Peter Jansen

Ruoyao Wang, Peter Jansen, Marc-Alexandre Côté et al. · microsoft-research

We present ScienceWorld, a benchmark to test agents' scientific reasoning abilities in a new interactive text environment at the level of a standard elementary school science curriculum. Despite the transformer-based progress seen in question-answering and scientific text processing, we find that current models cannot reason about or explain learned science concepts in novel contexts. For instance, models can easily answer what the conductivity of a known material is but struggle when asked how they would conduct an experiment in a grounded environment to find the conductivity of an unknown material. This begs the question of whether current models are simply retrieving answers by way of seeing a large number of similar examples or if they have learned to reason about concepts in a reusable manner. We hypothesize that agents need to be grounded in interactive environments to achieve such reasoning capabilities. Our experiments provide empirical evidence supporting this hypothesis -- showing that a 1.5 million parameter agent trained interactively for 100k steps outperforms a 11 billion parameter model statically trained for scientific question-answering and reasoning from millions of expert demonstrations.

Ruoyao Wang, Peter Jansen, Marc-Alexandre Côté et al. · microsoft-research

In this work, we explore techniques for augmenting interactive agents with information from symbolic modules, much like humans use tools like calculators and GPS systems to assist with arithmetic and navigation. We test our agent's abilities in text games -- challenging benchmarks for evaluating the multi-step reasoning abilities of game agents in grounded, language-based environments. Our experimental study indicates that injecting the actions from these symbolic modules into the action space of a behavior cloned transformer agent increases performance on four text game benchmarks that test arithmetic, navigation, sorting, and common sense reasoning by an average of 22%, allowing an agent to reach the highest possible performance on unseen games. This action injection technique is easily extended to new agents, environments, and symbolic modules.

Bodhisattwa Prasad Majumder, Bhavana Dalvi Mishra, Peter Jansen et al.

Language agents have shown some ability to interact with an external environment, e.g., a virtual world such as ScienceWorld, to perform complex tasks, e.g., growing a plant, without the startup costs of reinforcement learning. However, despite their zero-shot capabilities, these agents to date do not continually improve over time beyond performance refinement on a specific task. Here we present CLIN, the first language-based agent to achieve this, so that it continually improves over multiple trials, including when both the environment and task are varied, and without requiring parameter updates. Our approach is to use a persistent, dynamic, textual memory centered on causal abstractions (rather than general "helpful hints") that is regularly updated after each trial so that the agent gradually learns useful knowledge for new trials. In the ScienceWorld benchmark, CLIN is able to continually improve on repeated trials on the same task and environment, outperforming state-of-the-art reflective language agents like Reflexion by 23 absolute points. CLIN can also transfer its learning to new environments (or new tasks), improving its zero-shot performance by 4 points (13 for new tasks) and can further improve performance there through continual memory updates, enhancing performance by an additional 17 points (7 for new tasks). This suggests a new architecture for agents built on frozen models that can still continually and rapidly improve over time.

Peter Jansen, Peter Clark, Doug Downey et al.

Contemporary automated scientific discovery has focused on agents for generating scientific experiments, while systems that perform higher-level scientific activities such as theory building remain underexplored. In this work, we formulate the problem of synthesizing theories consisting of qualitative and quantitative laws from large corpora of scientific literature. We study theory generation at scale, using 13.7k source papers to synthesize 2.9k theories, examining how generation using literature-grounding versus parametric knowledge, and accuracy-focused versus novelty-focused generation objectives change theory properties. Our experiments show that, compared to using parametric LLM memory for generation, our literature-supported method creates theories that are significantly better at both matching existing evidence and at predicting future results from 4.6k subsequently-written papers

Peter Jansen, Marc-Alexandre Côté, Tushar Khot et al.

