Papers by Peter Jansen

24 papers
Self-Supervised Behavior Cloned Transformers are Path Crawlers for Text Games (2023.findings-emnlp)

Copied to clipboard

Challenge: Text game agents are often modeled using reinforcement learning, but their performance is limited.
Approach: They propose a self-supervised behavior cloning transformer for text games . they explore trajectories that lead to reward within the games and then train small models . their approach consistently uncovers generalizable training data, achieving 90% performance of supervised systems across three benchmark text games.
Outcome: The proposed model achieves 90% performance on three text games.
CodeDistiller: Automatically Generating Code Libraries for Scientific Coding Agents (2026.acl-demo)

Copied to clipboard

Challenge: Automated Scientific Discovery (ASD) systems rely on parametric knowledge to generate and run code-based experiments.
Approach: They propose a system that distills large collections of scientific Github repositories into a vetted library of working domain-specific code examples.
Outcome: The proposed system produces more accurate, complete, and scientifically sound experiments than an agent with only general materials-science code examples.
On the Challenges of Evaluating Compositional Explanations in Multi-Hop Inference: Relevance, Completeness, and Expert Ratings (2021.emnlp-main)

Copied to clipboard

Challenge: a large corpus of domain-expert relevance ratings augments a corpus for compositional explanations . a writer's study shows that the evaluations of compositional inference models underestimate performance .
Approach: They construct a corpus of 126k domain-expert relevance ratings that augment explanations to standardized science exam questions.
Outcome: The results show that evaluations underestimate performance of compositional explanations . they show that models regularly discover and produce valid explanations that are different than gold explanations.
Multi-class Hierarchical Question Classification for Multiple Choice Science Exams (2020.lrec-1)

Copied to clipboard

Challenge: Prior work has demonstrated that question classification (QC) can help answer a question more accurately.
Approach: They propose to use a large dataset for question classification (QC) that contains 7,787 science exam questions paired with detailed classification labels from a fine-grained hierarchical taxonomy of 406 problem domains to train a BERT-based model.
Outcome: The proposed model achieves a large (+0.12 MAP) gain while also achieving state-of-the-art performance on benchmark open-domain and biomedical QC datasets.
Visually-Grounded Planning without Vision: Language Models Infer Detailed Plans from High-level Instructions (2020.findings-emnlp)

Copied to clipboard

Challenge: Currently, the best performing systems can complete less than 1% of tasks successfully . we show that it is possible to generate gold multi-step plans from language directives alone without any visual input in 26% of unseen cases .
Approach: They propose a task that requires a virtual robot to translate natural language directives into detailed sequences of actions that accomplish a goal in a real environment.
Outcome: The proposed task generates gold multi-step plans from natural language directives in 26% of unseen cases.
ScienceExamCER: A High-Density Fine-Grained Science-Domain Corpus for Common Entity Recognition (2020.lrec-1)

Copied to clipboard

Challenge: Named entity recognition identifies common classes of noun phrases in text, but these entity labels are sparse, limiting utility to downstream tasks.
Approach: They propose a name-based named entity recognition model that annotates all content words with a fine-grained semantic class label.
Outcome: The proposed model achieves 0.85 F1 on the science exam domain domain . the proposed model is a powerful tool for question answering and inference .
Generating Literature-Driven Scientific Theories at Scale (2026.acl-long)

Copied to clipboard

Challenge: Contemporary automated scientific discovery systems focus on generating experiments, but higher-level activities such as theory building remain underexplored.
Approach: They propose to synthesize theories from scientific literature using literature-grounding versus parametric knowledge.
Outcome: The proposed method matches existing evidence better than parametric LLM memory generation.
Behavior Cloned Transformers are Neurosymbolic Reasoners (2023.eacl-main)

Copied to clipboard

Challenge: In injecting actions from 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% . contemporary agents struggle on tasks such as navigation, admetic and other tasks that humans make use of external tools.
Approach: They propose to inject actions from symbolic modules into the action space of a behavior cloned transformer agent to increase performance on four text game benchmarks that test arithmetic, navigation, sorting, and common sense reasoning by an average of 22% .
Outcome: The proposed method improves performance on four text game benchmarks that test arithmetic, navigation, sorting, and common sense reasoning by 22%, allowing an agent to reach the highest possible performance on unseen games.
Explaining Answers with Entailment Trees (2021.emnlp-main)

