cases.

Vilém Zouhar, Clara Meister, Juan Luis Gastaldi et al. · eth-zurich

Subword tokenization is a key part of many NLP pipelines. However, little is known about why some tokenizer and hyperparameter combinations lead to better downstream model performance than others. We propose that good tokenizers lead to \emph{efficient} channel usage, where the channel is the means by which some input is conveyed to the model and efficiency can be quantified in information-theoretic terms as the ratio of the Shannon entropy to the maximum possible entropy of the token distribution. Yet, an optimal encoding according to Shannon entropy assigns extremely long codes to low-frequency tokens and very short codes to high-frequency tokens. Defining efficiency in terms of Rényi entropy, on the other hand, penalizes distributions with either very high or very low-frequency tokens. In machine translation, we find that across multiple tokenizers, the Rényi entropy with $α= 2.5$ has a very strong correlation with \textsc{Bleu}: $0.78$ in comparison to just $-0.32$ for compressed length.

Vilém Zouhar, Kalvin Chang, Chenxuan Cui et al. · cmu, eth-zurich

Mapping words into a fixed-dimensional vector space is the backbone of modern NLP. While most word embedding methods successfully encode semantic information, they overlook phonetic information that is crucial for many tasks. We develop three methods that use articulatory features to build phonetically informed word embeddings. To address the inconsistent evaluation of existing phonetic word embedding methods, we also contribute a task suite to fairly evaluate past, current, and future methods. We evaluate both (1) intrinsic aspects of phonetic word embeddings, such as word retrieval and correlation with sound similarity, and (2) extrinsic performance on tasks such as rhyme and cognate detection and sound analogies. We hope our task suite will promote reproducibility and inspire future phonetic embedding research.

Vilém Zouhar, Clara Meister, Juan Luis Gastaldi et al. · eth-zurich

Byte-Pair Encoding (BPE) is a popular algorithm used for tokenizing data in NLP, despite being devised initially as a compression method. BPE appears to be a greedy algorithm at face value, but the underlying optimization problem that BPE seeks to solve has not yet been laid down. We formalize BPE as a combinatorial optimization problem. Via submodular functions, we prove that the iterative greedy version is a $\frac{1}{σ(\boldsymbolμ^\star)}(1-e^{-{σ(\boldsymbolμ^\star)}})$-approximation of an optimal merge sequence, where ${σ(\boldsymbolμ^\star)}$ is the total backward curvature with respect to the optimal merge sequence $\boldsymbolμ^\star$. Empirically the lower bound of the approximation is $\approx 0.37$. We provide a faster implementation of BPE which improves the runtime complexity from $\mathcal{O}\left(N M\right)$ to $\mathcal{O}\left(N \log M\right)$, where $N$ is the sequence length and $M$ is the merge count. Finally, we optimize the brute-force algorithm for optimal BPE using memoization.

Alessandro Stolfo, Zhijing Jin, Kumar Shridhar et al.

We have recently witnessed a number of impressive results on hard mathematical reasoning problems with language models. At the same time, the robustness of these models has also been called into question; recent works have shown that models can rely on shallow patterns in the problem description when generating a solution. Building on the idea of behavioral testing, we propose a novel framework, which pins down the causal effect of various factors in the input, e.g., the surface form of the problem text, the operands, and math operators on the output solution. By grounding the behavioral analysis in a causal graph describing an intuitive reasoning process, we study the behavior of language models in terms of robustness and sensitivity to direct interventions in the input space. We apply our framework on a test bed of math word problems. Our analysis shows that robustness does not appear to continuously improve as a function of size, but the GPT-3 Davinci models (175B) achieve a dramatic improvement in both robustness and sensitivity compared to all other GPT variants.

Furui Cheng, Vilém Zouhar, Simran Arora et al. · eth-zurich

Large Language Models (LLMs) are notorious for blending fact with fiction and generating non-factual content, known as hallucinations. To address this challenge, we propose an interactive system that helps users gain insight into the reliability of the generated text. Our approach is based on the idea that the self-consistency of multiple samples generated by the same LLM relates to its confidence in individual claims in the generated texts. Using this idea, we design RELIC, an interactive system that enables users to investigate and verify semantic-level variations in multiple long-form responses. This allows users to recognize potentially inaccurate information in the generated text and make necessary corrections. From a user study with ten participants, we demonstrate that our approach helps users better verify the reliability of the generated text. We further summarize the design implications and lessons learned from this research for future studies of reliable human-LLM interactions.