Automated scientific discovery promises to accelerate progress across scientific domains. However, developing and evaluating an AI agent's capacity for end-to-end scientific reasoning is challenging as running real-world experiments is often prohibitively expensive or infeasible. In this work we introduce DISCOVERYWORLD, the first virtual environment for developing and benchmarking an agent's ability to perform complete cycles of novel scientific discovery. DISCOVERYWORLD contains a variety of different challenges, covering topics as diverse as radioisotope dating, rocket science, and proteomics, to encourage development of general discovery skills rather than task-specific solutions. DISCOVERYWORLD itself is an inexpensive, simulated, text-based environment (with optional 2D visual overlay). It includes 120 different challenge tasks, spanning eight topics each with three levels of difficulty and several parametric variations. Each task requires an agent to form hypotheses, design and run experiments, analyze results, and act on conclusions. DISCOVERYWORLD further provides three automatic metrics for evaluating performance, based on (a) task completion, (b) task-relevant actions taken, and (c) the discovered explanatory knowledge. We find that strong baseline agents, that perform well in prior published environments, struggle on most DISCOVERYWORLD tasks, suggesting that DISCOVERYWORLD captures some of the novel challenges of discovery, and thus that DISCOVERYWORLD may help accelerate near-term development and assessment of scientific discovery competency in agents. Code available at: www.github.com/allenai/discoveryworld

Ruoyao Wang, Graham Todd, Eric Yuan et al.

In this work, we investigate the capacity of language models to generate explicit, interpretable, and interactive world models of scientific and common-sense reasoning tasks. We operationalize this as a task of generating text games, expressed as hundreds of lines of Python code. To facilitate this task, we introduce ByteSized32 (Code: github.com/cognitiveailab/BYTESIZED32), a corpus of 32 reasoning-focused text games totaling 20k lines of Python code. We empirically demonstrate that GPT-4 can use these games as templates for single-shot in-context learning, successfully producing runnable games on unseen topics in 28% of cases. When allowed to self-reflect on program errors, game runnability substantially increases to 57%. While evaluating simulation fidelity is labor-intensive, we introduce a suite of automated metrics to assess game fidelity, technical validity, adherence to task specifications, and winnability, showing a high degree of agreement with expert human ratings. We pose this as a challenge task to spur further development at the juncture of world modeling and code generation.

Peter Jansen, Oyvind Tafjord, Marissa Radensky et al. · allen-ai

Despite the surge of interest in autonomous scientific discovery (ASD) of software artifacts (e.g., improved ML algorithms), current ASD systems face two key limitations: (1) they largely explore variants of existing codebases or similarly constrained design spaces, and (2) they produce large volumes of research artifacts (such as automatically generated papers and code) that are typically evaluated using conference-style paper review with limited evaluation of code. In this work we introduce CodeScientist, a novel ASD system that frames ideation and experiment construction as a form of genetic search jointly over combinations of research articles and codeblocks defining common actions in a domain (like prompting a language model). We use this paradigm to conduct hundreds of automated experiments on machine-generated ideas broadly in the domain of agents and virtual environments, with the system returning 19 discoveries, 6 of which were judged as being both at least minimally sound and incrementally novel after a multi-faceted evaluation beyond that typically conducted in prior work, including external (conference-style) review, code review, and replication attempts. Moreover, the discoveries span new tasks, agents, metrics, and data, suggesting a qualitative shift from benchmark optimization to broader discoveries.

Nathaniel Weir, Kate Sanders, Orion Weller et al.

Recent language models enable new opportunities for structured reasoning with text, such as the construction of intuitive, proof-like textual entailment trees without relying on brittle formal logic. However, progress in this direction has been hampered by a long-standing lack of a clear protocol for determining what valid compositional entailment is. This absence causes noisy datasets and limited performance gains by modern neuro-symbolic engines. To address these problems, we formulate a consistent and theoretically grounded approach to annotating decompositional entailment and evaluate its impact on LLM-based textual inference. We find that our new dataset, RDTE (Recognizing Decompositional Textual Entailment), has a substantially higher internal consistency (+9%) than prior decompositional entailment datasets. We also find that training an RDTE-oriented entailment classifier via knowledge distillation and employing it in an entailment tree reasoning engine significantly improves both accuracy and proof quality, illustrating the practical benefit of this advance for textual inference.

Peter Jansen, Samiah Hassan, Pragnya Narasimha

Automated Scientific Discovery (ASD) systems can help automatically generate and run code-based experiments, but their capabilities are limited by the code they can reliably generate from parametric knowledge alone. As a result, current systems either mutate a small number of manually-crafted experiment examples, or operate solely from parametric knowledge, limiting quality and reach. We introduce CodeDistiller, a system that automatically distills large collections of scientific Github repositories into a vetted library of working domain-specific code examples, allowing ASD agents to expand their capabilities without manual effort. Using a combination of automatic and domain-expert evaluation on 250 materials science repositories, we find the best model is capable of producing functional examples for 74% of repositories, while our downstream evaluation shows an ASD agent augmented with a CodeDistiller generated library produces more accurate, complete, and scientifically sound experiments than an agent with only general materials-science code examples.

Nathaniel Weir, Bhavana Dalvi Mishra, Orion Weller et al.

Recent reasoning methods (e.g., chain-of-thought, entailment reasoning) help users understand how language models (LMs) answer a single question, but they do little to reveal the LM's overall understanding, or "theory," about the question's topic, making it still hard to trust the model. Our goal is to materialize such theories - here called microtheories (a linguistic analog of logical microtheories) - as a set of sentences encapsulating an LM's core knowledge about a topic. These statements systematically work together to entail answers to a set of questions to both engender trust and improve performance. Our approach is to first populate a knowledge store with (model-generated) sentences that entail answers to training questions and then distill those down to a core microtheory that is concise, general, and non-redundant. We show that, when added to a general corpus (e.g., Wikipedia), microtheories can supply critical, topical information not necessarily present in the corpus, improving both a model's ability to ground its answers to verifiable knowledge (i.e., show how answers are systematically entailed by documents in the corpus, fully grounding up to +8% more answers), and the accuracy of those grounded answers (up to +8% absolute). We also show that, in a human evaluation in the medical domain, our distilled microtheories contain a significantly higher concentration of topically critical facts than the non-distilled knowledge store. Finally, we show we can quantify the coverage of a microtheory for a topic (characterized by a dataset) using a notion of $p$-relevance. Together, these suggest that microtheories are an efficient distillation of an LM's topic-relevant knowledge, that they can usefully augment existing corpora, and can provide both performance gains and an interpretable, verifiable window into the model's knowledge of a topic.

Peter Jansen, Samiah Hassan, Ruoyao Wang

Contemporary approaches to assisted scientific discovery use language models to automatically generate large numbers of potential hypothesis to test, while also automatically generating code-based experiments to test those hypotheses. While hypotheses can be comparatively inexpensive to generate, automated experiments can be costly, particularly when run at scale (i.e. thousands of experiments). Developing the capacity to filter hypotheses based on their feasibility would allow discovery systems to run at scale, while increasing their likelihood of making significant discoveries. In this work we introduce Matter-of-Fact, a challenge dataset for determining the feasibility of hypotheses framed as claims, while operationalizing feasibility assessment as a temporally-filtered claim verification task using backtesting. Matter-of-Fact includes 8.4k claims extracted from scientific articles spanning four high-impact contemporary materials science topics, including superconductors, semiconductors, batteries, and aerospace materials, while including qualitative and quantitative claims from theoretical, experimental, and code/simulation results. We show that strong baselines that include retrieval augmented generation over scientific literature and code generation fail to exceed 72% performance on this task (chance performance is 50%), while domain-expert verification suggests nearly all are solvable -- highlighting both the difficulty of this task for current models, and the potential to accelerate scientific discovery by making near-term progress.

Ruoyao Wang, Peter Jansen

In this work, we introduce a self-supervised behavior cloning transformer for text games, which are challenging benchmarks for multi-step reasoning in virtual environments. Traditionally, Behavior Cloning Transformers excel in such tasks but rely on supervised training data. Our approach auto-generates training data by exploring trajectories (defined by common macro-action sequences) that lead to reward within the games, while determining the generality and utility of these trajectories by rapidly training small models then evaluating their performance on unseen development games. Through empirical analysis, we show our method consistently uncovers generalizable training data, achieving about 90\% performance of supervised systems across three benchmark text games.