Copied to clipboard

Challenge: ENTAILMENTBANK is the first dataset to contain multistep entailment trees.
Approach: They propose to generate explanations in the form of entailment trees, a tree of multipremise entanglements steps from facts that are known to the hypothesis of interest.
Outcome: The proposed model can generate explanations in the form of entailment trees . this is a tree of multipremise enttailment steps from facts known to the hypothesis of interest.
TextWorldExpress: Simulating Text Games at One Million Steps Per Second (2023.eacl-demo)

Copied to clipboard

Challenge: Text-based games are limited in speed due to legacy tooling and limited implementations.
Approach: They propose a text-game simulator that reimplements three common text game benchmarks and runs three orders of magnitude faster than their original implementations.
Outcome: The proposed simulator increases throughput by three orders of magnitude, reaching over one million steps per second on common desktop hardware.
Enhancing Systematic Decompositional Natural Language Inference Using Informal Logic (2024.emnlp-main)

Copied to clipboard

Challenge: Recent language models allow structured reasoning with text, but lack of a clear protocol for discerning entailment causes noisy datasets and limited performance gains.
Approach: They propose a consistent approach to annotating decompositional entailment and evaluate its impact on LLM-based textual inference.
Outcome: The proposed approach has higher internal consistency than prior decompositional entailment datasets and significantly improves proof quality and accuracy.
ByteSized32: A Corpus and Challenge Task for Generating Task-Specific World Models Expressed as Text Games (2023.emnlp-main)

Copied to clipboard

Challenge: We show that language models can generate explicit, interpretable, and interactive world models of scientific and common-sense reasoning tasks.
Approach: They propose a corpus of 32 reasoning-focused text games expressed as hundreds of lines of Python code to facilitate this task.
Outcome: The proposed games can generate runnable games on unseen topics in 28% of cases.
PDDLEGO: Iterative Planning in Textual Environments (2024.starsem-1)

Copied to clipboard

Challenge: Existing methods to plan in textual environments rely on a fully-observed environment where all entity states are known, but are not interpretable.
Approach: They propose to use LLMs to generate a formal representation of the environment that can be solved by a symbolic planner.
Outcome: The proposed model outperforms existing methods in the Coin Collector simulation and Cooking World simulations.
WorldTree: A Corpus of Explanation Graphs for Elementary Science Questions supporting Multi-hop Inference (L18-1)

Copied to clipboard

Challenge: Existing methods of automated inference do not provide enough gold explanations to train models . standardized science exams are a challenge task for question answering .
Approach: They propose to manually construct a corpus of explanations for standardized science exams . they also provide an explanation-centered tablestore that contains the knowledge to construct these explanations .
Outcome: The proposed model provides detailed explanations for standardized science exams . the authors show that the proposed model can be trained on the basis of gold explanations .
ScienceWorld: Is your Agent Smarter than a 5th Grader? (2022.emnlp-main)

Copied to clipboard

Challenge: Existing models cannot reason about or explain learned science concepts in novel contexts, despite transformer-based progress in question-answering and scientific text processing .
Approach: They propose a benchmark to test agents’ scientific reasoning abilities in a new interactive text environment at the level of a standard elementary school science curriculum.
Outcome: The proposed model outperforms a model trained for 100k steps in a standard elementary school science curriculum.
Extracting Common Inference Patterns from Semi-Structured Explanations (D19-60)

Copied to clipboard

Challenge: Multi-hop inference suffers from semantic drift, or the tendency for chains of reasoning to "drift"' to unrelated topics.
Approach: They propose to extract large high-confidence multi-hop inference patterns from a corpus of explanations by abstracting large-scale structure from logical sentences.
Outcome: The proposed method extracts large high-confidence multi-hop inference patterns from a “matter” subset of elementary science exam questions.
Matter-of-Fact: A Benchmark for Verifying the Feasibility of Literature-Supported Claims in Materials Science (2025.emnlp-main)