Nico Daheim, Jakub Macina, Manu Kapur et al. · eth-zurich

Large language models (LLMs) present an opportunity to scale high-quality personalized education to all. A promising approach towards this means is to build dialog tutoring models that scaffold students' problem-solving. However, even though existing LLMs perform well in solving reasoning questions, they struggle to precisely detect student's errors and tailor their feedback to these errors. Inspired by real-world teaching practice where teachers identify student errors and customize their response based on them, we focus on verifying student solutions and show how grounding to such verification improves the overall quality of tutor response generation. We collect a dataset of 1K stepwise math reasoning chains with the first error step annotated by teachers. We show empirically that finding the mistake in a student solution is challenging for current models. We propose and evaluate several verifiers for detecting these errors. Using both automatic and human evaluation we show that the student solution verifiers steer the generation model towards highly targeted responses to student errors which are more often correct with less hallucinations compared to existing baselines.

Shehzaad Dhuliawala, Vilém Zouhar, Mennatallah El-Assady et al. · eth-zurich

In a human-AI collaboration, users build a mental model of the AI system based on its reliability and how it presents its decision, e.g. its presentation of system confidence and an explanation of the output. Modern NLP systems are often uncalibrated, resulting in confidently incorrect predictions that undermine user trust. In order to build trustworthy AI, we must understand how user trust is developed and how it can be regained after potential trust-eroding events. We study the evolution of user trust in response to these trust-eroding events using a betting game. We find that even a few incorrect instances with inaccurate confidence estimates damage user trust and performance, with very slow recovery. We also show that this degradation in trust reduces the success of human-AI collaboration and that different types of miscalibration -- unconfidently correct and confidently incorrect -- have different negative effects on user trust. Our findings highlight the importance of calibration in user-facing AI applications and shed light on what aspects help users decide whether to trust the AI system.

Yu Fei, Ping Nie, Zhao Meng et al. · eth-zurich

Recent work has demonstrated that pre-trained language models (PLMs) are zero-shot learners. However, most existing zero-shot methods involve heavy human engineering or complicated self-training pipelines, hindering their application to new situations. In this work, we show that zero-shot text classification can be improved simply by clustering texts in the embedding spaces of PLMs. Specifically, we fit the unlabeled texts with a Bayesian Gaussian Mixture Model after initializing cluster positions and shapes using class names. Despite its simplicity, this approach achieves superior or comparable performance on both topic and sentiment classification datasets and outperforms prior works significantly on unbalanced datasets. We further explore the applicability of our clustering approach by evaluating it on 14 datasets with more diverse topics, text lengths, and numbers of classes. Our approach achieves an average of 20% absolute improvement over prompt-based zero-shot learning. Finally, we compare different PLM embedding spaces and find that texts are well-clustered by topics even if the PLM is not explicitly pre-trained to generate meaningful sentence embeddings. This work indicates that PLM embeddings can categorize texts without task-specific fine-tuning, thus providing a new way to analyze and utilize their knowledge and zero-shot learning ability.

Yifan Hou, Buse Giledereli, Yilei Tu et al. · eth-zurich

Visual language is a system of communication that conveys information through symbols, shapes, and spatial arrangements. Diagrams are a typical example of a visual language depicting complex concepts and their relationships in the form of an image. The symbolic nature of diagrams presents significant challenges for building models capable of understanding them. Recent studies suggest that Large Vision-Language Models (LVLMs) can even tackle complex reasoning tasks involving diagrams. In this paper, we investigate this phenomenon by developing a comprehensive test suite to evaluate the diagram comprehension capability of LVLMs. Our test suite uses a variety of questions focused on concept entities and their relationships over a set of synthetic as well as real diagrams across domains to evaluate the recognition and reasoning abilities of models. Our evaluation of LVLMs shows that while they can accurately identify and reason about entities, their ability to understand relationships is notably limited. Further testing reveals that the decent performance on diagram understanding largely stems from leveraging their background knowledge as shortcuts to identify and reason about the relational information. Thus, we conclude that LVLMs have a limited capability for genuine diagram understanding, and their impressive performance in diagram reasoning is an illusion emanating from other confounding factors, such as the background knowledge in the models.

