In various fields of knowledge creation, including science, new ideas often build on pre-existing information. In this work, we explore this concept within the context of language models. Specifically, we explore the potential of self-training models on their own outputs, akin to how humans learn and build on their previous thoughts and actions. While this approach is intuitively appealing, our research reveals its practical limitations. We find that extended self-training of the GPT-2 model leads to a significant degradation in performance, resulting in repetitive and collapsed token output.

Large Language Models (LLMs) have drawn a lot of attention due to their strong performance on a wide range of natural language tasks, since the release of ChatGPT in November 2022. LLMs' ability of general-purpose language understanding and generation is acquired by training billions of model's parameters on massive amounts of text data, as predicted by scaling laws \cite{kaplan2020scaling,hoffmann2022training}. The research area of LLMs, while very recent, is evolving rapidly in many different ways. In this paper, we review some of the most prominent LLMs, including three popular LLM families (GPT, LLaMA, PaLM), and discuss their characteristics, contributions and limitations. We also give an overview of techniques developed to build, and augment LLMs. We then survey popular datasets prepared for LLM training, fine-tuning, and evaluation, review widely used LLM evaluation metrics, and compare the performance of several popular LLMs on a set of representative benchmarks. Finally, we conclude the paper by discussing open challenges and future research directions.

Recent advancements in large language models (LLMs) have sparked a debate on the detection of artificial intelligence (AI)-generated text, a concern especially prevalent among academic institutions and publishers. While current detection tools claim high accuracy rates, some studies point to their unreliability. This paper contends that efforts to detect AI writing are fundamentally flawed because improved detection capabilities could inadvertently refine AI writing tools, leading to a technological arms race. Moreover, the rapid evolution of LLMs means detection methods may quickly become obsolete. We propose a focus on ethical guidelines rather than outright prohibitions, emphasizing that technological solutions should complement, not replace, the core ethical principles of scientific publishing.

How much is 56 times 37? Language models often make mistakes in these types of difficult calculations. This is usually explained by their inability to perform complex reasoning. Since language models rely on large training sets and great memorization capability, naturally they are not equipped to run complex calculations. However, one can argue that humans also cannot perform this calculation immediately and require a considerable amount of time to construct the solution. In order to enhance the generalization capability of language models, and as a parallel to human behavior, we propose to use special 'thinking tokens' which allow the model to perform much more calculations whenever a complex problem is encountered.

Who is the US President? The answer changes depending on when the question is asked. While large language models (LLMs) are evaluated on various reasoning tasks, they often miss a crucial dimension: time. In real-world scenarios, the correctness of answers is frequently tied to temporal context. To address this gap, we present a novel framework and dataset spanning over 8,000 events from 2018 to 2024, annotated with day-level granularity and sourced globally across domains such as politics, science, and business. Our TimeShift evaluation method systematically probes LLMs for temporal reasoning, revealing that base models often outperform instruction-tuned and synthetic-trained counterparts on time-sensitive recall. Additionally, we find that even large-scale models exhibit brittleness in handling paraphrased facts, highlighting unresolved challenges in temporal consistency. By identifying these limitations, our work provides a significant step toward advancing time-aware language models capable of adapting to the dynamic nature of real-world knowledge.

Generalization is arguably the most important goal of statistical language modeling research. Publicly available benchmarks and papers published with an open-source code have been critical to advancing the field. However, it is often very difficult, and sometimes even impossible, to reproduce the results fully as reported in publications. In this paper, we propose a simple framework that should help advance the state of the art in language modeling in terms of generalization. We propose to publish not just the code, but also probabilities on dev and test sets with future publications so that one can easily add the new model into an ensemble. This has crucial advantages: it is much easier to determine whether a newly proposed model is actually complementary to the current baseline. Therefore, instead of inventing new names for the old tricks, the scientific community can advance faster. Finally, this approach promotes diversity of ideas: one does not need to create an individual model that is the new state of the art to attract attention; it will be sufficient to develop a new model that learns patterns which other models do not. Thus, even a suboptimal model can be found to have value. Remarkably, our approach has yielded new state-of-the-art results across various language modeling benchmarks up to 10%.

The growth of hateful online content, or hate speech, has been associated with a global increase in violent crimes against minorities [23]. Harmful online content can be produced easily, automatically and anonymously. Even though, some form of auto-detection is already achieved through text classifiers in NLP, they can be fooled by adversarial attacks. To strengthen existing systems and stay ahead of attackers, we need better adversarial attacks. In this paper, we show that up to 70% of adversarial examples generated by adversarial attacks should be discarded because they do not preserve semantics. We address this core weakness and propose a new, fully supervised sentence embedding technique called Semantics-Preserving-Encoder (SPE). Our method outperforms existing sentence encoders used in adversarial attacks by achieving 1.2× ∼ 5.1× better real attack success rate. We release our code as a plugin that can be used in any existing adversarial attack to improve its quality and speed up its execution. (The code, datasets and test examples are available at https://github.com/DavidHerel/semantics-preserving-encoder.)

not applicable for Letter to the editor

Adversarial attacks in NLP challenge the way we look at language models. The goal of this kind of adversarial attack is to modify the input text to fool a classifier while maintaining the original meaning of the text. Although most existing adversarial attacks claim to fulfill the constraint of semantics preservation, careful scrutiny shows otherwise. We show that the problem lies in the text encoders used to determine the similarity of adversarial examples, specifically in the way they are trained. Unsupervised training methods make these encoders more susceptible to problems with antonym recognition. To overcome this, we introduce a simple, fully supervised sentence embedding technique called Semantics-Preserving-Encoder (SPE). The results show that our solution minimizes the variation in the meaning of the adversarial examples generated. It also significantly improves the overall quality of adversarial examples, as confirmed by human evaluators. Furthermore, it can be used as a component in any existing attack to speed up its execution while maintaining similar attack success.

One of the main goals of Artificial Life is to research the conditions for the emergence of life, not necessarily as it is, but as it could be. Artificial chemistries are one of the most important tools for this purpose because they provide us with a basic framework to investigate under which conditions metabolisms capable of reproducing themselves, and ultimately, of evolving, can emerge. While there have been successful attempts at producing examples of emergent self-reproducing metabolisms, the set of rules involved remain too complex to shed much light on the underlying principles at work. In this article, we hypothesize that the key property needed for self-reproducing metabolisms to emerge is the existence of an autocatalyzed subset of Turing-complete reactions. We validate this hypothesis with a minimalistic artificial chemistry with conservation laws, which is based on a Turing-complete rewriting system called combinatory logic. Our experiments show that a single run of this chemistry, starting from a tabula rasa state, discovers-with no external intervention-a wide range of emergent structures including ones that self-reproduce in each cycle. All of these structures take the form of recursive algorithms that acquire basic constituents from the environment and decompose them in a process that is remarkably similar to biological metabolisms.

One of the main problems of evolutionary algorithms is the convergence of the population to local minima. In this paper, we explore techniques that can avoid this problem by encouraging a diverse behavior of the agents through a shared reward system. The rewards are randomly distributed in the environment, and the agents are only rewarded for collecting them first. This leads to an emergence of a novel behavior of the agents. We introduce our approach to the maze problem and compare it to the previously proposed solution, denoted as Novelty Search (Lehman and Stanley, 2011a). We find that our solution leads to an improved performance while being significantly simpler. Building on that, we generalize the problem and apply our approach to a more advanced set of tasks, Atari Games, where we observe a similar performance quality with much less computational power needed.

It is common to evaluate the performance of a machine learning model by measuring its predictive power on a test dataset. This approach favors complicated models that can smoothly fit complex functions and generalize well from training data points. Although essential components of intelligence, speed and data efficiency of this learning process are rarely reported or compared between different candidate models. In this paper, we introduce a benchmark of increasingly difficult tasks together with a data efficiency metric to measure how quickly machine learning models learn from training data. We compare the learning speed of some established sequential supervised models, such as RNNs, LSTMs, or Transformers, with relatively less known alternative models based on reservoir computing. The proposed tasks require a wide range of computational primitives, such as memory or the ability to compute Boolean functions, to be effectively solved. Surprisingly, we observe that reservoir computing systems that rely on dynamically evolving feature maps learn faster than fully supervised methods trained with stochastic gradient optimization while achieving comparable accuracy scores. The code, benchmark, trained models, and results to reproduce our experiments are available at https://github.com/hugcis/benchmark_learning_efficiency/ .

Cellular automata are fascinating models of parallel computation, yet it remains a challenge to program them efficiently to solve complex tasks. In this paper, we propose a novel method of analyzing their computational capacity. It is based on measuring the richness of the automaton's dynamics by assessing how many other automata it can simulate. Using this method, we can identify automata that can perform many computations efficiently. Moreover, there is no need to choose an arbitrary set of computational tasks as they are self-defined by the cellular automata space itself. The novelty of our method is based on a new definition of automata simulation that generalizes previous approaches. Thus, we obtain a rich structure of relations that we demonstrate on the class of elementary cellular automata. We show that the number of elementary automata with unique dynamics can be reduced from 88 to 52. We further show that some automata, such as 14, 43, and 142, have in fact equivalent computational capacity. We believe that using similar approaches, the number of unique automata can be dramatically reduced in the future.

Readability assessment, which predicts how 001 difficult it is for the reader to understand a 002 text, has mostly been performed at the pas-003 sage level. There has been recent interest in 004 sentence-level assessment, given its applica-005 tions in downstream tasks such as text simplifi-006 cation and language exercise generation. While 007 research has shown that neural approaches can 008 improve the assessment accuracy on the pas-009 sage levels, they have yet to be applied on the 010 sentence level. This paper evaluates a neural 011 model on readability assessment at the sen-012 tence level, and shows that its performance can 013 be further improved by integrating linguistic 014 features. On a dataset of 13,809 Chinese sen-015 tences ranging from Grade 3 to 12, we obtained 016 the strongest performance with the a modified 017 BERT model that incorporates linguistic fea-018 tures. 019

In order to develop systems capable of artificial evolution, we need to identify which systems can produce complex behavior. We present a novel classification method applicable to any class of deterministic discrete space and time dynamical systems. The method is based on classifying the asymptotic behavior of the average computation time in a given system before entering a loop. We were able to identify a critical region of behavior that corresponds to a phase transition from ordered behavior to chaos across various classes of dynamical systems. To show that our approach can be applied to many different computational systems, we demonstrate the results of classifying cellular automata, Turing machines, and random Boolean networks. Further, we use this method to classify 2D cellular automata to automatically find those with interesting, complex dynamics. We believe that our work can be used to design systems in which complex structures emerge. Also, it can be used to compare various versions of existing attempts to model open-ended evolution (Channon, 2006; Ofria & Wilke, 2004; Ray, 1991).

The complexity of cellular automata is traditionally measured by their computational capacity. However, it is difficult to choose a challenging set of computational tasks suitable for the parallel nature of such systems. We study the ability of automata to emulate one another, and we use this notion to define such a set of naturally emerging tasks. We present the results for elementary cellular automata, although the core ideas can be extended to other computational systems. We compute a graph showing which elementary cellular automata can be emulated by which and show that certain chaotic automata are the only ones that cannot emulate any automata non-trivially. Finally, we use the emulation notion to suggest a novel definition of chaos that we believe is suitable for discrete computational systems. We believe our work can help design parallel computational systems that are Turing-complete and also computationally efficient.

In this paper, we propose that the dot product pairwise matching attention layer, which is widely used in Transformer-based models, is redundant for the model performance. Attention, in its original formulation, has to be seen rather as a human-level tool to explore and/or visualize relevancy scores in sequential data. However, the way how it is constructed leads to significant computational complexity. Instead, we present SimpleTRON: Simple Transformer with O(N) Complexity, a simple and fast alternative without any approximation that, unlike other approximation models, does not have any architecture-related overhead and therefore can be seen as a purely linear Transformer-like model. This architecture, to the best of our knowledge, outperforms existing sub-quadratic attention approximation models on several tasks from the Long-Range Arena benchmark. Moreover, we show, that SimpleTRON can benefit from weight transfer from pretrained large language models, as its parameters can be fully transferable.

Online Continual Learning (OCL) studies learning over a continuous data stream without observing any single example more than once, a setting that is closer to the experience of humans and systems that must learn "on-the-wild". Yet, commonly available benchmarks are far from these real-world conditions, because they explicitly signal different tasks, lack latent similarity structure or assume temporal independence between different examples. Here, we propose a new benchmark for OCL based on language modelling in which input alternates between different languages and domains without any explicit delimitation. Additionally, we propose new metrics to study catastrophic forgetting in this setting and evaluate multiple baseline models based on compositions of experts. Finally, we introduce a simple gating technique that learns the latent similarities between different inputs, improving the performance of a Products of Experts model.