Jonathan Bragg, Mike D'Arcy, Nishant Balepur et al. · allen-ai

AI agents hold the potential to revolutionize scientific productivity by automating literature reviews, replicating experiments, analyzing data, and even proposing new directions of inquiry; indeed, there are now many such agents, ranging from general-purpose "deep research" systems to specialized science-specific agents, such as AI Scientist and AIGS. Rigorous evaluation of these agents is critical for progress. Yet existing benchmarks fall short on several fronts: they (1) fail to provide holistic, product-informed measures of real-world use cases such as science research; (2) lack reproducible agent tools necessary for a controlled comparison of core agentic capabilities; (3) do not account for confounding variables such as model cost and tool access; (4) do not provide standardized interfaces for quick agent prototyping and evaluation; and (5) lack comprehensive baseline agents necessary to identify true advances. In response, we define principles and tooling for more rigorously benchmarking agents. Using these, we present AstaBench, a suite that provides the first holistic measure of agentic ability to perform scientific research, comprising 2400+ problems spanning the entire scientific discovery process and multiple scientific domains, and including many problems inspired by actual user requests to deployed Asta agents. Our suite comes with the first scientific research environment with production-grade search tools that enable controlled, reproducible evaluation, better accounting for confounders. Alongside, we provide a comprehensive suite of nine science-optimized classes of Asta agents and numerous baselines. Our extensive evaluation of 57 agents across 22 agent classes reveals several interesting findings, most importantly that despite meaningful progress on certain individual aspects, AI remains far from solving the challenge of science research assistance.

Rosni Vasu, Peter Jansen, Pao Siangliulue et al. · allen-ai

While there has been a surge of interest in automated scientific discovery (ASD), especially with the emergence of LLMs, it remains challenging for tools to generate hypotheses that are both testable and grounded in the scientific literature. Additionally, existing ideation tools are not adaptive to prior experimental outcomes. We developed HARPA to address these challenges by incorporating the ideation workflow inspired by human researchers. HARPA first identifies emerging research trends through literature mining, then explores hypothesis design spaces, and finally converges on precise, testable hypotheses by pinpointing research gaps and justifying design choices. Our evaluations show that HARPA-generated hypothesis-driven research proposals perform comparably to a strong baseline AI-researcher across most qualitative dimensions (e.g., specificity, novelty, overall quality), but achieve significant gains in feasibility(+0.78, p$<0.05$, bootstrap) and groundedness (+0.85, p$<0.01$, bootstrap) on a 10-point Likert scale. When tested with the ASD agent (CodeScientist), HARPA produced more successful executions (20 vs. 11 out of 40) and fewer failures (16 vs. 21 out of 40), showing that expert feasibility judgments track with actual execution success. Furthermore, to simulate how researchers continuously refine their understanding of what hypotheses are both testable and potentially interesting from experience, HARPA learns a reward model that scores new hypotheses based on prior experimental outcomes, achieving approx. a 28\% absolute gain over HARPA's untrained baseline scorer. Together, these methods represent a step forward in the field of AI-driven scientific discovery.

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.

Peter Jansen

In this work, we show that contemporary language models have a previously unknown skill -- the capacity for electronic circuit design from high-level textual descriptions, akin to code generation. We introduce two benchmarks: Pins100, assessing model knowledge of electrical components, and Micro25, evaluating a model's capability to design common microcontroller circuits and code in the Arduino ecosystem that involve input, output, sensors, motors, protocols, and logic -- with models such as GPT-4 and Claude-V1 achieving between 60% to 96% Pass@1 on generating full devices. We include six case studies of using language models as a design assistant for moderately complex devices, such as a radiation-powered random number generator, an emoji keyboard, a visible spectrometer, and several assistive devices, while offering a qualitative analysis performance, outlining evaluation challenges, and suggesting areas of development to improve complex circuit design and practical utility. With this work, we aim to spur research at the juncture of natural language processing and electronic design.

