Jonathan Cook, Tim Rocktäschel, Jakob Nicolaus Foerster, Dennis Aumiller, Alex Wang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Large Language Models (LLMs) and Vision Language Models (VLMs) possess extensive knowledge and exhibit promising reasoning abilities, however, they still struggle to perform well in complex, dynamic environments. Real-world tasks require handling intricate interactions, advanced spatial reasoning, long-term planning, and continuous exploration of new strategies—areas in which we lack effective methodologies for comprehensively evaluating these capabilities. To address this gap, we introduce BALROG, a novel benchmark designed to assess the agentic capabilities of LLMs and VLMs through a diverse set of challenging games. Our benchmark incorporates a range of existing reinforcement learning environments with varying levels of difficulty, including tasks that are solvable by non-expert humans in seconds to extremely challenging ones that may take years to master (e.g., the NetHack Learning Environment). We devise fine-grained metrics to measure performance and conduct an extensive evaluation of several popular open-source and closed-source LLMs and VLMs. Our findings indicate that while current models achieve partial success in the easier games, they struggle significantly with more challenging tasks. Notably, we observe severe deficiencies in vision-based decision-making, as several models perform worse when visual representations of the environments are provided. We release BALROG as an open and user-friendly benchmark to facilitate future research and development in the agentic community. Code and Leaderboard at balrogai.com

The capabilities and limitations of Large Language Models (LLMs) have been sketched out in great detail in recent years, providing an intriguing yet conflicting picture. On the one hand, LLMs demonstrate a general ability to solve problems. On the other hand, they show surprising reasoning gaps when compared to humans, casting doubt on the robustness of their generalisation strategies. The sheer volume of data used in the design of LLMs has precluded us from applying the method traditionally used to measure generalisation: train-test set separation. To overcome this, we study what kind of generalisation strategies LLMs employ when performing reasoning tasks by investigating the pretraining data they rely on. For two models of different sizes (7B and 35B) and 2.5B of their pretraining tokens, we identify what documents influence the model outputs for three simple mathematical reasoning tasks and contrast this to the data that are influential for answering factual questions. We find that, while the models rely on mostly distinct sets of data for each factual question, a document often has a similar influence across different reasoning questions within the same task, indicating the presence of procedural knowledge. We further find that the answers to factual questions often show up in the most influential data. However, for reasoning questions the answers usually do not show up as highly influential, nor do the answers to the intermediate reasoning steps. When we characterise the top ranked documents for the reasoning questions qualitatively, we confirm that the influential documents often contain procedural knowledge, like demonstrating how to obtain a solution using formulae or code. Our findings indicate that the approach to reasoning the models use is unlike retrieval, and more like a generalisable strategy that synthesises procedural knowledge from documents doing a similar form of reasoning.

Automatic evaluation methods based on large language models (LLMs) are emerging as the standard tool for assessing the instruction-following abilities of LLM-based agents. The most common method in this paradigm, pairwise comparisons with a baseline model, critically depends on the assumption of transitive preferences. However, the validity of this assumption remains largely unexplored. In this study, we investigate the presence of non-transitivity within the AlpacaEval framework and analyze its effects on model rankings. We find that LLM judges exhibit non-transitive preferences, leading to rankings that are sensitive to the choice of the baseline model. To mitigate this issue, we show that round-robin tournaments combined with Bradley-Terry models of preference can produce more reliable rankings. Notably, our method increases both the Spearman correlation and the Kendall correlation with Chatbot Arena (95.0% $\rightarrow$ 96.4% and 82.1% $\rightarrow$ 86.3% respectively). To address the computational cost of round-robin tournaments, we propose Swiss-Wise Iterative Matchmaking (Swim) tournaments, using a dynamic matching strategy to capture the benefits of round-robin tournaments while maintaining computational efficiency.

Training agents to act in embodied environments typically requires vast training data or access to accurate simulation, neither of which exists for many cases in the real world. Instead, world models are emerging as an alternative–leveraging offline, passively collected data, they make it possible to generate diverse worlds for training agents in simulation. In this work, we harness world models to generate “imagined” environments to train robust agents capable of generalizing to novel task variations. One of the challenges in doing this is ensuring the agent trains on useful generated data. We thus propose a novel approach IMAC (Imagined Autocurricula) leveraging Unsupervised Environment Design (UED), induces an automatic curriculum over generated worlds. In a series of challenging, procedurally generated environments, we show it is possible to achieve strong transfer performance on held-out environments having trained only inside a world model learned from a narrower dataset. We believe this opens the path to utilizing larger-scale, foundation world models for generally capable agents.

Fine-tuning large pre-trained models has become the de facto strategy for developing both task-specific and general-purpose machine learning systems, including developing models that are safe to deploy. Despite its clear importance, there has been minimal work that explains how fine-tuning alters the underlying capabilities learned by a model during pretraining: does fine-tuning yield entirely novel capabilities or does it just modulate existing ones? We address this question empirically in synthetic, controlled settings where we can use mechanistic interpretability tools (e.g., network pruning and probing) to understand how the model's underlying capabilities are changing. We perform an extensive analysis of the effects of fine-tuning in these settings, and show that: (i) fine-tuning rarely alters the underlying model capabilities; (ii) a minimal transformation, which we call a `wrapper', is typically learned on top of the underlying model capabilities, creating the illusion that they have been modified; and (iii) further fine-tuning on a task where such ``wrapped capabilities'' are relevant leads to sample-efficient revival of the capability, i.e., the model begins reusing these capabilities after only a few gradient steps. This indicates that practitioners can unintentionally remove a model's safety wrapper merely by fine-tuning it on a, e.g., superficially unrelated, downstream task. We additionally perform analysis on language models trained on the TinyStories dataset to support our claims in a more realistic setup.

Humanoid control is an important research challenge offering avenues for integration into human-centric infrastructures and enabling physics-driven humanoid animations. The daunting challenges in this field stem from the difficulty of optimizing in high-dimensional action spaces and the instability introduced by the bipedal morphology of humanoids. However, the extensive collection of human motion-captured data and the derived datasets of humanoid trajectories, such as MoCapAct, paves the way to tackle these challenges. In this context, we present Humanoid Generalist Autoencoding Planner (H-GAP), a state-action trajectory generative model trained on humanoid trajectories derived from human motion-captured data, capable of adeptly handling downstream control tasks with Model Predictive Control (MPC). For 56 degrees of freedom humanoid, we empirically demonstrate that H-GAP learns to represent and generate a wide range of motor behaviors. Further, without any learning from online interactions, it can also flexibly transfer these behaviours to solve novel downstream control tasks via planning. Notably, H-GAP excels established MPC baselines with access to the ground truth model, and is superior or comparable to offline RL methods trained for individual tasks. Finally, we do a series of empirical studies on the scaling properties of H-GAP, showing the potential for performance gains via additional data but not computing.

In recent years there has been a tremendous surge in the general capabilities of AI systems, mainly fuelled by training foundation models on internet-scale data. Nevertheless, the creation of open-ended, ever self-improving AI remains elusive. **In this position paper, we argue that the ingredients are now in place to achieve *open-endedness* in AI systems with respect to a human observer. Furthermore, we claim that such open-endedness is an essential property of any artificial superhuman intelligence (ASI).** We begin by providing a concrete formal definition of open-endedness through the lens of novelty and learnability. We then illustrate a path towards ASI via open-ended systems built on top of foundation models, capable of making novel, human-relevant discoveries. We conclude by examining the safety implications of generally-capable open-ended AI. We expect that open-ended foundation models will prove to be an increasingly fertile and safety-critical area of research in the near future.

We introduce Genie, the first *generative interactive environment* trained in an unsupervised manner from unlabelled Internet videos. The model can be prompted to generate an endless variety of action-controllable virtual worlds described through text, synthetic images, photographs, and even sketches. At 11B parameters, Genie can be considered a *foundation world model*. It is comprised of a spatiotemporal video tokenizer, an autoregressive dynamics model, and a simple and scalable latent action model. Genie enables users to act in the generated environments on a frame-by-frame basis *despite training without any ground-truth action labels* or other domain specific requirements typically found in the world model literature. Further the resulting learned latent action space facilitates training agents to imitate behaviors from unseen videos, opening the path for training generalist agents of the future.

Popular prompt strategies like Chain-of-Thought Prompting can dramatically improve the reasoning abilities of Large Language Models (LLMs) in various domains. However, such hand-crafted prompt-strategies are often sub-optimal. In this paper, we present Promptbreeder, a general-purpose self-referential self-improvement mechanism that evolves and adapts prompts for a given domain. Driven by an LLM, Promptbreeder mutates a population of task-prompts, evaluates them for fitness on a training set, and repeats this process over multiple generations to evolve task-prompts. Crucially, the mutation of these task-prompts is governed by mutation-prompts that the LLM generates and improves throughout evolution in a self-referential way. That is, Promptbreeder is not just improving task-prompts, but it is also improving the mutation-prompts that improve these task-prompts. Promptbreeder outperforms state-of-the-art prompt strategies such as Chain-of-Thought and Plan-and-Solve Prompting on commonly used arithmetic and commonsense reasoning benchmarks. Furthermore, Promptbreeder is able to evolve intricate task-prompts for the challenging problem of hate speech classification.

Common methods for aligning large language models (LLMs) with desired behaviour heavily rely on human-labelled data. However, as models grow increasingly sophisticated, they will surpass human expertise, and the role of human evaluation will evolve into non-experts overseeing experts. In anticipation of this, we ask: can weaker models assess the correctness of stronger models? We investigate this question in an analogous setting, where stronger models (experts) possess the necessary information to answer questions and weaker models (non-experts) lack this information. The method we evaluate is debate, where two LLM experts each argue for a different answer, and a non-expert selects the answer. We find that debate consistently helps both non-expert models and humans answer questions, achieving 76% and 88% accuracy respectively (naive baselines obtain 48% and 60%). Furthermore, optimising expert debaters for persuasiveness in an unsupervised manner improves non-expert ability to identify the truth in debates. Our results provide encouraging empirical evidence for the viability of aligning models with debate in the absence of ground truth.

Benchmarks are crucial in the development of machine learning algorithms, significantly influencing reinforcement learning (RL) research through the available environments. Traditionally, RL environments run on the CPU, which limits their scalability with the computational resources typically available in academia. However, recent advancements in JAX have enabled the wider use of hardware acceleration, enabling massively parallel RL training pipelines and environments. While this has been successfully applied to single-agent RL, it has not yet been widely adopted for multi-agent scenarios. In this paper, we present JaxMARL, the first open-source, easy-to-use code base that combines GPU-enabled efficiency with support for a large number of commonly used MARL environments and popular baseline algorithms. Our experiments show that, in terms of wall clock time, our JAX-based training pipeline is up to 12,500 times faster than existing approaches. This enables efficient and thorough evaluations, potentially alleviating the evaluation crisis in the field. We also introduce and benchmark SMAX, a vectorised, simplified version of the popular StarCraft Multi-Agent Challenge, which removes the need to run the StarCraft II game engine. This not only enables GPU acceleration, but also provides a more flexible MARL environment, unlocking the potential for self-play, meta-learning, and other future applications in MARL. The code is available at https://github.com/flairox/jaxmarl.

