CANA framework boosts LLM historical analogy retrieval by 10%
LLMs match surface features, not deep mechanisms — CANA changes that for better foresight.
A new research paper from Yongqiang Chen, Guangyi Chen, Yuewen Sun, and Kun Zhang introduces Analogical Deep Research (ADR) — a task that tests whether LLM agents can find and leverage historical analogies for foresight analysis. The authors built the first benchmark, ADR-bench, and discovered a critical flaw: LLMs consistently match surface-level features (e.g., similar names or dates) rather than the underlying causal mechanisms that make analogies truly useful. This surface-matching behavior leads to poor quality analogies and unreliable foresight predictions.
To solve this, the team developed the Causal Analogical Researcher (CANA) framework. CANA is built on two theoretical principles: mechanism alignment and cross-analogy confirmation. It uses a structural decomposition representation to break events into causal components, then applies structural feedback to iteratively refine the analogy search. In tests on ADR-bench, CANA achieved up to 10% improvement in historical analogy generation and outperformed existing deep research agents. Case studies with ongoing real-world events confirmed that CANA's analogies are more structurally sound, offering a practical path toward AI that can learn from history more like human experts do.
- LLMs fail at historical analogy because they match surface features, not causal mechanisms.
- CANA uses structural decomposition and reflective feedback to find deeper analogies.
- CANA achieves a 10% improvement over state-of-the-art deep research agents on the ADR-bench benchmark.
Why It Matters
Enables AI to learn from history more accurately, improving strategic decision-making and foresight analysis.