A team of six researchers spent a year reading over 300 papers on Learning from Demonstration (LfD) to assess the scale of the robotics publication explosion and evaluate genuine research progress. Their findings: only ~20% of papers offered highly notable contributions, with the rest being incremental improvements or new application domains. Notably, citation and download counts did not correlate with true contribution quality. The study also tested LLMs as a reading aid, finding them useful for summarization and quantitative extraction but unable to assess true novelty or detect overstated claims. The authors recommend building a research engine that weights peer-review scores, adopting blind publication models, and using LLMs as a complement to — not replacement for — expert review.