Yifan Hou, Wenxiang Jiao, Meizhen Liu et al. · eth-zurich

Large language models appear to learn facts from the large text corpora they are trained on. Such facts are encoded implicitly within their many parameters, making it difficult to verify or manipulate what knowledge has been learned. Language models have recently been extended to multilingual language models (MLLMs), enabling knowledge to be learned across hundreds of languages. Meanwhile, knowledge graphs contain facts in an explicit triple format, which require careful and costly curation and are only available in a few high-resource languages, restricting their research and application. To address these issues, we propose to enhance MLLMs with knowledge from multilingual knowledge graphs (MLKGs) so as to tackle language and knowledge graph tasks across many languages, including low-resource ones. Specifically, we introduce a lightweight adapter set to enhance MLLMs with cross-lingual entity alignment and facts from MLKGs for many languages. Experiments on common benchmarks show that such enhancement benefits both MLLMs and MLKGs, achieving: (1) comparable or improved performance for knowledge graph completion and entity alignment relative to baselines, especially for low-resource languages (for which knowledge graphs are unavailable); and (2) improved MLLM performance on language understanding tasks that require multilingual factual knowledge; all while maintaining performance on other general language tasks.

Andreas Opedal, Niklas Stoehr, Abulhair Saparov et al. · eth-zurich

Solving math story problems is a complex task for students and NLP models alike, requiring them to understand the world as described in the story and reason over it to compute an answer. Recent years have seen impressive performance on automatically solving these problems with large pre-trained language models and innovative techniques to prompt them. However, it remains unclear if these models possess accurate representations of mathematical concepts. This leads to lack of interpretability and trustworthiness which impedes their usefulness in various applications. In this paper, we consolidate previous work on categorizing and representing math story problems and develop MathWorld, which is a graph-based semantic formalism specific for the domain of math story problems. With MathWorld, we can assign world models to math story problems which represent the situations and actions introduced in the text and their mathematical relationships. We combine math story problems from several existing datasets and annotate a corpus of 1,019 problems and 3,204 logical forms with MathWorld. Using this data, we demonstrate the following use cases of MathWorld: (1) prompting language models with synthetically generated question-answer pairs to probe their reasoning and world modeling abilities, and (2) generating new problems by using the world models as a design space.

Gentiana Rashiti, Geethan Karunaratne, Mrinmaya Sachan et al.

The retrieval augmented generation (RAG) system such as Retro has been shown to improve language modeling capabilities and reduce toxicity and hallucinations by retrieving from a database of non-parametric memory containing trillions of entries. We introduce Retro-li that shows retrieval can also help using a small-scale database, but it demands more accurate and better neighbors when searching in a smaller hence sparser non-parametric memory. This can be met by using a proper semantic similarity search. We further propose adding a regularization to the non-parametric memory for the first time: it significantly reduces perplexity when the neighbor search operations are noisy during inference, and it improves generalization when a domain shift occurs. We also show that Retro-li's non-parametric memory can potentially be implemented on analog in-memory computing hardware, exhibiting O(1) search time while causing noise in retrieving neighbors, with minimal (<1%) performance loss. Our code is available at: https://github.com/IBM/Retrieval-Enhanced-Transformer-Little.

Ishan Kumar, Zhijing Jin, Ehsan Mokhtarian et al.

Citation count of a paper is a commonly used proxy for evaluating the significance of a paper in the scientific community. Yet citation measures are widely criticized for failing to accurately reflect the true impact of a paper. Thus, we propose CausalCite, a new way to measure the significance of a paper by assessing the causal impact of the paper on its follow-up papers. CausalCite is based on a novel causal inference method, TextMatch, which adapts the traditional matching framework to high-dimensional text embeddings. TextMatch encodes each paper using text embeddings from large language models (LLMs), extracts similar samples by cosine similarity, and synthesizes a counterfactual sample as the weighted average of similar papers according to their similarity values. We demonstrate the effectiveness of CausalCite on various criteria, such as high correlation with paper impact as reported by scientific experts on a previous dataset of 1K papers, (test-of-time) awards for past papers, and its stability across various subfields of AI. We also provide a set of findings that can serve as suggested ways for future researchers to use our metric for a better understanding of the quality of a paper. Our code is available at https://github.com/causalNLP/causal-cite.