Copied to clipboard

Challenge: Existing systems generate hypothesis, run experiments, analyze data, and write or review papers, but they are costly and impractical.
Approach: They propose a challenge dataset for determining the feasibility of hypotheses framed as claims and a temporally-filtered claim verification task using backtesting to test the validity of claims.
Outcome: The proposed model performs well on retrieval augmented generation and code generation while performing 50% of the task.
WorldTree V2: A Corpus of Science-Domain Structured Explanations and Inference Patterns supporting Multi-Hop Inference (2020.lrec-1)

Copied to clipboard

Challenge: Standardized science questions require combining an average of 6 facts and as many as 16 facts to answer and explain.
Approach: They propose to combine an average of 6 facts and as many as 16 facts to produce an answer for complex questions.
Outcome: The proposed model is based on a corpus of 5,114 standardized science exam questions . it uses multi-fact explanations that combine science knowledge and world knowledge .
CodeScientist: End-to-End Semi-Automated Scientific Discovery with Code-based Experimentation (2025.findings-acl)

Copied to clipboard

Challenge: Automated scientific discovery (ASD) systems are limited in their evaluation of software artifacts and large volumes of research artifs are typically evaluated using conference-style paper review with limited evaluation of code.
Approach: They propose 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.
Outcome: The proposed system returns 19 discoveries on machine-generated ideas in the domain of agents and virtual environments.
TextGraphs 2019 Shared Task on Multi-Hop Inference for Explanation Regeneration (D19-53)

Copied to clipboard

Challenge: Detailed extended analyses of all submitted systems showed large relative improvements in accessing the most challenging multi-hop inference problems, while absolute performance remains low.
Approach: The Shared Task on Multi-Hop Inference for Explanation Regeneration asks participants to regenerate detailed gold explanations for elementary science questions by selecting facts from a knowledge base of semi-structured tables.
Outcome: The top-performing system achieved a mean average precision of 0.56 . the task combines facts from a knowledge base and supervised training data .
CoSaTa: A Constraint Satisfaction Solver and Interpreted Language for Semi-Structured Tables of Sentences (2020.emnlp-demos)

Copied to clipboard

Challenge: Existing methods for inference using multiple facts are unable to combine more than two together.
Approach: They present an intuitive constraint satisfaction solver and interpreted language for knowledge bases of semi-structured tables expressed as text.
Outcome: The solver and interpreter are released as open source . the solver can express complex compositional "inference patterns"
Can Language Models Serve as Text-Based World Simulators? (2024.acl-short)

Copied to clipboard

Challenge: Recent advances in large language models (LLMs) have pointed towards an alternative approach by leveraging the huge amount of knowledge contained in their pre-training datasets.
Approach: They build and use a benchmark to quantify how well text-based simulators can serve as text-driven world simulators.
Outcome: The proposed benchmark aims to quantify how well language models can serve as world simulators.
Extracting Space Situational Awareness Events from News Text (2022.lrec-1)

Copied to clipboard

Challenge: Space situational awareness is the decisionmaking knowledge required to predict, avoid, operate through, or recover from the loss, disruption, or degradation of space services, capabilities, or activities.
Approach: They construct a corpus of 48.5k news articles spanning all known active satellites between 2009 and 2020 that are annotated by humans with 15.9k labels for event slots.
Outcome: The proposed system achieves an overall F1 between 53 and 91 per slot for event extraction in this low-resource, high-impact domain.
From Words to Wires: Generating Functioning Electronic Devices from Natural Language Descriptions (2023.findings-emnlp)

Copied to clipboard

Challenge: Recent work shows that language models have a capacity for electronic circuit design from high-level textual descriptions, akin to code generation.
Approach: They propose two benchmarks to assess a model's ability to design common microcontroller circuits and code in the Arduino ecosystem.
Outcome: The proposed model knowledge of electrical components and the model's ability to generate full circuits and code in the Arduino ecosystem are evaluated.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations