Xingdi Yuan

Ziang Xiao, Xingdi Yuan, Q. Vera Liao et al. · microsoft-research

Qualitative analysis of textual contents unpacks rich and valuable information by assigning labels to the data. However, this process is often labor-intensive, particularly when working with large datasets. While recent AI-based tools demonstrate utility, researchers may not have readily available AI resources and expertise, let alone be challenged by the limited generalizability of those task-specific models. In this study, we explored the use of large language models (LLMs) in supporting deductive coding, a major category of qualitative analysis where researchers use pre-determined codebooks to label the data into a fixed set of codes. Instead of training task-specific models, a pre-trained LLM could be used directly for various tasks without fine-tuning through prompt learning. Using a curiosity-driven questions coding task as a case study, we found, by combining GPT-3 with expert-drafted codebooks, our proposed approach achieved fair to substantial agreements with expert-coded results. We lay out challenges and opportunities in using LLMs to support qualitative coding and beyond.

Rui Meng, Tong Wang, Xingdi Yuan et al. · microsoft-research

Training keyphrase generation (KPG) models require a large amount of annotated data, which can be prohibitively expensive and often limited to specific domains. In this study, we first demonstrate that large distribution shifts among different domains severely hinder the transferability of KPG models. We then propose a three-stage pipeline, which gradually guides KPG models' learning focus from general syntactical features to domain-related semantics, in a data-efficient manner. With Domain-general Phrase pre-training, we pre-train Sequence-to-Sequence models with generic phrase annotations that are widely available on the web, which enables the models to generate phrases in a wide range of domains. The resulting model is then applied in the Transfer Labeling stage to produce domain-specific pseudo keyphrases, which help adapt models to a new domain. Finally, we fine-tune the model with limited data with true labels to fully adapt it to the target domain. Our experiment results show that the proposed process can produce good-quality keyphrases in new domains and achieve consistent improvements after adaptation with limited in-domain annotated data. All code and datasets are available at https://github.com/memray/OpenNMT-kpg-release.

Rania Abdelghani, Yen-Hsiang Wang, Xingdi Yuan et al. · microsoft-research

In order to train children's ability to ask curiosity-driven questions, previous research has explored designing specific exercises relying on providing semantic and linguistic cues to help formulate such questions. But despite showing pedagogical efficiency, this method is still limited as it relies on generating the said cues by hand, which can be a very costly process. In this context, we propose to leverage advances in the natural language processing field (NLP) and investigate the efficiency of using a large language model (LLM) for automating the production of the pedagogical content of a curious question-asking (QA) training. We study generating the said content using the "prompt-based" method that consists of explaining the task to the LLM in natural text. We evaluate the output using human experts annotations and comparisons with hand-generated content. Results suggested indeed the relevance and usefulness of this content. We also conduct a field study in primary school (75 children aged 9-10), where we evaluate children's QA performance when having this training. We compare 3 types of content : 1) hand-generated content that proposes "closed" cues leading to predefined questions; 2) GPT-3-generated content that proposes the same type of cues; 3) GPT-3-generated content that proposes "open" cues leading to several possible questions. We see a similar QA performance between the two "closed" trainings (showing the scalability of the approach using GPT-3), and a better one for participants with the "open" training. These results suggest the efficiency of using LLMs to support children in generating more curious questions, using a natural language prompting approach that affords usability by teachers and other users not specialists of AI techniques. Furthermore, results also show that open-ended content may be more suitable for training curious question-asking skills.

Xingdi Yuan, Tong Wang, Yen-Hsiang Wang et al. · microsoft-research

Large Language Models (LLMs) have in recent years demonstrated impressive prowess in natural language generation. A common practice to improve generation diversity is to sample multiple outputs from the model. However, there lacks a simple and robust way of selecting the best output from these stochastic samples. As a case study framed in the context of question generation, we propose two prompt-based approaches to selecting high-quality questions from a set of LLM-generated candidates. Our method works under the constraints of 1) a black-box (non-modifiable) question generation model and 2) lack of access to human-annotated references -- both of which are realistic limitations for real-world deployment of LLMs. With automatic as well as human evaluations, we empirically demonstrate that our approach can effectively select questions of higher qualities than greedy generation.

Iou-Jen Liu, Xingdi Yuan, Marc-Alexandre Côté et al. · microsoft-research

To solve difficult tasks, humans ask questions to acquire knowledge from external sources. In contrast, classical reinforcement learning agents lack such an ability and often resort to exploratory behavior. This is exacerbated as few present-day environments support querying for knowledge. In order to study how agents can be taught to query external knowledge via language, we first introduce two new environments: the grid-world-based Q-BabyAI and the text-based Q-TextWorld. In addition to physical interactions, an agent can query an external knowledge source specialized for these environments to gather information. Second, we propose the "Asking for Knowledge" (AFK) agent, which learns to generate language commands to query for meaningful knowledge that helps solve the tasks. AFK leverages a non-parametric memory, a pointer mechanism and an episodic exploration bonus to tackle (1) irrelevant information, (2) a large query language space, (3) delayed reward for making meaningful queries. Extensive experiments demonstrate that the AFK agent outperforms recent baselines on the challenging Q-BabyAI and Q-TextWorld environments.

Alessandro Sordoni, Xingdi Yuan, Marc-Alexandre Côté et al. · microsoft-research

Large language models (LLMs) can be seen as atomic units of computation mapping sequences to a distribution over sequences. Thus, they can be seen as stochastic language layers in a language network, where the learnable parameters are the natural language prompts at each layer. By stacking two such layers and feeding the output of one layer to the next, we obtain a Deep Language Network (DLN). We first show how to effectively perform prompt optimization for a 1-Layer language network (DLN-1). Then, we present an extension that applies to 2-layer DLNs (DLN-2), where two prompts must be learned. The key idea is to consider the output of the first layer as a latent variable, which requires inference, and prompts to be learned as the parameters of the generative distribution. We first test the effectiveness of DLN-1 in multiple reasoning and natural language understanding tasks. Then, we show that DLN-2 can reach higher performance than a single layer, showing promise that we might reach comparable performance to GPT-4, even when each LLM in the network is smaller and less powerful.

Nathaniel Weir, Xingdi Yuan, Marc-Alexandre Côté et al. · microsoft-research

Humans have the capability, aided by the expressive compositionality of their language, to learn quickly by demonstration. They are able to describe unseen task-performing procedures and generalize their execution to other contexts. In this work, we introduce DescribeWorld, an environment designed to test this sort of generalization skill in grounded agents, where tasks are linguistically and procedurally composed of elementary concepts. The agent observes a single task demonstration in a Minecraft-like grid world, and is then asked to carry out the same task in a new map. To enable such a level of generalization, we propose a neural agent infused with hierarchical latent language--both at the level of task inference and subtask planning. Our agent first generates a textual description of the demonstrated unseen task, then leverages this description to replicate it. Through multiple evaluation scenarios and a suite of generalization tests, we find that agents that perform text-based inference are better equipped for the challenge under a random split of tasks.

Laetitia Teodorescu, Xingdi Yuan, Marc-Alexandre Côté et al. · microsoft-research

Building open-ended agents that can autonomously discover a diversity of behaviours is one of the long-standing goals of artificial intelligence. This challenge can be studied in the framework of autotelic RL agents, i.e. agents that learn by selecting and pursuing their own goals, self-organizing a learning curriculum. Recent work identified language as a key dimension of autotelic learning, in particular because it enables abstract goal sampling and guidance from social peers for hindsight relabelling. Within this perspective, we study the following open scientific questions: What is the impact of hindsight feedback from a social peer (e.g. selective vs. exhaustive)? How can the agent learn from very rare language goal examples in its experience replay? How can multiple forms of exploration be combined, and take advantage of easier goals as stepping stones to reach harder ones? To address these questions, we use ScienceWorld, a textual environment with rich abstract and combinatorial physics. We show the importance of selectivity from the social peer's feedback; that experience replay needs to over-sample examples of rare goals; and that following self-generated goal sequences where the agent's competence is intermediate leads to significant improvements in final performance.

Zhiyuan Sun, Haochen Shi, Marc-Alexandre Côté et al. · microsoft-research

Large language models (LLMs) have been increasingly applied to tasks in language understanding and interactive decision-making, with their impressive performance largely attributed to the extensive domain knowledge embedded within them. However, the depth and breadth of this knowledge can vary across domains. Many existing approaches assume that LLMs possess a comprehensive understanding of their environment, often overlooking potential gaps in their grasp of actual world dynamics. To address this, we introduce Discover, Verify, and Evolve (DiVE), a framework that discovers world dynamics from a small number of demonstrations, verifies the accuracy of these dynamics, and evolves new, advanced dynamics tailored to the current situation. Through extensive evaluations, we assess the impact of each component on performance and compare the dynamics generated by DiVE to human-annotated dynamics. Our results show that LLMs guided by DiVE make more informed decisions, achieving rewards comparable to human players in the Crafter environment and surpassing methods that require prior task-specific training in the MiniHack environment.

Xinyi Wang, Lucas Caccia, Oleksiy Ostapenko et al.

Large language models (LLMs) have recently attracted considerable interest for their ability to perform complex reasoning tasks, such as chain-of-thought (CoT) reasoning. However, most of the existing approaches to enhance this ability rely heavily on data-driven methods, while neglecting the structural aspects of the model's reasoning capacity. To encourage a more structural generation of CoT steps, we propose a hierarchical generation scheme: we let the LM generate a planning token at the start of each reasoning step, intuitively serving as a high-level plan of the current step, and add their embeddings to the model parameters. Our approach requires a negligible increase in trainable parameters (0.001%) and can be applied through either full fine-tuning or a more parameter-efficient scheme. We demonstrate our method's effectiveness by applying it to three different LLMs, showing notable accuracy improvements across three math word problem datasets and one multihop QA dataset with respect to standard fine-tuning baselines.

Baolin Peng, Wenlin Yao, Qianhui Wu et al.

Agentic modeling aims to transform LLMs into autonomous agents capable of solving complex tasks through planning, reasoning, tool use, and multi-turn interaction with environments. Despite major investment, open research remains constrained by infrastructure and training gaps. Many high-performing systems rely on proprietary codebases, models, or services, while most open-source frameworks focus on orchestration and evaluation rather than scalable agent training. We present Orchard, an open-source framework for scalable agentic modeling. At its core is Orchard Env, a lightweight environment service providing reusable primitives for sandbox lifecycle management across task domains, agent harnesses, and pipeline stages. On top of Orchard Env, we build three agentic modeling recipes. Orchard-SWE targets coding agents. We distill 107K trajectories from MiniMax-M2.5 and Qwen3.5-397B, introduce credit-assignment SFT to learn from productive segments of unresolved trajectories, and apply Balanced Adaptive Rollout for RL. Starting from Qwen3-30B-A3B-Thinking, Orchard-SWE achieves 64.3% on SWE-bench Verified after SFT and 67.5% after SFT+RL, setting a new state of the art among open-source models of comparable size. Orchard-GUI trains a 4B vision-language computer-use agent using only 0.4K distilled trajectories and 2.2K open-ended tasks. It achieves 74.1%, 67.0%, and 64.0% success rates on WebVoyager, Online-Mind2Web, and DeepShop, respectively, making it the strongest open-source model while remaining competitive with proprietary systems. Orchard-Claw targets personal assistant agents. Trained with only 0.2K synthetic tasks, it achieves 59.6% pass@3 on Claw-Eval and 73.9% when paired with a stronger ZeroClaw harness. Collectively, these results show that a lightweight, open, harness-agnostic environment layer enables reusable agentic data, training recipes, and evaluations across domains.