Nico Daheim, Nouha Dziri, Mrinmaya Sachan et al.

Ideally, dialogue systems should generate responses that are faithful to the knowledge contained in relevant documents. However, many models generate hallucinated responses instead that contradict it or contain unverifiable information. To mitigate such undesirable behaviour, it has been proposed to fine-tune a `negative expert' on negative examples and subtract its parameters from those of a pre-trained model. However, intuitively, this does not take into account that some parameters are more responsible than others in causing hallucinations. Thus, we propose to weigh their individual importance via (an approximation of) the Fisher Information matrix, which measures the uncertainty of their estimate. We call this method Elastic Weight Removal (EWR). We evaluate our method -- using different variants of Flan-T5 as a backbone language model -- on multiple datasets for information-seeking dialogue generation and compare our method with state-of-the-art techniques for faithfulness, such as CTRL, Quark, DExperts, and Noisy Channel reranking. Extensive automatic and human evaluation shows that EWR systematically increases faithfulness at minor costs in terms of other metrics. However, we notice that only discouraging hallucinations may increase extractiveness, i.e. shallow copy-pasting of document spans, which can be undesirable. Hence, as a second main contribution, we show that our method can be extended to simultaneously discourage hallucinations and extractive responses. We publicly release the code for reproducing EWR and all baselines.

Janvijay Singh, Vilém Zouhar, Mrinmaya Sachan · eth-zurich

Textbooks are one of the main mediums for delivering high-quality education to students. In particular, explanatory and illustrative visuals play a key role in retention, comprehension and general transfer of knowledge. However, many textbooks lack these interesting visuals to support student learning. In this paper, we investigate the effectiveness of vision-language models to automatically enhance textbooks with images from the web. We collect a dataset of e-textbooks in the math, science, social science and business domains. We then set up a text-image matching task that involves retrieving and appropriately assigning web images to textbooks, which we frame as a matching optimization problem. Through a crowd-sourced evaluation, we verify that (1) while the original textbook images are rated higher, automatically assigned ones are not far behind, and (2) the precise formulation of the optimization problem matters. We release the dataset of textbooks with an associated image bank to inspire further research in this intersectional area of computer vision and NLP for education.

Mubashara Akhtar, Anka Reuel, Prajna Soni et al. · meta-ai

Artificial Intelligence (AI) benchmarks play a central role in measuring progress in model development and guiding deployment decisions. However, many benchmarks quickly become saturated, meaning that they can no longer differentiate between the best-performing models, diminishing their long-term value. In this study, we analyze benchmark saturation across 60 Large Language Model (LLM) benchmarks selected from technical reports by major model developers. To identify factors driving saturation, we characterize benchmarks along 14 properties spanning task design, data construction, and evaluation format. We test five hypotheses examining how each property contributes to saturation rates. Our analysis reveals that nearly half of the benchmarks exhibit saturation, with rates increasing as benchmarks age. Notably, hiding test data (i.e., public vs. private) shows no protective effect, while expert-curated benchmarks resist saturation better than crowdsourced ones. Our findings highlight which design choices extend benchmark longevity and inform strategies for more durable evaluation.

Yijie Tong, Yifan Hou, Shaobo Cui et al.

Test-time compute (TTC) strategies have emerged as a lightweight approach to boost reasoning in large language models (LLMs). However, their application and benefits for vision-language models (VLMs) remain underexplored. We present a systematic study of TTC across seven VLMs and six benchmarks, specifically analyzing feature-based scoring and majority voting methods. We find that feature heuristics fail and voting yields only modest gains in single-model settings. We theoretically show that this limitation stems from a lack of prediction diversity: when outputs are highly correlated, voting provides little benefit. In contrast, multi-model ensembles offer richer diversity, yet standard majority voting fails to account for varying model capabilities. To address this, we propose Entropy-based TTC (ETTC), which selects the most confident prediction based on predictive entropy. Our method reduces to majority voting in the single-model case, but in model ensembles, it leverages confidence disparities to prioritize stronger models. We prove that ETTC outperforms majority voting under mild assumptions and empirically demonstrate that it consistently surpasses both voting and the best individual model. Crucially, our results show that smaller models can synergistically enhance larger ones, unlocking ensembling gains not achievable with standard strategies.