As large language models (LLMs) become increasingly prevalent across many real-world applications, understanding and enhancing their robustness to adversarial attacks is of paramount importance. Existing methods for identifying adversarial prompts tend to focus on specific domains, lack diversity, or require extensive human annotations. To address these limitations, we present Rainbow Teaming, a novel black-box approach for producing a diverse collection of adversarial prompts. Rainbow Teaming casts adversarial prompt generation as a quality-diversity problem and uses open-ended search to generate prompts that are both effective and diverse. Focusing on the safety domain, we use Rainbow Teaming to target various state-of-the-art LLMs, including the Llama 2 and Llama 3 models. Our approach reveals hundreds of effective adversarial prompts, with an attack success rate exceeding 90% across all tested models. Furthermore, we demonstrate that prompts generated by Rainbow Teaming are highly transferable and that fine-tuning models with synthetic data generated by our method significantly enhances their safety without sacrificing general performance or helpfulness. We additionally explore the versatility of Rainbow Teaming by applying it to question answering and cybersecurity, showcasing its potential to drive robust open-ended self-improvement in a wide range of applications.

Foundation models have shown impressive adaptation and scalability in supervised and self-supervised learning problems, but so far these successes have not fully translated to reinforcement learning (RL). In this work, we demonstrate that training an RL agent at scale leads to a general in-context learning algorithm that can adapt to open-ended novel embodied 3D problems as quickly as humans. In a vast space of held-out environment dynamics, our adaptive agent (AdA) displays on-the-fly hypothesis-driven exploration, efficient exploitation of acquired knowledge, and can successfully be prompted with first-person demonstrations. Adaptation emerges from three ingredients: (1) meta-reinforcement learning across a vast, smooth and diverse task distribution, (2) a policy parameterised as a large-scale attention-based memory architecture, and (3) an effective automated curriculum that prioritises tasks at the frontier of an agent’s capabilities. We demonstrate characteristic scaling laws with respect to network size, memory length, and richness of the training task distribution. We believe our results lay the foundation for increasingly general and adaptive RL agents that perform well across ever-larger open-ended domains.

Despite widespread use of LLMs as conversational agents, evaluations of performance fail to capture a crucial aspect of communication: interpreting language in context---incorporating its pragmatics. Humans interpret language using beliefs and prior knowledge about the world. For example, we intuitively understand the response "I wore gloves" to the question "Did you leave fingerprints?" as meaning "No". To investigate whether LLMs have the ability to make this type of inference, known as an implicature, we design a simple task and evaluate four categories of widely used state-of-the-art models. We find that, despite only evaluating on utterances that require a binary inference (yes or no), models in three of these categories perform close to random. However, LLMs instruction-tuned at the example-level perform significantly better. These results suggest that certain fine-tuning strategies are far better at inducing pragmatic understanding in models. We present our findings as the starting point for further research into evaluating how LLMs interpret language in context and to drive the development of more pragmatic and useful models of human discourse.

The study of zero-shot generalisation (ZSG) in deep Reinforcement Learning (RL) aims to produce RL algorithms whose policies generalise well to novel unseen situations at deployment time, avoiding overfitting to their training environments. Tackling this is vital if we are to deploy reinforcement learning algorithms in real world scenarios, where the environment will be diverse, dynamic and unpredictable. This survey is an overview of this nascent field. We rely on a unifying formalism and terminology for discussing different ZSG problems, building upon previous works. We go on to categorise existing benchmarks for ZSG, as well as current methods for tackling these problems. Finally, we provide a critical discussion of the current state of the field, including recommendations for future work. Among other conclusions, we argue that taking a purely procedural content generation approach to benchmark design is not conducive to progress in ZSG, we suggest fast online adaptation and tackling RL-specific problems as some areas for future work on methods for ZSG, and we recommend building benchmarks in underexplored problem settings such as offline RL ZSG and reward-function variation.

Training generally-capable agents with reinforcement learning (RL) remains a significant challenge. A promising avenue for improving the robustness of RL agents is through the use of curricula. One such class of methods frames environment design as a game between a student and a teacher, using regret-based objectives to produce environment instantiations (or levels) at the frontier of the student agent’s capabilities. These methods benefit from theoretical robustness guarantees at equilibrium, yet they often struggle to find effective levels in challenging design spaces in practice. By contrast, evolutionary approaches incrementally alter environment complexity, resulting in potentially open-ended learning, but often rely on domain-specific heuristics and vast amounts of computational resources. This work proposes harnessing the power of evolution in a principled, regret-based curriculum. Our approach, which we call Adversarially Compounding Complexity by Editing Levels (ACCEL), seeks to constantly produce levels at the frontier of an agent’s capabilities, resulting in curricula that start simple but become increasingly complex. ACCEL maintains the theoretical benefits of prior regret-based methods, while providing significant empirical gains in a diverse set of environments. An interactive version of this paper is available at https://accelagent.github.io.