Haochen Shi, Xingdi Yuan, Bang Liu

We study continual skill acquisition in open-ended embodied environments where an agent must construct, refine, and reuse an expanding library of executable skills. We introduce the Programmatic Skill Network (PSN), a framework in which skills are executable symbolic programs forming a compositional network that evolves through experience. PSN defines three core mechanisms instantiated via large language models: (1)REFLECT for structured fault localization over skill compositions, (2) progressive optimization with maturity-aware update gating that stabilizes reliable skills while maintaining plasticity for uncertain ones, and (3) canonical structural refactoring under rollback validation that maintains network compactness. We further show that PSN's learning dynamics exhibit structural parallels to neural network training. Experiments on MineDojo and Crafter demonstrate robust skill reuse, rapid adaptation, and strong generalization across open-ended task distributions.\footnote{We plan to open-source the code.

Minseon Kim, Lucas Caccia, Zhengyan Shi et al.

User prompts to large language models (LLMs) are often ambiguous or under-specified, and subtle contextual cues shaped by user intentions, prior knowledge, and risk factors strongly influence what constitutes an appropriate response. Misinterpreting intent or risks may lead to unsafe outputs, while overly cautious interpretations can cause unnecessary refusal of benign requests. In this paper, we question the conventional framework in which LLMs generate immediate responses to requests without considering broader contextual factors. User requests are situated within broader contexts such as intentions, knowledge, and prior experience, which strongly influence what constitutes an appropriate answer. We propose a framework that extracts and leverages such contextual information from the user prompt itself. Specifically, a reinforcement learning based context generator, designed in an autoencoder-like fashion, is trained to infer contextual signals grounded in the prompt and use them to guide response generation. This approach is particularly important for safety tasks, where ambiguous requests may bypass safeguards while benign but confusing requests can trigger unnecessary refusals. Experiments show that our method reduces harmful responses by an average of 5.6% on the SafetyInstruct dataset across multiple foundation models and improves the harmonic mean of attack success rate and compliance on benign prompts by 6.2% on XSTest and WildJailbreak. These results demonstrate the effectiveness of context extraction for safer and more reliable LLM inferences.

Hyunji Lee, Minseon Kim, Chinmay Singh et al.

As coding agents are increasingly deployed in large codebases, the need to automatically design challenging, codebase-level evaluation is central. We propose Gistify, a task where a coding LLM must create a single, minimal, self-contained file that can reproduce a specific functionality of a codebase. The coding LLM is given full access to a codebase along with a specific entrypoint (e.g., a python command), and the generated file must replicate the output of the same command ran under the full codebase, while containing only the essential components necessary to execute the provided command. Success on Gistify requires both structural understanding of the codebase, accurate modeling of its execution flow as well as the ability to produce potentially large code patches. Our findings show that current state-of-the-art models struggle to reliably solve Gistify tasks, especially ones with long executions traces.

Weijia Xu, Alessandro Sordoni, Chandan Singh et al.

We introduce EvoLib, a test-time learning framework that enables large language models to accumulate, reuse, and evolve knowledge across problem instances without parameter updates or external supervision. Instead of adapting model parameters, our approach maintains a shared library of knowledge abstractions, including modular skills and reflective insights, automatically extracted from the model's own inference trajectories. To support continual improvement, we introduce a principled weighting and consolidation mechanism that jointly optimizes for immediate utility and long-term value. This allows simple, instance-specific abstractions to evolve into more general and reusable ones over time. Across challenging benchmarks in mathematical reasoning, code generation, and multi-turn agentic environments, EvoLib improves substantially over the top test-time scaling and learning methods without ground-truth feedback.

Arian Hosseini, Xingdi Yuan, Nikolay Malkin et al. · deepmind

Common self-improvement approaches for large language models (LLMs), such as STaR, iteratively fine-tune LLMs on self-generated solutions to improve their problem-solving ability. However, these approaches discard the large amounts of incorrect solutions generated during this process, potentially neglecting valuable information in such solutions. To address this shortcoming, we propose V-STaR that utilizes both the correct and incorrect solutions generated during the self-improvement process to train a verifier using DPO that judges correctness of model-generated solutions. This verifier is used at inference time to select one solution among many candidate solutions. Running V-STaR for multiple iterations results in progressively better reasoners and verifiers, delivering a 4% to 17% test accuracy improvement over existing self-improvement and verification approaches on common code generation and math reasoning benchmarks with LLaMA2 models.

Jie Fu, Xue Geng, Zhijian Duan et al.