Yuchen Eleanor Jiang, Tianyu Liu, Shuming Ma et al. · eth-zurich

Machine translation (MT) has almost achieved human parity at sentence-level translation. In response, the MT community has, in part, shifted its focus to document-level translation. However, the development of document-level MT systems is hampered by the lack of parallel document corpora. This paper describes BWB, a large parallel corpus first introduced in Jiang et al. (2022), along with an annotated test set. The BWB corpus consists of Chinese novels translated by experts into English, and the annotated test set is designed to probe the ability of machine translation systems to model various discourse phenomena. Our resource is freely available, and we hope it will serve as a guide and inspiration for more work in document-level machine translation.

Justus Mattern, Zhijing Jin, Benjamin Weggenmann et al.

To protect the privacy of individuals whose data is being shared, it is of high importance to develop methods allowing researchers and companies to release textual data while providing formal privacy guarantees to its originators. In the field of NLP, substantial efforts have been directed at building mechanisms following the framework of local differential privacy, thereby anonymizing individual text samples before releasing them. In practice, these approaches are often dissatisfying in terms of the quality of their output language due to the strong noise required for local differential privacy. In this paper, we approach the problem at hand using global differential privacy, particularly by training a generative language model in a differentially private manner and consequently sampling data from it. Using natural language prompts and a new prompt-mismatch loss, we are able to create highly accurate and fluent textual datasets taking on specific desired attributes such as sentiment or topic and resembling statistical properties of the training data. We perform thorough experiments indicating that our synthetic datasets do not leak information from our original data and are of high language quality and highly suitable for training models for further analysis on real-world data. Notably, we also demonstrate that training classifiers on private synthetic data outperforms directly training classifiers on real data with DP-SGD.

Tianyu Liu, Yuchen Jiang, Nicholas Monath et al.

Recent years have seen a paradigm shift in NLP towards using pretrained language models ({PLM}) for a wide range of tasks. However, there are many difficult design decisions to represent structures (e.g. tagged text, coreference chains) in a way such that they can be captured by PLMs. Prior work on structured prediction with PLMs typically flattens the structured output into a sequence, which limits the quality of structural information being learned and leads to inferior performance compared to classic discriminative models. In this work, we describe an approach to model structures as sequences of actions in an autoregressive manner with PLMs, allowing in-structure dependencies to be learned without any loss. Our approach achieves the new state-of-the-art on all the structured prediction tasks we looked at, namely, named entity recognition, end-to-end relation extraction, and coreference resolution.

Minjing Shi, Junling Wang, Jingwei Ni et al.

Identifying logical fallacies in everyday discourse is challenging for many people. This challenge is amplified in the era of Large Language Models (LLMs), where malicious agents can deploy fallacious arguments to disseminate misinformation at scale. In this work, we explore the potential of LLMs as part of the solution. We introduce LFTutor, an intelligent tutoring system which uses LLMs to tutor laypeople and help them learn about logical fallacies. LFTutor integrates intent-driven Socratic questioning and critical argumentation principles to actively engage learners to reflect on their reasoning. Through both automatic and human evaluations, we demonstrate that LFTutor significantly outperforms baseline LLMs lacking these pedagogical strategies. This work highlights the promise of combining LLMs with pedagogical scaffolding to foster critical thinking and argument literacy in the age of AI.

Peng Cui, Vilém Zouhar, Xiaoyu Zhang et al. · eth-zurich

Using questions in written text is an effective strategy to enhance readability. However, what makes an active reading question good, what the linguistic role of these questions is, and what is their impact on human reading remains understudied. We introduce GuidingQ, a dataset of 10K in-text questions from textbooks and scientific articles. By analyzing the dataset, we present a comprehensive understanding of the use, distribution, and linguistic characteristics of these questions. Then, we explore various approaches to generate such questions using language models. Our results highlight the importance of capturing inter-question relationships and the challenge of question position identification in generating these questions. Finally, we conduct a human study to understand the implication of such questions on reading comprehension. We find that the generated questions are of high quality and are almost as effective as human-written questions in terms of improving readers' memorization and comprehension.

Ruida Wang, Wangchunshu Zhou, Mrinmaya Sachan

*Data Synthesis* is a promising way to train a small model with very little labeled data. One approach for data synthesis is to leverage the rich knowledge from large language models to synthesize pseudo training examples for small models, making it possible to achieve both data and compute efficiency at the same time. However, a key challenge in data synthesis is that the synthesized dataset often suffers from a large distributional discrepancy from the *real task* data distribution. Thus, in this paper, we propose *Synthesis Step by Step* (**S3**), a data synthesis framework that shrinks this distribution gap by iteratively extrapolating the errors made by a small model trained on the synthesized dataset on a small real-world validation dataset using a large language model. Extensive experiments on multiple NLP tasks show that our approach improves the performance of a small model by reducing the gap between the synthetic dataset and the real data, resulting in significant improvement compared to several baselines: 9.48% improvement compared to ZeroGen and 2.73% compared to GoldGen, and at most 15.17% improvement compared to the small model trained on human-annotated data.

Peng Cui, Mrinmaya Sachan

Adaptive learning aims to provide customized educational activities (e.g., exercises) to address individual learning needs. However, manual construction and delivery of such activities is a laborious process. Thus, in this paper, we study a novel task of adaptive and personalized exercise generation for online language learning. To this end, we combine a knowledge tracing model that estimates each student's evolving knowledge states from their learning history and a controlled text generation model that generates exercise sentences based on the student's current estimated knowledge state and instructor requirements of desired properties (e.g., domain knowledge and difficulty). We train and evaluate our model on real-world learner interaction data from Duolingo and demonstrate that LMs guided by student states can generate superior exercises. Then, we discuss the potential use of our model in educational applications using various simulations. These simulations show that our model can adapt to students' individual abilities and can facilitate their learning efficiency by personalizing learning sequences.

Đorđe Miladinović, Kumar Shridhar, Kushal Jain et al.

In principle, applying variational autoencoders (VAEs) to sequential data offers a method for controlled sequence generation, manipulation, and structured representation learning. However, training sequence VAEs is challenging: autoregressive decoders can often explain the data without utilizing the latent space, known as posterior collapse. To mitigate this, state-of-the-art models weaken the powerful decoder by applying uniformly random dropout to the decoder input. We show theoretically that this removes pointwise mutual information provided by the decoder input, which is compensated for by utilizing the latent space. We then propose an adversarial training strategy to achieve information-based stochastic dropout. Compared to uniform dropout on standard text benchmark datasets, our targeted approach increases both sequence modeling performance and the information captured in the latent space.

Daphna Keidar, Andreas Opedal, Zhijing Jin et al.

Languages are continuously undergoing changes, and the mechanisms that underlie these changes are still a matter of debate. In this work, we approach language evolution through the lens of causality in order to model not only how various distributional factors associate with language change, but how they causally affect it. In particular, we study slang, which is an informal language that is typically restricted to a specific group or social setting. We analyze the semantic change and frequency shift of slang words and compare them to those of standard, nonslang words. With causal discovery and causal inference techniques, we measure the effect that word type (slang/nonslang) has on both semantic change and frequency shift, as well as its relationship to frequency, polysemy and part of speech. Our analysis provides some new insights in the study of language change, e.g., we show that slang words undergo less semantic change but tend to have larger frequency shifts over time.

Junling Wang, Boqi Chen, Heejin Do et al.

AI systems are increasingly used to support educational content creation, yet it remains unclear whether they can generate outputs that faithfully represent the pedagogical concepts they are intended to teach. Thus, we introduce equation-to-visual generation, a task that, in contrast to conventional image generation, requires producing pedagogically meaningful visuals from arithmetic equations while precisely preserving their numerical and relational structure. Informed by interviews with teachers and an analysis of educational materials, we construct E2V-Bench, a benchmark spanning four pedagogically grounded visual types, along with automatic metrics for evaluating visual correctness. Our evaluation reveals that recent text-to-image (T2I) models frequently fail on this task, with errors dominated by incorrect object counts and broken relational structure. Building on this, we explore benchmark-guided enhancement strategies. These strategies improve representative models, while the remaining gap calls for stronger numerical and relational grounding in future T2I models.

