TCA-SIR: Learning Target-Conditioned Abstractions for Scientific Inspiration Retrieval
TCA-SIR retrieves reusable scientific principles, not just papers with similar topics.
The paper reframes Scientific Inspiration Retrieval as target-conditioned abstraction: extracting how a candidate paper’s idea could transfer to a specific target problem. Its model generates those abstractions and uses them to predict transferability, aimed especially at “remote” inspirations with weak topical overlap. On ResearchBench, it beats prior SIR methods and direct LLM retrieval, with HitRate@top4% more than 10 points above MOOSE-Chem. ArXiv · AI/CL/LG's note
The paper reframes Scientific Inspiration Retrieval as target-conditioned abstraction: extracting how a candidate paper’s idea could transfer to a specific target problem. Its model generates those abstractions and uses them to predict transferability, aimed especially at “remote” inspirations with weak topical overlap. On ResearchBench, it beats prior SIR methods and direct LLM retrieval, with HitRate@top4% more than 10 points above MOOSE-Chem. ArXiv · AI/CL/LG's note
score 4