Knowledge Distillation (KD) is a common method for transferring the ``knowledge'' learned by one machine learning model (the \textit{teacher}) into another model (the \textit{student}), where typically, the teacher has a greater capacity (e.g., more parameters or higher bit-widths). To our knowledge, existing methods overlook the fact that although the student absorbs extra knowledge from the teacher, both models share the same input data -- and this data is the only medium by which the teacher's knowledge can be demonstrated. Due to the difference in model capacities, the student may not benefit fully from the same data points on which the teacher is trained. On the other hand, a human teacher may demonstrate a piece of knowledge with individualized examples adapted to a particular student, for instance, in terms of her cultural background and interests. Inspired by this behavior, we design data augmentation agents with distinct roles to facilitate knowledge distillation. Our data augmentation agents generate distinct training data for the teacher and student, respectively. We find empirically that specially tailored data points enable the teacher's knowledge to be demonstrated more effectively to the student. We compare our approach with existing KD methods on training popular neural architectures and demonstrate that role-wise data augmentation improves the effectiveness of KD over strong prior approaches. The code for reproducing our results can be found at https://github.com/bigaidream-projects/role-kd

Haochen Shi, Zhiyuan Sun, Xingdi Yuan et al. · microsoft-research

Embodied Instruction Following (EIF) is a crucial task in embodied learning, requiring agents to interact with their environment through egocentric observations to fulfill natural language instructions. Recent advancements have seen a surge in employing large language models (LLMs) within a framework-centric approach to enhance performance in embodied learning tasks, including EIF. Despite these efforts, there exists a lack of a unified understanding regarding the impact of various components-ranging from visual perception to action execution-on task performance. To address this gap, we introduce OPEx, a comprehensive framework that delineates the core components essential for solving embodied learning tasks: Observer, Planner, and Executor. Through extensive evaluations, we provide a deep analysis of how each component influences EIF task performance. Furthermore, we innovate within this space by deploying a multi-agent dialogue strategy on a TextWorld counterpart, further enhancing task performance. Our findings reveal that LLM-centric design markedly improves EIF outcomes, identify visual perception and low-level action execution as critical bottlenecks, and demonstrate that augmenting LLMs with a multi-agent framework further elevates performance.

Victor Zhong, Dipendra Misra, Xingdi Yuan et al. · microsoft-research

We introduce Language Feedback Models (LFMs) that identify desirable behaviour - actions that help achieve tasks specified in the instruction - for imitation learning in instruction following. To train LFMs, we obtain feedback from Large Language Models (LLMs) on visual trajectories verbalized to language descriptions. First, by using LFMs to identify desirable behaviour to imitate, we improve in task-completion rate over strong behavioural cloning baselines on three distinct language grounding environments (Touchdown, ScienceWorld, and ALFWorld). Second, LFMs outperform using LLMs as experts to directly predict actions, when controlling for the number of LLM output tokens. Third, LFMs generalize to unseen environments, improving task-completion rate by 3.5-12.0% through one round of adaptation. Finally, LFM can be modified to provide human-interpretable feedback without performance loss, allowing human verification of desirable behaviour for imitation learning.

Haotian Fu, Pratyusha Sharma, Elias Stengel-Eskin et al. · microsoft-research

We present an algorithm for skill discovery from expert demonstrations. The algorithm first utilizes Large Language Models (LLMs) to propose an initial segmentation of the trajectories. Following that, a hierarchical variational inference framework incorporates the LLM-generated segmentation information to discover reusable skills by merging trajectory segments. To further control the trade-off between compression and reusability, we introduce a novel auxiliary objective based on the Minimum Description Length principle that helps guide this skill discovery process. Our results demonstrate that agents equipped with our method are able to discover skills that help accelerate learning and outperform baseline skill learning approaches on new long-horizon tasks in BabyAI, a grid world navigation environment, as well as ALFRED, a household simulation environment.

Christopher Zhang Cui, Xingdi Yuan, Ziang Xiao et al. · microsoft-research

Reasoning is an essential skill to enable Large Language Models (LLMs) to interact with the world. As tasks become more complex, they demand increasingly sophisticated and diverse reasoning capabilities for sequential decision-making, requiring structured reasoning over the context history to determine the next best action. We introduce TALES, a diverse collection of synthetic and human-written text-adventure games designed to challenge and evaluate diverse reasoning capabilities. We present results over a range of LLMs, open- and closed-weights, performing a qualitative analysis on the top performing models. Despite an impressive showing on synthetic games, even the top LLM-driven agents fail to achieve 15% on games designed for human enjoyment. Code and visualization of the experiments can be found at https://microsoft.github.io/tale-suite.

Xingdi Yuan, Morgane M Moss, Charbel El Feghali et al. · microsoft-research

Large Language Models (LLMs) are increasingly relied upon for coding tasks, yet in most scenarios it is assumed that all relevant information can be either accessed in context or matches their training data. We posit that LLMs can benefit from the ability to interactively explore a codebase to gather the information relevant to their task. To achieve this, we present a textual environment, namely debug-gym, for developing LLM-based agents in an interactive coding setting. Our environment is lightweight and provides a preset of useful tools, such as a Python debugger (pdb), designed to facilitate an LLM-based agent's interactive debugging. Beyond coding and debugging tasks, this approach can be generalized to other tasks that would benefit from information-seeking behavior by an LLM agent.

Atharv Sonwane, Isadora White, Hyunji Lee et al.