Progress in reinforcement learning (RL) research is often driven by the design of new, challenging environments---a costly undertaking requiring skills orthogonal to that of a typical machine learning researcher. The complexity of environment development has only increased with the rise of procedural-content generation (PCG) as the prevailing paradigm for producing varied environments capable of testing the robustness and generalization of RL agents. Moreover, existing environments often require complex build processes, making reproducing results difficult. To address these issues, we introduce GriddlyJS, a web-based Integrated Development Environment (IDE) based on the Griddly engine. GriddlyJS allows researchers to easily design and debug arbitrary, complex PCG grid-world environments, as well as visualize, evaluate, and record the performance of trained agent models. By connecting the RL workflow to the advanced functionality enabled by modern web standards, GriddlyJS allows publishing interactive agent-environment demos that reproduce experimental results directly to the web. To demonstrate the versatility of GriddlyJS, we use it to quickly develop a complex compositional puzzle-solving environment alongside arbitrary human-designed environment configurations and their solutions for use in a automatic curriculum learning and offline RL context. The GriddlyJS IDE is open source and freely available at https://griddly.ai.

Recent breakthroughs in the development of agents to solve challenging sequential decision making problems such as Go, StarCraft, or DOTA, have relied on both simulated environments and large-scale datasets. However, progress on this research has been hindered by the scarcity of open-sourced datasets and the prohibitive computational cost to work with them. Here we present the NetHack Learning Dataset (NLD), a large and highly-scalable dataset of trajectories from the popular game of NetHack, which is both extremely challenging for current methods and very fast to run. NLD consists of three parts: 10 billion state transitions from 1.5 million human trajectories collected on the NAO public NetHack server from 2009 to 2020; 3 billion state-action-score transitions from 100,000 trajectories collected from the symbolic bot winner of the NetHack Challenge 2021; and, accompanying code for users to record, load and stream any collection of such trajectories in a highly compressed form. We evaluate a wide range of existing algorithms for learning from demonstrations, showing that significant research advances are needed to fully leverage large-scale datasets for challenging sequential decision making tasks.

In recent years, a number of reinforcement learning (RL) methods have been pro- posed to explore complex environments which differ across episodes. In this work, we show that the effectiveness of these methods critically relies on a count-based episodic term in their exploration bonus. As a result, despite their success in relatively simple, noise-free settings, these methods fall short in more realistic scenarios where the state space is vast and prone to noise. To address this limitation, we introduce Exploration via Elliptical Episodic Bonuses (E3B), a new method which extends count-based episodic bonuses to continuous state spaces and encourages an agent to explore states that are diverse under a learned embed- ding within each episode. The embedding is learned using an inverse dynamics model in order to capture controllable aspects of the environment. Our method sets a new state-of-the-art across 16 challenging tasks from the MiniHack suite, without requiring task-specific inductive biases. E3B also outperforms existing methods in reward-free exploration on Habitat, demonstrating that it can scale to high-dimensional pixel-based observations and realistic environments.

Adaptive curricula in reinforcement learning (RL) have proven effective for producing policies robust to discrepancies between the train and test environment. Recently, the Unsupervised Environment Design (UED) framework generalized RL curricula to generating sequences of entire environments, leading to new methods with robust minimax regret properties. Problematically, in partially-observable or stochastic settings, optimal policies may depend on the ground-truth distribution over aleatoric parameters of the environment in the intended deployment setting, while curriculum learning necessarily shifts the training distribution. We formalize this phenomenon as curriculum-induced covariate shift (CICS), and describe how its occurrence in aleatoric parameters can lead to suboptimal policies. Directly sampling these parameters from the ground-truth distribution avoids the issue, but thwarts curriculum learning. We propose SAMPLR, a minimax regret UED method that optimizes the ground-truth utility function, even when the underlying training data is biased due to CICS. We prove, and validate on challenging domains, that our approach preserves optimality under the ground-truth distribution, while promoting robustness across the full range of environment settings.

Reinforcement learning (RL) agents are particularly hard to train when rewards are sparse. One common solution is to use intrinsic rewards to encourage agents to explore their environment. However, recent intrinsic exploration methods often use state-based novelty measures which reward low-level exploration and may not scale to domains requiring more abstract skills. Instead, we explore natural language as a general medium for highlighting relevant abstractions in an environment. Unlike previous work, we evaluate whether language can improve over existing exploration methods by directly extending (and comparing to) competitive intrinsic exploration baselines: AMIGo (Campero et al., 2021) and NovelD (Zhang et al., 2021). These language-based variants outperform their non-linguistic forms by 47-85% across 13 challenging tasks from the MiniGrid and MiniHack environment suites.

Building generally capable agents is a grand challenge for deep reinforcement learning (RL). To approach this challenge practically, we outline two key desiderata: 1) to facilitate generalization, exploration should be task agnostic; 2) to facilitate scalability, exploration policies should collect large quantities of data without costly centralized retraining. Combining these two properties, we introduce the reward-free deployment efficiency setting, a new paradigm for RL research. We then present CASCADE, a novel approach for self-supervised exploration in this new setting. CASCADE seeks to learn a world model by collecting data with a population of agents, using an information theoretic objective inspired by Bayesian Active Learning. CASCADE achieves this by specifically maximizing the diversity of trajectories sampled by the population through a novel cascading objective. We provide theoretical intuition for CASCADE which we show in a tabular setting improves upon naïve approaches that do not account for population diversity. We then demonstrate that CASCADE collects diverse task-agnostic datasets and learns agents that generalize zero-shot to novel, unseen downstream tasks on Atari, MiniGrid, Crafter and the DM Control Suite. Code and videos are available at https://ycxuyingchen.github.io/cascade/

Recent work has shown that augmenting environments with language descriptions improves policy learning. However, for environments with complex language abstractions, learning how to ground language to observations is difficult due to sparse, delayed rewards. We propose Language Dynamics Distillation (LDD), which pretrains a model to predict environment dynamics given demonstrations with language descriptions, and then fine-tunes these language-aware pretrained representations via reinforcement learning (RL). In this way, the model is trained to both maximize expected reward and retain knowledge about how language relates to environment dynamics. On SILG, a benchmark of five tasks with language descriptions that evaluate distinct generalization challenges on unseen environments (NetHack, ALFWorld, RTFM, Messenger, and Touchdown), LDD outperforms tabula-rasa RL, VAE pretraining, and methods that learn from unlabeled demonstrations in inverse RL and reward shaping with pretrained experts. In our analyses, we show that language descriptions in demonstrations improve sample-efficiency and generalization across environments, and that dynamics modeling with expert demonstrations is more effective than with non-experts.

Environments with procedurally generated content serve as important benchmarks for testing systematic generalization in deep reinforcement learning. In this setting, each level is an algorithmically created environment instance with a unique configuration of its factors of variation. Training on a prespecified subset of levels allows for testing generalization to unseen levels. What can be learned from a level depends on the current policy, yet prior work defaults to uniform sampling of training levels independently of the policy. We introduce Prioritized Level Replay (PLR), a general framework for selectively sampling the next training level by prioritizing those with higher estimated learning potential when revisited in the future. We show TD-errors effectively estimate a level’s future learning potential and, when used to guide the sampling procedure, induce an emergent curriculum of increasingly difficult levels. By adapting the sampling of training levels, PLR significantly improves sample-efficiency and generalization on Procgen Benchmark—matching the previous state-of-the-art in test return—and readily combines with other methods. Combined with the previous leading method, PLR raises the state-of-the-art to over 76% improvement in test return relative to standard RL baselines.

Prithviraj Ammanabrolu, Jack Urbanek, Margaret Li, Arthur Szlam, Tim Rocktäschel, Jason Weston. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick Lewis, Majid Yazdani, Nicola De Cao, James Thorne, Yacine Jernite, Vladimir Karpukhin, Jean Maillard, Vassilis Plachouras, Tim Rocktäschel, Sebastian Riedel. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Deep reinforcement learning (RL) agents may successfully generalize to new settings if trained on an appropriately diverse set of environment and task configurations. Unsupervised Environment Design (UED) is a promising self-supervised RL paradigm, wherein the free parameters of an underspecified environment are automatically adapted during training to the agent's capabilities, leading to the emergence of diverse training environments. Here, we cast Prioritized Level Replay (PLR), an empirically successful but theoretically unmotivated method that selectively samples randomly-generated training levels, as UED. We argue that by curating completely random levels, PLR, too, can generate novel and complex levels for effective training. This insight reveals a natural class of UED methods we call Dual Curriculum Design (DCD). Crucially, DCD includes both PLR and a popular UED algorithm, PAIRED, as special cases and inherits similar theoretical guarantees. This connection allows us to develop novel theory for PLR, providing a version with a robustness guarantee at Nash equilibria. Furthermore, our theory suggests a highly counterintuitive improvement to PLR: by stopping the agent from updating its policy on uncurated levels (training on less data), we can improve the convergence to Nash equilibria. Indeed, our experiments confirm that our new method, PLR$^{\perp}$, obtains better results on a suite of out-of-distribution, zero-shot transfer tasks, in addition to demonstrating that PLR$^{\perp}$ improves the performance of PAIRED, from which it inherited its theoretical framework.

Procedurally generating cohesive and interesting game environments is challenging and time-consuming. In order for the relationships between the game elements to be natural, common-sense has to be encoded into arrangement of the elements. In this work, we investigate a machine learning approach for world creation using content from the multi-player text adventure game environment LIGHT (Urbanek et al. 2019). We introduce neural network based models to compositionally arrange locations, characters, and objects into a coherent whole. In addition to creating worlds based on existing elements, our models can generate new game content. Humans can also leverage our models to interactively aid in worldbuilding. We show that the game environments created with our approach are cohesive, diverse, and preferred by human evaluators compared to other machine learning based world construction algorithms.

Reasoning with knowledge expressed in natural language and Knowledge Bases (KBs) is a major challenge for Artificial Intelligence, with applications in machine reading, dialogue, and question answering. General neural architectures that jointly learn representations and transformations of text are very data-inefficient, and it is hard to analyse their reasoning process. These issues are addressed by end-to-end differentiable reasoning systems such as Neural Theorem Provers (NTPs), although they can only be used with small-scale symbolic KBs. In this paper we first propose Greedy NTPs (GNTPs), an extension to NTPs addressing their complexity and scalability limitations, thus making them applicable to real-world datasets. This result is achieved by dynamically constructing the computation graph of NTPs and including only the most promising proof paths during inference, thus obtaining orders of magnitude more efficient models 1. Then, we propose a novel approach for jointly reasoning over KBs and textual mentions, by embedding logic facts and natural language sentences in a shared embedding space. We show that GNTPs perform on par with NTPs at a fraction of their cost while achieving competitive link prediction results on large datasets, providing explanations for predictions, and inducing interpretable models.

Natural Language Inference (NLI) datasets contain annotation artefacts resulting in spurious correlations between the natural language utterances and their respective entailment classes.These artefacts are exploited by neural networks even when only considering the hypothesis and ignoring the premise, leading to unwanted biases.Belinkov et al. (2019b) proposed tackling this problem via adversarial training, but this can lead to learned sentence representations that still suffer from the same biases.We show that the bias can be reduced in the sentence representations by using an ensemble of adversaries, encouraging the model to jointly decrease the accuracy of these different adversaries while fitting the data.This approach produces more robust NLI models, outperforming previous de-biasing efforts when generalised to 12 other NLI datasets (Belinkov et al., 2019a;Mahabadi et al., 2020).In addition, we find that the optimal number of adversarial classifiers depends on the dimensionality of the sentence representations, with larger sentence representations being more difficult to de-bias while benefiting from using a greater number of adversaries.

Attempts to render deep learning models interpretable, data-efficient, and robust have seen some success through hybridisation with rule-based systems, for example, in Neural Theorem Provers (NTPs). These neuro-symbolic models can induce interpretable rules and learn representations from data via back-propagation, while providing logical explanations for their predictions. However, they are restricted by their computational complexity, as they need to consider all possible proof paths for explaining a goal, thus rendering them unfit for large-scale applications. We present Conditional Theorem Provers (CTPs), an extension to NTPs that learns an optimal rule selection strategy via gradient-based optimisation. We show that CTPs are scalable and yield state-of-the-art results on the CLUTRR dataset, which tests systematic generalisation of neural models by learning to reason over smaller graphs and evaluating on larger ones. Finally, CTPs show better link prediction results on standard benchmarks in comparison with other neural-symbolic models, while being explainable.

Progress in Reinforcement Learning (RL) algorithms goes hand-in-hand with the development of challenging environments that test the limits of current methods. While existing RL environments are either sufficiently complex or based on fast simulation, they are rarely both. Here, we present the NetHack Learning Environment (NLE), a scalable, procedurally generated, stochastic, rich, and challenging environment for RL research based on the popular single-player terminal-based roguelike game, NetHack. We argue that NetHack is sufficiently complex to drive long-term research on problems such as exploration, planning, skill acquisition, and language-conditioned RL, while dramatically reducing the computational resources required to gather a large amount of experience. We compare NLE and its task suite to existing alternatives, and discuss why it is an ideal medium for testing the robustness and systematic generalization of RL agents. We demonstrate empirical success for early stages of the game using a distributed Deep RL baseline and Random Network Distillation exploration, alongside qualitative analysis of various agents trained in the environment. NLE is open source and available at https://github.com/facebookresearch/nle.

Large pre-trained language models have been shown to store factual knowledge in their parameters, and achieve state-of-the-art results when fine-tuned on downstream NLP tasks. However, their ability to access and precisely manipulate knowledge is still limited, and hence on knowledge-intensive tasks, their performance lags behind task-specific architectures. Additionally, providing provenance for their decisions and updating their world knowledge remain open research problems. Pre-trained models with a differentiable access mechanism to explicit non-parametric memory can overcome this issue, but have so far been only investigated for extractive downstream tasks. We explore a general-purpose fine-tuning recipe for retrieval-augmented generation (RAG) -- models which combine pre-trained parametric and non-parametric memory for language generation. We introduce RAG models where the parametric memory is a pre-trained seq2seq model and the non-parametric memory is a dense vector index of Wikipedia, accessed with a pre-trained neural retriever. We compare two RAG formulations, one which conditions on the same retrieved passages across the whole generated sequence, the other can use different passages per token. We fine-tune and evaluate our models on a wide range of knowledge-intensive NLP tasks and set the state-of-the-art on three open domain QA tasks, outperforming parametric seq2seq models and task-specific retrieve-and-extract architectures. For language generation tasks, we find that RAG models generate more specific, diverse and factual language than a state-of-the-art parametric-only seq2seq baseline.

MOTIVATION: Several recent studies showed that the application of deep neural networks advanced the state-of-the-art in named entity recognition (NER), including biomedical NER. However, the impact on performance and the robustness of improvements crucially depends on the availability of sufficiently large training corpora, which is a problem in the biomedical domain with its often rather small gold standard corpora. RESULTS: We evaluate different methods for alleviating the data sparsity problem by pretraining a deep neural network (LSTM-CRF), followed by a rather short fine-tuning phase focusing on a particular corpus. Experiments were performed using 34 different corpora covering five different biomedical entity types, yielding an average increase in F1-score of ∼2 pp compared to learning without pretraining. We experimented both with supervised and semi-supervised pretraining, leading to interesting insights into the precision/recall trade-off. Based on our results, we created the stand-alone NER tool HUNER incorporating fully trained models for five entity types. On the independent CRAFT corpus, which was not used for creating HUNER, it outperforms the state-of-the-art tools GNormPlus and tmChem by 5-13 pp on the entity types chemicals, species and genes. AVAILABILITY AND IMPLEMENTATION: HUNER is freely available at https://hu-ner.github.io. HUNER comes in containers, making it easy to install and use, and it can be applied off-the-shelf to arbitrary texts. We also provide an integrated tool for obtaining and converting all 34 corpora used in our evaluation, including fixed training, development and test splits to enable fair comparisons in the future. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Rule-based models are attractive for various tasks because they inherently lead to interpretable and explainable decisions and can easily incorporate prior knowledge.However, such systems are difficult to apply to problems involving natural language, due to its linguistic variability.In contrast, neural models can cope very well with ambiguity by learning distributed representations of words and their composition from data, but lead to models that are difficult to interpret.In this paper, we describe a model combining neural networks with logic programming in a novel manner for solving multi-hop reasoning tasks over natural language.Specifically, we propose to use a Prolog prover which we extend to utilize a similarity function over pretrained sentence encoders.We fine-tune the representations for the similarity function via backpropagation.This leads to a system that can apply rulebased reasoning to natural language, and induce domain-specific rules from training data.We evaluate the proposed system on two different question answering tasks, showing that it outperforms two baselines -BIDAF (Seo et al., 2016a) and FASTQA (Weissenborn et al., 2017b) on a subset of the WIKIHOP corpus and achieves competitive results on the MEDHOP data set (Welbl et al., 2017).

Fabio Petroni, Tim Rocktäschel, Sebastian Riedel, Patrick Lewis, Anton Bakhtin, Yuxiang Wu, Alexander Miller. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Jack Urbanek, Angela Fan, Siddharth Karamcheti, Saachi Jain, Samuel Humeau, Emily Dinan, Tim Rocktäschel, Douwe Kiela, Arthur Szlam, Jason Weston. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

By enabling correct differentiation in Stochastic Computation Graphs (SCGs), the infinitely differentiable Monte-Carlo estimator (DiCE) can generate correct estimates for the higher order gradients that arise in, e.g., multi-agent reinforcement learning and meta-learning. However, the baseline term in DiCE that serves as a control variate for reducing variance applies only to first order gradient estimation, limiting the utility of higher-order gradient estimates. To improve the sample efficiency of DiCE, we propose a new baseline term for higher order gradient estimation. This term may be easily included in the objective, and produces unbiased variance-reduced estimators under (automatic) differentiation, without affecting the estimate of the objective itself or of the first order gradient estimate. It reuses the same baseline function (e.g., the state-value function in reinforcement learning) already used for the first order baseline. We provide theoretical analysis and numerical evaluations of this new baseline, which demonstrate that it can dramatically reduce the variance of DiCE’s second order gradient estimators and also show empirically that it reduces the variance of third and fourth order gradients. This computational tool can be easily used to estimate higher order gradients with unprecedented efficiency and simplicity wherever automatic differentiation is utilised, and it has the potential to unlock applications of higher order gradients in reinforcement learning and meta-learning.

To be successful in real-world tasks, Reinforcement Learning (RL) needs to exploit the compositional, relational, and hierarchical structure of the world, and learn to transfer it to the task at hand. Recent advances in representation learning for language make it possible to build models that acquire world knowledge from text corpora and integrate this knowledge into downstream decision making problems. We thus argue that the time is right to investigate a tight integration of natural language understanding into RL in particular. We survey the state of the field, including work on instruction following, text games, and learning from textual domain knowledge. Finally, we call for the development of new environments as well as further investigation into the potential uses of recent Natural Language Processing (NLP) techniques for such tasks.

Marzieh Saeidi, Max Bartolo, Patrick Lewis, Sameer Singh, Tim Rocktäschel, Mike Sheldon, Guillaume Bouchard, Sebastian Riedel. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018.

The score function estimator is widely used for estimating gradients of stochastic objectives in stochastic computation graphs (SCG), eg., in reinforcement learning and meta-learning. While deriving the first-order gradient estimators by differentiating a surrogate loss (SL) objective is computationally and conceptually simple, using the same approach for higher-order derivatives is more challenging. Firstly, analytically deriving and implementing such estimators is laborious and not compliant with automatic differentiation. Secondly, repeatedly applying SL to construct new objectives for each order derivative involves increasingly cumbersome graph manipulations. Lastly, to match the first-order gradient under differentiation, SL treats part of the cost as a fixed sample, which we show leads to missing and wrong terms for estimators of higher-order derivatives. To address all these shortcomings in a unified way, we introduce DiCE, which provides a single objective that can be differentiated repeatedly, generating correct estimators of derivatives of any order in SCGs. Unlike SL, DiCE relies on automatic differentiation for performing the requisite graph manipulations. We verify the correctness of DiCE both through a proof and numerical evaluation of the DiCE derivative estimates. We also use DiCE to propose and evaluate a novel approach for multi-agent learning. Our code is available at https://github.com/alshedivat/lola.

In order for machine learning to garner widespread public adoption, models must be able to provide interpretable and robust explanations for their decisions, as well as learn from human-provided explanations at train time. In this work, we extend the Stanford Natural Language Inference dataset with an additional layer of human-annotated natural language explanations of the entailment relations. We further implement models that incorporate these explanations into their training process and output them at test time. We show how our corpus of explanations, which we call e-SNLI, can be used for various goals, such as obtaining full sentence justifications of a model’s decisions, improving universal sentence representations and transferring to out-of-domain NLI datasets. Our dataset thus opens up a range of research directions for using natural language explanations, both for improving models and for asserting their trust

Given that in practice training data is scarce for all but a small set of problems, a core question is how to incorporate prior knowledge into a model. In this paper, we consider the case of prior procedural knowledge for neural networks, such as knowing how a program should traverse a sequence, but not what local actions should be performed at each step. To this end, we present an end-to-end differentiable interpreter for the programming language Forth which enables programmers to write program sketches with slots that can be filled with behaviour trained from program input-output data. We can optimise this behaviour directly through gradient descent techniques on user-specified objectives, and also integrate the program into any larger neural computation graph. We show empirically that our interpreter is able to effectively leverage different levels of prior program structure and learn complex behaviours such as sequence sorting and addition. When connected to outputs of an LSTM and trained jointly, our interpreter achieves state-of-the-art accuracy for end-to-end reasoning about quantities expressed in natural language stories.