Zhengnan Xie, Alice Saebom Kwak, Enfa George et al.

Space situational awareness typically makes use of physical measurements from radar, telescopes, and other assets to monitor satellites and other spacecraft for operational, navigational, and defense purposes. In this work we explore using textual input for the space situational awareness task. We construct a corpus of 48.5k news articles spanning all known active satellites between 2009 and 2020. Using a dependency-rule-based extraction system designed to target three high-impact events -- spacecraft launches, failures, and decommissionings, we identify 1,787 space-event sentences that are then annotated by humans with 15.9k labels for event slots. We empirically demonstrate a state-of-the-art neural extraction system achieves an overall F1 between 53 and 91 per slot for event extraction in this low-resource, high-impact domain.

Peter Jansen, Kelly Smith, Dan Moreno et al.

Building compositional explanations requires models to combine two or more facts that, together, describe why the answer to a question is correct. Typically, these "multi-hop" explanations are evaluated relative to one (or a small number of) gold explanations. In this work, we show these evaluations substantially underestimate model performance, both in terms of the relevance of included facts, as well as the completeness of model-generated explanations, because models regularly discover and produce valid explanations that are different than gold explanations. To address this, we construct a large corpus of 126k domain-expert (science teacher) relevance ratings that augment a corpus of explanations to standardized science exam questions, discovering 80k additional relevant facts not rated as gold. We build three strong models based on different methodologies (generation, ranking, and schemas), and empirically show that while expert-augmented ratings provide better estimates of explanation quality, both original (gold) and expert-augmented automatic evaluations still substantially underestimate performance by up to 36% when compared with full manual expert judgements, with different models being disproportionately affected. This poses a significant methodological challenge to accurately evaluating explanations produced by compositional reasoning models.

Peter Jansen, Jordan Boyd-Graber

Tamarian, a fictional language introduced in the Star Trek episode Darmok, communicates meaning through utterances of metaphorical references, such as "Darmok and Jalad at Tanagra" instead of "We should work together." This work assembles a Tamarian-English dictionary of utterances from the original episode and several follow-on novels, and uses this to construct a parallel corpus of 456 English-Tamarian utterances. A machine translation system based on a large language model (T5) is trained using this parallel corpus, and is shown to produce an accuracy of 76% when translating from English to Tamarian on known utterances.

Bhavana Dalvi, Peter Jansen, Oyvind Tafjord et al.

Our goal, in the context of open-domain textual question-answering (QA), is to explain answers by showing the line of reasoning from what is known to the answer, rather than simply showing a fragment of textual evidence (a "rationale'"). If this could be done, new opportunities for understanding and debugging the system's reasoning become possible. Our approach is to generate explanations in the form of entailment trees, namely a tree of multipremise entailment steps from facts that are known, through intermediate conclusions, to the hypothesis of interest (namely the question + answer). To train a model with this skill, we created ENTAILMENTBANK, the first dataset to contain multistep entailment trees. Given a hypothesis (question + answer), we define three increasingly difficult explanation tasks: generate a valid entailment tree given (a) all relevant sentences (b) all relevant and some irrelevant sentences, or (c) a corpus. We show that a strong language model can partially solve these tasks, in particular when the relevant sentences are included in the input (e.g., 35% of trees for (a) are perfect), and with indications of generalization to other domains. This work is significant as it provides a new type of dataset (multistep entailments) and baselines, offering a new avenue for the community to generate richer, more systematic explanations.

Hannah Smith, Zeyu Zhang, John Culnan et al.

Named entity recognition identifies common classes of entities in text, but these entity labels are generally sparse, limiting utility to downstream tasks. In this work we present ScienceExamCER, a densely-labeled semantic classification corpus of 133k mentions in the science exam domain where nearly all (96%) of content words have been annotated with one or more fine-grained semantic class labels including taxonomic groups, meronym groups, verb/action groups, properties and values, and synonyms. Semantic class labels are drawn from a manually-constructed fine-grained typology of 601 classes generated through a data-driven analysis of 4,239 science exam questions. We show an off-the-shelf BERT-based named entity recognition model modified for multi-label classification achieves an accuracy of 0.85 F1 on this task, suggesting strong utility for downstream tasks in science domain question answering requiring densely-labeled semantic classification.