High quality bugs are key to training the next generation of language model based software engineering (SWE) agents. We introduce a novel method for synthetic generation of difficult and diverse bugs. Our method instructs SWE Agents to introduce a feature into the codebase whereby they may unintentionally break tests, resulting in bugs. Prior approaches often induce an out-of-distribution effect by generating bugs intentionally (e.g. by introducing local perturbation to existing code), which does not reflect realistic development processes. We perform qualitative analysis to demonstrate that our approach for generating bugs more closely reflects the patterns found in human-authored edits. Through extensive experiments, we demonstrate that our bugs provide more efficient training data for supervised fine-tuning, outperforming other bug datasets by 2% with half the training data (1.2k vs. 3k bugs). We train on our newly generated bugs in addition to existing bug datasets to get FrogBoss a state-of-the-art 32B parameter model on SWE-bench Verified with a pass@1 of 54.6% and FrogMini a state-of-the-art 14B model on SWE-bench Verified with a pass@1 of 45.3% on SWE-bench Verified averaged over three seeds.

Haonan Wang, Junfeng Sun, Xingdi Yuan et al.

Simulating interactive world models remains a core challenge in Large Language Models(LLMs). In this work, we introduce the ByteSized32Refactored, a refactored, modular, and extensible implementation of the original ByteSized32 corpus to explore the task of text game generation. We further optimize the code structure of each text game and create the GameBasic.py foundation library, which centralizes common logic across all 32 games by abstracting 7 base classes (GameObject, etc.) into reusable modules, thereby reducing from 20k to 10k total lines of Python code compared to the original Bytesized32. Our refactored implementation enables extendability - with our centralized design, ByteSized32Refactored can be more efficiently extended to include text games of new scenarios and specifications by reusing the shared logic and functionalities. Extensive experiments with GPT-4o demonstrate a mix of performance - with Bytesized32Refactored, the generated text games for unseen scenarios showcase quality improvements on two of the four evaluation dimensions while decreases on the other two, indicating that the hierarchical structure of the refactored code presents new challenges for LLMs. Overall, we highlight that our extensible code structure, centered on the foundation library and the modular optimization, not only facilitates LLM adaptation to environment specifications but also establishes a scalable environment that supports future extensions.

Jingcheng Niu, Xingdi Yuan, Tong Wang et al. · utoronto

We observe a novel phenomenon, contextual entrainment, across a wide range of language models (LMs) and prompt settings, providing a new mechanistic perspective on how LMs become distracted by ``irrelevant'' contextual information in the input prompt. Specifically, LMs assign significantly higher logits (or probabilities) to any tokens that have previously appeared in the context prompt, even for random tokens. This suggests that contextual entrainment is a mechanistic phenomenon, occurring independently of the relevance or semantic relation of the tokens to the question or the rest of the sentence. We find statistically significant evidence that the magnitude of contextual entrainment is influenced by semantic factors. Counterfactual prompts have a greater effect compared to factual ones, suggesting that while contextual entrainment is a mechanistic phenomenon, it is modulated by semantic factors. We hypothesise that there is a circuit of attention heads -- the entrainment heads -- that corresponds to the contextual entrainment phenomenon. Using a novel entrainment head discovery method based on differentiable masking, we identify these heads across various settings. When we ``turn off'' these heads, i.e., set their outputs to zero, the effect of contextual entrainment is significantly attenuated, causing the model to generate output that capitulates to what it would produce if no distracting context were provided. Our discovery of contextual entrainment, along with our investigation into LM distraction via the entrainment heads, marks a key step towards the mechanistic analysis and mitigation of the distraction problem.

Ruoyao Wang, Graham Todd, Ziang Xiao et al.

Virtual environments play a key role in benchmarking advances in complex planning and decision-making tasks but are expensive and complicated to build by hand. Can current language models themselves serve as world simulators, correctly predicting how actions change different world states, thus bypassing the need for extensive manual coding? Our goal is to answer this question in the context of text-based simulators. Our approach is to build and use a new benchmark, called ByteSized32-State-Prediction, containing a dataset of text game state transitions and accompanying game tasks. We use this to directly quantify, for the first time, how well LLMs can serve as text-based world simulators. We test GPT-4 on this dataset and find that, despite its impressive performance, it is still an unreliable world simulator without further innovations. This work thus contributes both new insights into current LLM's capabilities and weaknesses, as well as a novel benchmark to track future progress as new models appear.

Jikun Kang, Romain Laroche, Xingdi Yuan et al.

Decision Transformer-based decision-making agents have shown the ability to generalize across multiple tasks. However, their performance relies on massive data and computation. We argue that this inefficiency stems from the forgetting phenomenon, in which a model memorizes its behaviors in parameters throughout training. As a result, training on a new task may deteriorate the model's performance on previous tasks. In contrast to LLMs' implicit memory mechanism, the human brain utilizes distributed memory storage, which helps manage and organize multiple skills efficiently, mitigating the forgetting phenomenon. Inspired by this, we propose a working memory module to store, blend, and retrieve information for different downstream tasks. Evaluation results show that the proposed method improves training efficiency and generalization in Atari games and Meta-World object manipulation tasks. Moreover, we demonstrate that memory fine-tuning further enhances the adaptability of the proposed architecture.

Cédric Colas, Laetitia Teodorescu, Pierre-Yves Oudeyer et al.

Humans learn to master open-ended repertoires of skills by imagining and practicing their own goals. This autotelic learning process, literally the pursuit of self-generated (auto) goals (telos), becomes more and more open-ended as the goals become more diverse, abstract and creative. The resulting exploration of the space of possible skills is supported by an inter-individual exploration: goal representations are culturally evolved and transmitted across individuals, in particular using language. Current artificial agents mostly rely on predefined goal representations corresponding to goal spaces that are either bounded (e.g. list of instructions), or unbounded (e.g. the space of possible visual inputs) but are rarely endowed with the ability to reshape their goal representations, to form new abstractions or to imagine creative goals. In this paper, we introduce a language model augmented autotelic agent (LMA3) that leverages a pretrained language model (LM) to support the representation, generation and learning of diverse, abstract, human-relevant goals. The LM is used as an imperfect model of human cultural transmission; an attempt to capture aspects of humans' common-sense, intuitive physics and overall interests. Specifically, it supports three key components of the autotelic architecture: 1)~a relabeler that describes the goals achieved in the agent's trajectories, 2)~a goal generator that suggests new high-level goals along with their decomposition into subgoals the agent already masters, and 3)~reward functions for each of these goals. Without relying on any hand-coded goal representations, reward functions or curriculum, we show that LMA3 agents learn to master a large diversity of skills in a task-agnostic text-based environment.

Arjun Subramonian, Xingdi Yuan, Hal Daumé et al.

Progress in NLP is increasingly measured through benchmarks; hence, contextualizing progress requires understanding when and why practitioners may disagree about the validity of benchmarks. We develop a taxonomy of disagreement, drawing on tools from measurement modeling, and distinguish between two types of disagreement: 1) how tasks are conceptualized and 2) how measurements of model performance are operationalized. To provide evidence for our taxonomy, we conduct a meta-analysis of relevant literature to understand how NLP tasks are conceptualized, as well as a survey of practitioners about their impressions of different factors that affect benchmark validity. Our meta-analysis and survey across eight tasks, ranging from coreference resolution to question answering, uncover that tasks are generally not clearly and consistently conceptualized and benchmarks suffer from operationalization disagreements. These findings support our proposed taxonomy of disagreement. Finally, based on our taxonomy, we present a framework for constructing benchmarks and documenting their limitations.

Xingdi Yuan

Interactive machine reading comprehension (iMRC) is machine comprehension tasks where knowledge sources are partially observable. An agent must interact with an environment sequentially to gather necessary knowledge in order to answer a question. We hypothesize that graph representations are good inductive biases, which can serve as an agent's memory mechanism in iMRC tasks. We explore four different categories of graphs that can capture text information at various levels. We describe methods that dynamically build and update these graphs during information gathering, as well as neural models to encode graph representations in RL agents. Extensive experiments on iSQuAD suggest that graph representations can result in significant performance improvements for RL agents.

Rui Meng, Khushboo Thaker, Lei Zhang et al.

Faceted summarization provides briefings of a document from different perspectives. Readers can quickly comprehend the main points of a long document with the help of a structured outline. However, little research has been conducted on this subject, partially due to the lack of large-scale faceted summarization datasets. In this study, we present FacetSum, a faceted summarization benchmark built on Emerald journal articles, covering a diverse range of domains. Different from traditional document-summary pairs, FacetSum provides multiple summaries, each targeted at specific sections of a long document, including the purpose, method, findings, and value. Analyses and empirical results on our dataset reveal the importance of bringing structure into summaries. We believe FacetSum will spur further advances in summarization research and foster the development of NLP systems that can leverage the structured information in both long texts and summaries.

Mohit Shridhar, Xingdi Yuan, Marc-Alexandre Côté et al.

Given a simple request like Put a washed apple in the kitchen fridge, humans can reason in purely abstract terms by imagining action sequences and scoring their likelihood of success, prototypicality, and efficiency, all without moving a muscle. Once we see the kitchen in question, we can update our abstract plans to fit the scene. Embodied agents require the same abilities, but existing work does not yet provide the infrastructure necessary for both reasoning abstractly and executing concretely. We address this limitation by introducing ALFWorld, a simulator that enables agents to learn abstract, text based policies in TextWorld (Côté et al., 2018) and then execute goals from the ALFRED benchmark (Shridhar et al., 2020) in a rich visual environment. ALFWorld enables the creation of a new BUTLER agent whose abstract knowledge, learned in TextWorld, corresponds directly to concrete, visually grounded actions. In turn, as we demonstrate empirically, this fosters better agent generalization than training only in the visually grounded environment. BUTLER's simple, modular design factors the problem to allow researchers to focus on models for improving every piece of the pipeline (language understanding, planning, navigation, and visual scene understanding).

Rui Meng, Xingdi Yuan, Tong Wang et al.

Recent years have seen a flourishing of neural keyphrase generation (KPG) works, including the release of several large-scale datasets and a host of new models to tackle them. Model performance on KPG tasks has increased significantly with evolving deep learning research. However, there lacks a comprehensive comparison among different model designs, and a thorough investigation on related factors that may affect a KPG system's generalization performance. In this empirical study, we aim to fill this gap by providing extensive experimental results and analyzing the most crucial factors impacting the generalizability of KPG models. We hope this study can help clarify some of the uncertainties surrounding the KPG task and facilitate future research on this topic.

Shengding Hu, Zheng Xiong, Meng Qu et al.

Graph neural networks (GNNs) have been attracting increasing popularity due to their simplicity and effectiveness in a variety of fields. However, a large number of labeled data is generally required to train these networks, which could be very expensive to obtain in some domains. In this paper, we study active learning for GNNs, i.e., how to efficiently label the nodes on a graph to reduce the annotation cost of training GNNs. We formulate the problem as a sequential decision process on graphs and train a GNN-based policy network with reinforcement learning to learn the optimal query strategy. By jointly training on several source graphs with full labels, we learn a transferable active learning policy which can directly generalize to unlabeled target graphs. Experimental results on multiple datasets from different domains prove the effectiveness of the learned policy in promoting active learning performance in both settings of transferring between graphs in the same domain and across different domains.

