List of Concepts | Summarize concepts and examples across multiple papers

List of Concepts is a computationally intensive workflow in Elicit that searches a shared corpus of roughly 200 million academic papers and then extracts, groups, and organizes recurring concepts or examples across multiple studies. Instead of forcing users to scan paper-by-paper tables, it surfaces a consolidated set of concepts (such as “data sets” for a given topic) and pairs each item with supporting quotes and direct links to the underlying papers. The practical payoff is faster literature navigation: users can get decision-relevant answers up front and only open the most relevant sources for context.

The workflow works by first identifying a set of relevant papers for a user’s query—an example given was “data sets for Quality Control in MRI,” which returned 40 relevant papers. It then processes the accessible text (full text when available, otherwise abstracts) to detect whether specific concepts or examples are discussed. When matches appear, it extracts the relevant quotes and performs a deduplication step to collapse near-identical descriptions that show up across multiple segments of text. Users can inspect intermediate steps if grouping looks wrong, then proceed to the final grouped results.

A key feature is provenance and uncertainty handling. For each extracted concept or example, Elicit provides the quote and a link to the paper where it was found. But the workflow can also incorporate answers generated by a language model—such as asking ChatGPT for “data sets for quality control and MRI.” Those model-suggested items are still displayed, yet flagged when no supporting evidence is found in the searched academic corpus. The intent is to keep potentially useful ideas visible while signaling the risk of hallucination or gaps caused by limited coverage.

Because List of Concepts combines multiple underlying Elicit building blocks—paper retrieval, concept extraction, and structural grouping—it tends to be more expensive than simpler workflows. The transcript notes that it processes significantly more papers, includes multiple processing steps, and therefore costs more credits; users can check credit usage in their credit history page. Results can also be downloaded as a CSV.

Beyond dataset-finding, the workflow is positioned as useful for techniques (e.g., “techniques for producing hallucination in language models”), effects (e.g., long-term effects of regular melatonin usage), and interdomain exploration. One example aimed at “principal-agent problems” that generalize to alignment and machine learning, then translated into concepts spanning economics, decision theory, and machine learning—such as “simplification optimal incentive schemes” and “overfitting to trading data.” The overall message is that the grouped, quote-backed structure helps users generate ideas across domains, while the flagged language-model outputs and inspectable steps help manage quality and trust.