Shehzaad Dhuliawala, Leonard Adolphs, Rajarshi Das et al.

In typical machine learning systems, an estimate of the probability of the prediction is used to assess the system's confidence in the prediction. This confidence measure is usually uncalibrated; i.e.\ the system's confidence in the prediction does not match the true probability of the predicted output. In this paper, we present an investigation into calibrating open setting machine reading systems such as open-domain question answering and claim verification systems. We show that calibrating such complex systems which contain discrete retrieval and deep reading components is challenging and current calibration techniques fail to scale to these settings. We propose simple extensions to existing calibration approaches that allows us to adapt them to these settings. Our experimental results reveal that the approach works well, and can be useful to selectively predict answers when question answering systems are posed with unanswerable or out-of-the-training distribution questions.

Lingzhi Wang, Mrinmaya Sachan, Xingshan Zeng et al.

Conversational tutoring systems (CTSs) aim to help students master educational material with natural language interaction in the form of a dialog. CTSs have become a key pillar in educational data mining research. A key challenge in CTSs is to engage the student in the conversation while exposing them to a diverse set of teaching strategies, akin to a human teacher, thereby, helping them learn in the process. Different from previous work that generates responses given the strategies as input, we propose to jointly predict teaching strategies and generate tutor responses accordingly, which fits a more realistic application scenario. We benchmark several competitive models on three dialog tutoring datasets and propose a unified framework that combines teaching response generation and pedagogical strategy prediction, where a self-distillation mechanism is adopted to guide the teaching strategy learning and facilitate tutor response generation. Our experiments and analyses shed light on how teaching strategies affect dialog tutoring.

Jiaoda Li, Ryan Cotterell, Mrinmaya Sachan

Probing is a popular method to discern what linguistic information is contained in the representations of pre-trained language models. However, the mechanism of selecting the probe model has recently been subject to intense debate, as it is not clear if the probes are merely extracting information or modeling the linguistic property themselves. To address this challenge, this paper introduces a novel model-free approach to probing, by formulating probing as a prompting task. We conduct experiments on five probing tasks and show that our approach is comparable or better at extracting information than diagnostic probes while learning much less on its own. We further combine the probing via prompting approach with attention head pruning to analyze where the model stores the linguistic information in its architecture. We then examine the usefulness of a specific linguistic property for pre-training by removing the heads that are essential to that property and evaluating the resulting model's performance on language modeling.

Xingzhou Pang, Yifan Hou, Junling Wang et al.

While Large Vision-Language Models (VLMs) excel at interpolation, they suffer catastrophic failures in systematic generalization, most notably in visual counting. In this work, we investigate this extrapolation bottleneck by deconstructing visual counting into three cognitive stages: visual individuation, magnitude awareness, and symbolic mapping. Using synthetic Go boards and linear probes, we demonstrate that visual backbones maintain robust, linearly separable representations of quantity well into the extrapolation regime, ruling out perceptual failure. Furthermore, models retain latent magnitude awareness, successfully performing comparative reasoning on quantities they fail to enumerate. We pinpoint the collapse to the symbolic mapping stage, where the model fails to project valid visual magnitudes onto symbolic tokens. Our findings support a frac tured magnitude hypothesis: VLMs fail to acquire a universal number space, instead learning disjoint, modality-specific statistical manifolds that prevent cross-modal grounding for unseen quantities. Validated on the state-of-the-art foundation model, our results suggest that bridging this gap requires inductive priors enforcing unified representations, as data scaling alone is insufficient.

Mattia Atzeni, Mrinmaya Sachan, Andreas Loukas

The Abstraction and Reasoning Corpus (ARC) (Chollet, 2019) and its most recent language-complete instantiation (LARC) has been postulated as an important step towards general AI. Yet, even state-of-the-art machine learning models struggle to achieve meaningful performance on these problems, falling behind non-learning based approaches. We argue that solving these tasks requires extreme generalization that can only be achieved by proper accounting for core knowledge priors. As a step towards this goal, we focus on geometry priors and introduce LatFormer, a model that incorporates lattice symmetry priors in attention masks. We show that, for any transformation of the hypercubic lattice, there exists a binary attention mask that implements that group action. Hence, our study motivates a modification to the standard attention mechanism, where attention weights are scaled using soft masks generated by a convolutional network. Experiments on synthetic geometric reasoning show that LatFormer requires 2 orders of magnitude fewer data than standard attention and transformers. Moreover, our results on ARC and LARC tasks that incorporate geometric priors provide preliminary evidence that these complex datasets do not lie out of the reach of deep learning models.

Tianyu Liu, Yuchen Eleanor Jiang, Ryan Cotterell et al.

Many natural language processing tasks, e.g., coreference resolution and semantic role labeling, require selecting text spans and making decisions about them. A typical approach to such tasks is to score all possible spans and greedily select spans for task-specific downstream processing. This approach, however, does not incorporate any inductive bias about what sort of spans ought to be selected, e.g., that selected spans tend to be syntactic constituents. In this paper, we propose a novel grammar-based structured span selection model which learns to make use of the partial span-level annotation provided for such problems. Compared to previous approaches, our approach gets rid of the heuristic greedy span selection scheme, allowing us to model the downstream task on an optimal set of spans. We evaluate our model on two popular span prediction tasks: coreference resolution and semantic role labeling. We show empirical improvements on both.

Yuchen Eleanor Jiang, Ryan Cotterell, Mrinmaya Sachan

Centering theory (CT; Grosz et al., 1995) provides a linguistic analysis of the structure of discourse. According to the theory, local coherence of discourse arises from the manner and extent to which successive utterances make reference to the same entities. In this paper, we investigate the connection between centering theory and modern coreference resolution systems. We provide an operationalization of centering and systematically investigate if neural coreference resolvers adhere to the rules of centering theory by defining various discourse metrics and developing a search-based methodology. Our information-theoretic analysis reveals a positive dependence between coreference and centering; but also shows that high-quality neural coreference resolvers may not benefit much from explicitly modeling centering ideas. Our analysis further shows that contextualized embeddings contain much of the coherence information, which helps explain why CT can only provide little gains to modern neural coreference resolvers which make use of pretrained representations. Finally, we discuss factors that contribute to coreference which are not modeled by CT such as world knowledge and recency bias. We formulate a version of CT that also models recency and show that it captures coreference information better compared to vanilla CT.

Saüc Abadal Lloret, Shehzaad Dhuliawala, Keerthiram Murugesan et al.

We present ALT (ALignment with Textual feedback), an approach that aligns language models with user preferences expressed in text. We argue that text offers greater expressiveness, enabling users to provide richer feedback than simple comparative preferences and this richer feedback can lead to more efficient and effective alignment. ALT aligns the model by conditioning its generation on the textual feedback. Our method relies solely on language modeling techniques and requires minimal hyper-parameter tuning, though it still presents the main benefits of RL-based alignment algorithms and can effectively learn from textual feedback. We explore the efficacy and efficiency of textual feedback across different tasks such as toxicity reduction, summarization, and dialog response generation. We find that ALT outperforms PPO for the task of toxicity reduction while being able to match its performance on summarization with only 20% of the samples. We also explore how ALT can be used with feedback provided by an existing LLM where we explore an LLM providing constrained and unconstrained textual feedback. We also outline future directions to align models with natural language feedback.

Yanick Zengaffinen, Andreas Opedal, Donya Rooein et al.

Modeling plausible student misconceptions is critical for AI in education. In this work, we examine how large language models (LLMs) reason about misconceptions when generating multiple-choice distractors, a task that requires modeling incorrect yet plausible answers by coordinating solution knowledge, simulating student misconceptions, and evaluating plausibility. We introduce a taxonomy for analyzing the strategies used by state-of-the-art LLMs, examining their reasoning procedures and comparing them to established best practices in the learning sciences. Our structured analysis reveals a surprising alignment between their processes and best practices: the models typically solve the problem correctly first, then articulate and simulate multiple potential misconceptions, and finally select a set of distractors. An analysis of failure modes reveals that errors arise primarily from failures in recovering the correct solution and selecting among response candidates, rather than simulating errors or structuring the process. Consistent with these results, we find that providing the correct solution in the prompt improves alignment with human-authored distractors by 8%, highlighting the critical role of anchoring to the correct solution when generating plausible incorrect student reasoning. Overall, our analysis offers a structured and interpretable lens into LLMs' ability to model incorrect student reasoning and produce high-quality distractors.