We introduce deep neural networks for end-to-end differentiable theorem proving that operate on dense vector representations of symbols. These neural networks are recursively constructed by following the backward chaining algorithm as used in Prolog. Specifically, we replace symbolic unification with a differentiable computation on vector representations of symbols using a radial basis function kernel, thereby combining symbolic reasoning with learning subsymbolic vector representations. The resulting neural network can be trained to infer facts from a given incomplete knowledge base using gradient descent. By doing so, it learns to (i) place representations of similar symbols in close proximity in a vector space, (ii) make use of such similarities to prove facts, (iii) induce logical rules, and (iv) it can use provided and induced logical rules for complex multi-hop reasoning. On four benchmark knowledge bases we demonstrate that this architecture outperforms ComplEx, a state-of-the-art neural link prediction model, while at the same time inducing interpretable function-free first-order logic rules.

Stance detection is the task of classifying the attitude expressed in a text towards a target such as Hillary Clinton to be "positive", negative" or "neutral". Previous work has assumed that either the target is mentioned in the text or that training data for every target is given. This paper considers the more challenging version of this task, where targets are not always mentioned and no training data is available for the test targets. We experiment with conditional LSTM encoding, which builds a representation of the tweet that is dependent on the target, and demonstrate that it outperforms encoding the tweet and the target independently. Performance is improved further when the conditional model is augmented with bidirectional encoding. We evaluate our approach on the SemEval 2016 Task 6 Twitter Stance Detection corpus achieving performance second best only to a system trained on semi-automatically labelled tweets for the test target. When such weak supervision is added, our approach achieves state-of-the-art results.

Methods based on representation learning currently hold the state-of-the-art in many natural language processing and knowledge base inference tasks.Yet, a major challenge is how to efficiently incorporate commonsense knowledge into such models.A recent approach regularizes relation and entity representations by propositionalization of first-order logic rules.However, propositionalization does not scale beyond domains with only few entities and rules.In this paper we present a highly efficient method for incorporating implication rules into distributed representations for automated knowledge base construction.We map entity-tuple embeddings into an approximately Boolean space and encourage a partial ordering over relation embeddings based on implication rules mined from WordNet.Surprisingly, we find that the strong restriction of the entity-tuple embedding space does not hurt the expressiveness of the model and even acts as a regularizer that improves generalization.By incorporating few commonsense rules, we achieve an increase of 2 percentage points mean average precision over a matrix factorization baseline, while observing a negligible increase in runtime.

UNLABELLED: : Descriptions of genetic variations and their effect are widely spread across the biomedical literature. However, finding all mentions of a specific variation, or all mentions of variations in a specific gene, is difficult to achieve due to the many ways such variations are described. Here, we describe SETH, a tool for the recognition of variations from text and their subsequent normalization to dbSNP or UniProt. SETH achieves high precision and recall on several evaluation corpora of PubMed abstracts. It is freely available and encompasses stand-alone scripts for isolated application and evaluation as well as a thorough documentation for integration into other applications. AVAILABILITY AND IMPLEMENTATION: SETH is released under the Apache 2.0 license and can be downloaded from http://rockt.github.io/SETH/ CONTACT: [email protected] or [email protected].

Matrix factorization approaches to relation extraction provide several attractive features: they support distant supervision, handle open schemas, and leverage unlabeled data.Unfortunately, these methods share a shortcoming with all other distantly supervised approaches: they cannot learn to extract target relations without existing data in the knowledge base, and likewise, these models are inaccurate for relations with sparse data.Rule-based extractors, on the other hand, can be easily extended to novel relations and improved for existing but inaccurate relations, through first-order formulae that capture auxiliary domain knowledge.However, usually a large set of such formulae is necessary to achieve generalization.In this paper, we introduce a paradigm for learning low-dimensional embeddings of entity-pairs and relations that combine the advantages of matrix factorization with first-order logic domain knowledge.We introduce simple approaches for estimating such embeddings, as well as a novel training algorithm to jointly optimize over factual and first-order logic information.Our results show that this method is able to learn accurate extractors with little or no distant supervision alignments, while at the same time generalizing to textual patterns that do not appear in the formulae.

Sameer Singh, Tim Rocktäschel, Luke Hewitt, Jason Naradowsky, Sebastian Riedel. Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations. 2015.

MOTIVATION: The accurate identification of chemicals in text is important for many applications, including computer-assisted reconstruction of metabolic networks or retrieval of information about substances in drug development. But due to the diversity of naming conventions and traditions for such molecules, this task is highly complex and should be supported by computational tools. RESULTS: We present ChemSpot, a named entity recognition (NER) tool for identifying mentions of chemicals in natural language texts, including trivial names, drugs, abbreviations, molecular formulas and International Union of Pure and Applied Chemistry entities. Since the different classes of relevant entities have rather different naming characteristics, ChemSpot uses a hybrid approach combining a Conditional Random Field with a dictionary. It achieves an F(1) measure of 68.1% on the SCAI corpus, outperforming the only other freely available chemical NER tool, OSCAR4, by 10.8 percentage points. AVAILABILITY: ChemSpot is freely available at: http://www.informatik.hu-berlin.de/wbi/resources.