ComplexityMT: Benchmarking the Interaction Between Text Complexity and Machine Translation

When a text is translated, does the translation retain the complexity of the original? We introduce ComplexityMT, a new challenge for assessing how text complexity and machine translation interact with and influence each other, using the Common European Framework of Reference for Languages (CEFR) levels as the measure of text complexity. Across six languages, including Arabic, Dutch, English, French, Hindi, and Russian, we evaluate three open-weight models, one closed model, and a commercial machine translation system on two tasks: i) correlation of CEFR with translation difficulty, and ii) shifts in CEFR levels of the source texts. Our experiments show that higher CEFR levels make texts more difficult to translate, and that machine translation shifts the CEFR level of the target text compared to the original source, for most languages. These findings provide new insights for researchers and practitioners working on multilingual pedagogical content generation and machine translation difficulty estimation.

Linguistic Productivity in Large Language Models: Models Coerce, but do not Preempt

Usage-based theories of grammars posit that creative productivity of the structures of language is both bolstered and constrained by two distinct frequency signals: entrenchment, stemming from high frequency usage, and preemption, stemming from having never observed a particular linguistic structure in a context where one might expect that structure to appear. Large Language Models are also usage-based, in the sense that the structures of language are learned through exposure to vast amounts of text. Here, we test whether or not the opposing statistical forces of entrenchment and preemption also encourage and constrain linguistic productivity in LLMs. We demonstrate across model architectures that larger models recognize and can reproduce with nonce words constructional productivity (entrenchment) in cases of coercion, wherein the broader constructional context coerces an atypical interpretation of a lexical item. However, we also show that even the largest models do not extend negative evidence to novel language, and statistical preemption does not enable models to avoid overgeneralization of patterns that are semantically felicitous, but never observed in data.

Safer Policy Compliance with Dynamic Epistemic Fallback

Humans develop a series of cognitive defenses, known as epistemic vigilance, to combat risks of deception and misinformation from everyday interactions. Developing safeguards for LLMs inspired by this mechanism might be particularly helpful for their application in high-stakes tasks such as automating compliance with data privacy laws. In this paper, we introduce Dynamic Epistemic Fallback (DEF), a dynamic safety protocol for improving an LLM's inference-time defenses against deceptive attacks that make use of maliciously perturbed policy texts. Through various levels of one-sentence textual cues, DEF nudges LLMs to flag inconsistencies, refuse compliance, and fallback to their parametric knowledge upon encountering perturbed policy texts. Using globally recognized legal policies such as HIPAA and GDPR, our empirical evaluations report that DEF effectively improves the capability of frontier LLMs to detect and refuse perturbed versions of policies, with DeepSeek-R1 achieving a 100% detection rate in one setting. This work encourages further efforts to develop cognitively inspired defenses to improve LLM robustness against forms of harm and deception that exploit legal artifacts.

Neither Stochastic Parroting nor AGI: LLMs Solve Tasks through Context-Directed Extrapolation from Training Data Priors

In this position paper we raise critical awareness of a realistic view of LLM capabilities that eschews extreme alternative views that LLMs are either "stochastic parrots" or in possession of "emergent" advanced reasoning capabilities, which, due to their unpredictable emergence, constitute an existential threat. Our middle-ground view is that LLMs extrapolate from priors from their training data, and that a mechanism akin to in-context learning enables the targeting of the appropriate information from which to extrapolate. We call this "context-directed extrapolation." Under this view, substantiated though existing literature, while reasoning capabilities go well beyond stochastic parroting, such capabilities are predictable, controllable, not indicative of advanced reasoning akin to high-level cognitive capabilities in humans, and not infinitely scalable with additional training. As a result, fears of uncontrollable emergence of agency are allayed, while research advances are appropriately refocused on the processes of context-directed extrapolation and how this interacts with training data to produce valuable capabilities in LLMs. Future work can therefore explore alternative augmenting techniques that do not rely on inherent advanced reasoning in LLMs.

Illusion or Algorithm? Investigating Memorization, Emergence, and Symbolic Processing in In-Context Learning

Large-scale Transformer language models (LMs) trained solely on next-token prediction with web-scale data can solve a wide range of tasks after seeing just a few examples. The mechanism behind this capability, known as in-context learning (ICL), remains both controversial and poorly understood. Some studies argue that it is merely the result of memorizing vast amounts of data, while others contend that it reflects a fundamental, symbolic algorithmic development in LMs. In this work, we introduce a suite of investigative tasks and a novel method to systematically investigate ICL by leveraging the full Pythia scaling suite, including interim checkpoints that capture progressively larger amount of training data. By carefully exploring ICL performance on downstream tasks and simultaneously conducting a mechanistic analysis of the residual stream's subspace, we demonstrate that ICL extends beyond mere "memorization" of the training corpus, yet does not amount to the implementation of an independent symbolic algorithm. Our results also clarify several aspects of ICL, including the influence of training dynamics, model capabilities, and elements of mechanistic interpretability. Overall, our work advances the understanding of ICL and its implications, offering model developers insights into potential improvements and providing AI security practitioners with a basis for more informed guidelines.