Tushar Khot, Peter Clark, Michal Guerquin et al.

Composing knowledge from multiple pieces of texts is a key challenge in multi-hop question answering. We present a multi-hop reasoning dataset, Question Answering via Sentence Composition(QASC), that requires retrieving facts from a large corpus and composing them to answer a multiple-choice question. QASC is the first dataset to offer two desirable properties: (a) the facts to be composed are annotated in a large corpus, and (b) the decomposition into these facts is not evident from the question itself. The latter makes retrieval challenging as the system must introduce new concepts or relations in order to discover potential decompositions. Further, the reasoning model must then learn to identify valid compositions of these retrieved facts using common-sense reasoning. To help address these challenges, we provide annotation for supporting facts as well as their composition. Guided by these annotations, we present a two-step approach to mitigate the retrieval challenges. We use other multiple-choice datasets as additional training data to strengthen the reasoning model. Our proposed approach improves over current state-of-the-art language models by 11% (absolute). The reasoning and retrieval problems, however, remain unsolved as this model still lags by 20% behind human performance.

Dongfang Xu, Peter Jansen, Jaycie Martin et al.

Prior work has demonstrated that question classification (QC), recognizing the problem domain of a question, can help answer it more accurately. However, developing strong QC algorithms has been hindered by the limited size and complexity of annotated data available. To address this, we present the largest challenge dataset for QC, containing 7,787 science exam questions paired with detailed classification labels from a fine-grained hierarchical taxonomy of 406 problem domains. We then show that a BERT-based model trained on this dataset achieves a large (+0.12 MAP) gain compared with previous methods, while also achieving state-of-the-art performance on benchmark open-domain and biomedical QC datasets. Finally, we show that using this model's predictions of question topic significantly improves the accuracy of a question answering system by +1.7% P@1, with substantial future gains possible as QC performance improves.

Peter Jansen

Question Answering for complex questions is often modeled as a graph construction or traversal task, where a solver must build or traverse a graph of facts that answer and explain a given question. This "multi-hop" inference has been shown to be extremely challenging, with few models able to aggregate more than two facts before being overwhelmed by "semantic drift", or the tendency for long chains of facts to quickly drift off topic. This is a major barrier to current inference models, as even elementary science questions require an average of 4 to 6 facts to answer and explain. In this work we empirically characterize the difficulty of building or traversing a graph of sentences connected by lexical overlap, by evaluating chance sentence aggregation quality through 9,784 manually-annotated judgments across knowledge graphs built from three free-text corpora (including study guides and Simple Wikipedia). We demonstrate semantic drift tends to be high and aggregation quality low, at between 0.04% and 3%, and highlight scenarios that maximize the likelihood of meaningfully combining information.

Rebecca Sharp, Mihai Surdeanu, Peter Jansen et al.

A common model for question answering (QA) is that a good answer is one that is closely related to the question, where relatedness is often determined using general-purpose lexical models such as word embeddings. We argue that a better approach is to look for answers that are related to the question in a relevant way, according to the information need of the question, which may be determined through task-specific embeddings. With causality as a use case, we implement this insight in three steps. First, we generate causal embeddings cost-effectively by bootstrapping cause-effect pairs extracted from free text using a small set of seed patterns. Second, we train dedicated embeddings over this data, by using task-specific contexts, i.e., the context of a cause is its effect. Finally, we extend a state-of-the-art reranking approach for QA to incorporate these causal embeddings. We evaluate the causal embedding models both directly with a casual implication task, and indirectly, in a downstream causal QA task using data from Yahoo! Answers. We show that explicitly modeling causality improves performance in both tasks. In the QA task our best model achieves 37.3% P@1, significantly outperforming a strong baseline by 7.7% (relative).