Ashutosh Adhikari, Xingdi Yuan, Marc-Alexandre Côté et al.

Playing text-based games requires skills in processing natural language and sequential decision making. Achieving human-level performance on text-based games remains an open challenge, and prior research has largely relied on hand-crafted structured representations and heuristics. In this work, we investigate how an agent can plan and generalize in text-based games using graph-structured representations learned end-to-end from raw text. We propose a novel graph-aided transformer agent (GATA) that infers and updates latent belief graphs during planning to enable effective action selection by capturing the underlying game dynamics. GATA is trained using a combination of reinforcement and self-supervised learning. Our work demonstrates that the learned graph-based representations help agents converge to better policies than their text-only counterparts and facilitate effective generalization across game configurations. Experiments on 500+ unique games from the TextWorld suite show that our best agent outperforms text-based baselines by an average of 24.2%.

Mikuláš Zelinka, Xingdi Yuan, Marc-Alexandre Côté et al.

We are interested in learning how to update Knowledge Graphs (KG) from text. In this preliminary work, we propose a novel Sequence-to-Sequence (Seq2Seq) architecture to generate elementary KG operations. Furthermore, we introduce a new dataset for KG extraction built upon text-based game transitions (over 300k data points). We conduct experiments and discuss the results.

Matthew Hausknecht, Prithviraj Ammanabrolu, Marc-Alexandre Côté et al.

A hallmark of human intelligence is the ability to understand and communicate with language. Interactive Fiction games are fully text-based simulation environments where a player issues text commands to effect change in the environment and progress through the story. We argue that IF games are an excellent testbed for studying language-based autonomous agents. In particular, IF games combine challenges of combinatorial action spaces, language understanding, and commonsense reasoning. To facilitate rapid development of language-based agents, we introduce Jericho, a learning environment for man-made IF games and conduct a comprehensive study of text-agents across a rich set of games, highlighting directions in which agents can improve.

Rui Meng, Xingdi Yuan, Tong Wang et al.

Recently, concatenating multiple keyphrases as a target sequence has been proposed as a new learning paradigm for keyphrase generation. Existing studies concatenate target keyphrases in different orders but no study has examined the effects of ordering on models' behavior. In this paper, we propose several orderings for concatenation and inspect the important factors for training a successful keyphrase generation model. By running comprehensive comparisons, we observe one preferable ordering and summarize a number of empirical findings and challenges, which can shed light on future research on this line of work.

Xingdi Yuan, Marc-Alexandre Cote, Jie Fu et al.

Humans observe and interact with the world to acquire knowledge. However, most existing machine reading comprehension (MRC) tasks miss the interactive, information-seeking component of comprehension. Such tasks present models with static documents that contain all necessary information, usually concentrated in a single short substring. Thus, models can achieve strong performance through simple word- and phrase-based pattern matching. We address this problem by formulating a novel text-based question answering task: Question Answering with Interactive Text (QAit). In QAit, an agent must interact with a partially observable text-based environment to gather information required to answer questions. QAit poses questions about the existence, location, and attributes of objects found in the environment. The data is built using a text-based game generator that defines the underlying dynamics of interaction with the environment. We propose and evaluate a set of baseline models for the QAit task that includes deep reinforcement learning agents. Experiments show that the task presents a major challenge for machine reading systems, while humans solve it with relative ease.

Xingdi Yuan, Jie Fu, Marc-Alexandre Cote et al.

Existing machine reading comprehension (MRC) models do not scale effectively to real-world applications like web-level information retrieval and question answering (QA). We argue that this stems from the nature of MRC datasets: most of these are static environments wherein the supporting documents and all necessary information are fully observed. In this paper, we propose a simple method that reframes existing MRC datasets as interactive, partially observable environments. Specifically, we "occlude" the majority of a document's text and add context-sensitive commands that reveal "glimpses" of the hidden text to a model. We repurpose SQuAD and NewsQA as an initial case study, and then show how the interactive corpora can be used to train a model that seeks relevant information through sequential decision making. We believe that this setting can contribute in scaling models to web-level QA scenarios.

Yi Tay, Shuohang Wang, Luu Anh Tuan et al.

This paper tackles the problem of reading comprehension over long narratives where documents easily span over thousands of tokens. We propose a curriculum learning (CL) based Pointer-Generator framework for reading/sampling over large documents, enabling diverse training of the neural model based on the notion of alternating contextual difficulty. This can be interpreted as a form of domain randomization and/or generative pretraining during training. To this end, the usage of the Pointer-Generator softens the requirement of having the answer within the context, enabling us to construct diverse training samples for learning. Additionally, we propose a new Introspective Alignment Layer (IAL), which reasons over decomposed alignments using block-based self-attention. We evaluate our proposed method on the NarrativeQA reading comprehension benchmark, achieving state-of-the-art performance, improving existing baselines by $51\%$ relative improvement on BLEU-4 and $17\%$ relative improvement on Rouge-L. Extensive ablations confirm the effectiveness of our proposed IAL and CL components.

Ruo Yu Tao, Marc-Alexandre Côté, Xingdi Yuan et al.