The Inherent Limits of Pretrained LLMs: The Unexpected Convergence of Instruction Tuning and In-Context Learning Capabilities

Large Language Models (LLMs), trained on extensive web-scale corpora, have demonstrated remarkable abilities across diverse tasks, especially as they are scaled up. Nevertheless, even state-of-the-art models struggle in certain cases, sometimes failing at problems solvable by young children, indicating that traditional notions of task complexity are insufficient for explaining LLM capabilities. However, exploring LLM capabilities is complicated by the fact that most widely-used models are also "instruction-tuned" to respond appropriately to prompts. With the goal of disentangling the factors influencing LLM performance, we investigate whether instruction-tuned models possess fundamentally different capabilities from base models that are prompted using in-context examples. Through extensive experiments across various model families, scales and task types, which included instruction tuning 90 different LLMs, we demonstrate that the performance of instruction-tuned models is significantly correlated with the in-context performance of their base counterparts. By clarifying what instruction-tuning contributes, we extend prior research into in-context learning, which suggests that base models use priors from pretraining data to solve tasks. Specifically, we extend this understanding to instruction-tuned models, suggesting that their pretraining data similarly sets a limiting boundary on the tasks they can solve, with the added influence of the instruction-tuning dataset.

Adapting Whisper for Regional Dialects: Enhancing Public Services for Vulnerable Populations in the United Kingdom

We collect novel data in the public service domain to evaluate the capability of the state-of-the-art automatic speech recognition (ASR) models in capturing regional differences in accents in the United Kingdom (UK), specifically focusing on two accents from Scotland with distinct dialects. This study addresses real-world problems where biased ASR models can lead to miscommunication in public services, disadvantaging individuals with regional accents particularly those in vulnerable populations. We first examine the out-of-the-box performance of the Whisper large-v3 model on a baseline dataset and our data. We then explore the impact of fine-tuning Whisper on the performance in the two UK regions and investigate the effectiveness of existing model evaluation techniques for our real-world application through manual inspection of model errors. We observe that the Whisper model has a higher word error rate (WER) on our test datasets compared to the baseline data and fine-tuning on a given data improves performance on the test dataset with the same domain and accent. The fine-tuned models also appear to show improved performance when applied to the test data outside of the region it was trained on suggesting that fine-tuned models may be transferable within parts of the UK. Our manual analysis of model outputs reveals the benefits and drawbacks of using WER as an evaluation metric and fine-tuning to adapt to regional dialects.

Assessing Language Comprehension in Large Language Models Using Construction Grammar

Large Language Models, despite their significant capabilities, are known to fail in surprising and unpredictable ways. Evaluating their true `understanding' of language is particularly challenging due to the extensive web-scale data they are trained on. Therefore, we construct an evaluation to systematically assess natural language understanding (NLU) in LLMs by leveraging Construction Grammar (CxG), which provides insights into the meaning captured by linguistic elements known as constructions (Cxns). CxG is well-suited for this purpose because provides a theoretical basis to construct targeted evaluation sets. These datasets are carefully constructed to include examples which are unlikely to appear in pre-training data, yet intuitive and easy for humans to understand, enabling a more targeted and reliable assessment. Our experiments focus on downstream natural language inference and reasoning tasks by comparing LLMs' understanding of the underlying meanings communicated through 8 unique Cxns with that of humans. The results show that while LLMs demonstrate some knowledge of constructional information, even the latest models including GPT-o1 struggle with abstract meanings conveyed by these Cxns, as demonstrated in cases where test sentences are dissimilar to their pre-training data. We argue that such cases provide a more accurate test of true language understanding, highlighting key limitations in LLMs' semantic capabilities. We make our novel dataset and associated experimental data including prompts and model responses publicly available.

SpeciaLex: A Benchmark for In-Context Specialized Lexicon Learning

Specialized lexicons are collections of words with associated constraints such as special definitions, specific roles, and intended target audiences. These constraints are necessary for content generation and documentation tasks (e.g., writing technical manuals or children's books), where the goal is to reduce the ambiguity of text content and increase its overall readability for a specific group of audience. Understanding how large language models can capture these constraints can help researchers build better, more impactful tools for wider use beyond the NLP community. Towards this end, we introduce SpeciaLex, a benchmark for evaluating a language model's ability to follow specialized lexicon-based constraints across 18 diverse subtasks with 1,285 test instances covering core tasks of Checking, Identification, Rewriting, and Open Generation. We present an empirical evaluation of 15 open and closed-source LLMs and discuss insights on how factors such as model scale, openness, setup, and recency affect performance upon evaluating with the benchmark.

Fine-Tuning with Divergent Chains of Thought Boosts Reasoning Through Self-Correction in Language Models

Requiring a Large Language Model to generate intermediary reasoning steps has been shown to be an effective way of boosting performance. In fact, it has been found that instruction tuning on these intermediary reasoning steps improves model performance. In this work, we present a novel method of further improving performance by requiring models to compare multiple reasoning chains before generating a solution in a single inference step. We call this method Divergent CoT (DCoT). We find that instruction tuning on DCoT datasets boosts the performance of even smaller, and therefore more accessible, LLMs. Through a rigorous set of experiments spanning a wide range of tasks that require various reasoning types, we show that fine-tuning on DCoT consistently improves performance over the CoT baseline across model families and scales (1.3B to 70B). Through a combination of empirical and manual evaluation, we additionally show that these performance gains stem from models generating multiple divergent reasoning chains in a single inference step, indicative of the enabling of self-correction in language models. Our code and data are publicly available at https://github.com/UKPLab/arxiv2024-divergent-cot.