To solve a text-based game, an agent needs to formulate valid text commands for a given context and find the ones that lead to success. Recent attempts at solving text-based games with deep reinforcement learning have focused on the latter, i.e., learning to act optimally when valid actions are known in advance. In this work, we propose to tackle the first task and train a model that generates the set of all valid commands for a given context. We try three generative models on a dataset generated with Textworld. The best model can generate valid commands which were unseen at training and achieve high $F_1$ score on the test set.

Rajarshi Das, Tsendsuren Munkhdalai, Xingdi Yuan et al.

We propose a neural machine-reading model that constructs dynamic knowledge graphs from procedural text. It builds these graphs recurrently for each step of the described procedure, and uses them to track the evolving states of participant entities. We harness and extend a recently proposed machine reading comprehension (MRC) model to query for entity states, since these states are generally communicated in spans of text and MRC models perform well in extracting entity-centric spans. The explicit, structured, and evolving knowledge graph representations that our model constructs can be used in downstream question answering tasks to improve machine comprehension of text, as we demonstrate empirically. On two comprehension tasks from the recently proposed PROPARA dataset (Dalvi et al., 2018), our model achieves state-of-the-art results. We further show that our model is competitive on the RECIPES dataset (Kiddon et al., 2015), suggesting it may be generally applicable. We present some evidence that the model's knowledge graphs help it to impose commonsense constraints on its predictions.

Xingdi Yuan, Tong Wang, Rui Meng et al.

Different texts shall by nature correspond to different number of keyphrases. This desideratum is largely missing from existing neural keyphrase generation models. In this study, we address this problem from both modeling and evaluation perspectives. We first propose a recurrent generative model that generates multiple keyphrases as delimiter-separated sequences. Generation diversity is further enhanced with two novel techniques by manipulating decoder hidden states. In contrast to previous approaches, our model is capable of generating diverse keyphrases and controlling number of outputs. We further propose two evaluation metrics tailored towards the variable-number generation. We also introduce a new dataset StackEx that expands beyond the only existing genre (i.e., academic writing) in keyphrase generation tasks. With both previous and new evaluation metrics, our model outperforms strong baselines on all datasets.

Marc-Alexandre Côté, Ákos Kádár, Xingdi Yuan et al.

We introduce TextWorld, a sandbox learning environment for the training and evaluation of RL agents on text-based games. TextWorld is a Python library that handles interactive play-through of text games, as well as backend functions like state tracking and reward assignment. It comes with a curated list of games whose features and challenges we have analyzed. More significantly, it enables users to handcraft or automatically generate new games. Its generative mechanisms give precise control over the difficulty, scope, and language of constructed games, and can be used to relax challenges inherent to commercial text games like partial observability and sparse rewards. By generating sets of varied but similar games, TextWorld can also be used to study generalization and transfer learning. We cast text-based games in the Reinforcement Learning formalism, use our framework to develop a set of benchmark games, and evaluate several baseline agents on this set and the curated list.

Xingdi Yuan, Marc-Alexandre Côté, Alessandro Sordoni et al.

We propose a recurrent RL agent with an episodic exploration mechanism that helps discovering good policies in text-based game environments. We show promising results on a set of generated text-based games of varying difficulty where the goal is to collect a coin located at the end of a chain of rooms. In contrast to previous text-based RL approaches, we observe that our agent learns policies that generalize to unseen games of greater difficulty.

Tsendsuren Munkhdalai, Xingdi Yuan, Soroush Mehri et al.

We describe a mechanism by which artificial neural networks can learn rapid adaptation - the ability to adapt on the fly, with little data, to new tasks - that we call conditionally shifted neurons. We apply this mechanism in the framework of metalearning, where the aim is to replicate some of the flexibility of human learning in machines. Conditionally shifted neurons modify their activation values with task-specific shifts retrieved from a memory module, which is populated rapidly based on limited task experience. On metalearning benchmarks from the vision and language domains, models augmented with conditionally shifted neurons achieve state-of-the-art results.

Sandeep Subramanian, Tong Wang, Xingdi Yuan et al.

We propose a two-stage neural model to tackle question generation from documents. First, our model estimates the probability that word sequences in a document are ones that a human would pick when selecting candidate answers by training a neural key-phrase extractor on the answers in a question-answering corpus. Predicted key phrases then act as target answers and condition a sequence-to-sequence question-generation model with a copy mechanism. Empirically, our key-phrase extraction model significantly outperforms an entity-tagging baseline and existing rule-based approaches. We further demonstrate that our question generation system formulates fluent, answerable questions from key phrases. This two-stage system could be used to augment or generate reading comprehension datasets, which may be leveraged to improve machine reading systems or in educational settings.

Tong Wang, Xingdi Yuan, Adam Trischler

We propose a generative machine comprehension model that learns jointly to ask and answer questions based on documents. The proposed model uses a sequence-to-sequence framework that encodes the document and generates a question (answer) given an answer (question). Significant improvement in model performance is observed empirically on the SQuAD corpus, confirming our hypothesis that the model benefits from jointly learning to perform both tasks. We believe the joint model's novelty offers a new perspective on machine comprehension beyond architectural engineering, and serves as a first step towards autonomous information seeking.

Xingdi Yuan, Tong Wang, Caglar Gulcehre et al.

We propose a recurrent neural model that generates natural-language questions from documents, conditioned on answers. We show how to train the model using a combination of supervised and reinforcement learning. After teacher forcing for standard maximum likelihood training, we fine-tune the model using policy gradient techniques to maximize several rewards that measure question quality. Most notably, one of these rewards is the performance of a question-answering system. We motivate question generation as a means to improve the performance of question answering systems. Our model is trained and evaluated on the recent question-answering dataset SQuAD.