When AI Learns to Think Like a Historian: CANA and Analogical Deep Research

Something AI Is Bad At

Ask ChatGPT: "Iran blockades the Strait of Hormuz in 2026 — which historical event is most analogous?"

It will likely answer: "The 1973 oil crisis." Or "The 1980 Iran-Iraq War."

Both answers sound plausible, but they are surface matches — overlaps of keywords like "Middle East + oil + conflict." A real historian wouldn't think this way. They would ask: what is the event's causal structure? Supply chain disruption? Geopolitical games? Market panic? Then they would find historical events with similar causal structures, even if they look unrelated on the surface.

Yongqiang Chen et al. (MBZUAI and CMU), in the paper *Analogical Deep Research: Retrieving and Integrating Historical Analogies for Foresight Analysis*, reveal a key finding: LLM agents systematically fail at finding historical analogies because they match surface features instead of underlying mechanisms.

The Core of the Problem: Analogy Is a Causal Question

The paper's central claim: Analogical Deep Research (ADR) is fundamentally a causal reasoning problem.

Why? Because judging whether two events are "analogous" requires understanding why events happen, not just what they are. The similarity between the 1973 oil crisis and a 2026 Hormuz blockade lies not in "both being Middle East oil crises," but in:

• Mechanism 1: Supply chain disruption causing short-term price shocks
• Mechanism 2: Market panic causing overshoot, followed by regression
• Mechanism 3: Deterrence and counter-deterrence in geopolitical games

Only by identifying these underlying mechanisms can you find true analogies — e.g., the 1956 Suez Crisis (mechanisms 1+3), the 2020 COVID supply chain rupture (mechanisms 1+2), or the 1962 Cuban Missile Crisis (mechanism 3).

From this analysis, the paper proposes two principles:

1. Mechanism Alignment: Analogies must match at the causal mechanism level, not on surface features. 2. Cross-Analogy Confirmation: An event should be supported by multiple independent analogies, not a single one.

The CANA Framework

Based on these principles, the authors propose CANA (Causal Analogical Researcher) — an agent framework that guides LLMs to do analogical deep research.

CANA's core innovation is Structural Decomposition: breaking a complex event into multiple causal mechanisms, finding analogies for each independently, then integrating them.

The workflow:

1. Structural decomposition: Split the current event into N causal sub-mechanisms (e.g., "supply disruption," "market panic," "geopolitical deterrence") 2. Independent retrieval: For each sub-mechanism, retrieve structurally similar events from historical corpora 3. Structural feedback: Verify in reverse — "Does this event really match on causal mechanisms?" 4. Reflective refinement: If not, re-decompose or re-retrieve 5. Cross-analogy integration: Aggregate analogy events across sub-mechanisms, identifying historical events pointed to by multiple mechanisms

The key is step 3's structural feedback. It's not simple "retrieve-return" but "retrieve-verify-refine" — mirroring how human historians think: form a hypothesis, test it against sources, and revise if the sources don't support it.

A Concrete Case

The paper includes a compelling case study: "Why is the S&P near all-time highs despite the energy shock?" (Scenario: May 2026, Iran war, Strait of Hormuz closed, oil at $118/barrel)

Standard deep research agents (Gemini DR, ChatGPT DR, Qwen DR):

• Gemini DR produced a 39KB report citing zero historical events, saying: "This is a notable anomaly... rapid V-shaped recovery in equities... lagged effects of the energy shock, costs not yet transmitted to earnings." — generic analysis with no historical anchors.
• ChatGPT DR and Qwen DR likewise cited zero events, just listing risks: "ceasefire collapse, OPEC splits."

CANA (Claude):

Identified 4 mechanisms × 3+ historical events:

> "Equity market inflections lead physical supply normalization by 6–18 months, when..."

CANA's historical analogies included:

• Post-1973 oil crisis stock market rebound patterns
• Market responses during the 1990 Gulf crisis
• Recovery paths after the 2020 COVID supply shocks

Each analogy maps to a specific causal mechanism, not a vague "oil crisis" label.

Experimental Results

The paper introduces ADR-bench — the first analogical deep research benchmark. On it:

• CANA delivers up to 10% improvement in historical analogy generation
• CANA outperforms current state-of-the-art deep research agents
• For ongoing-event case studies, experts rated CANA's analogies as more insightful

ADR-bench evaluation dimensions include causal isomorphism (C. Isom.), directional consequences (Dir. Con.), mapping consistency (Map. Con.), coverage (Cov.), system alignment (Sys. Alig.), and novelty (Nov.), with GPT-5.4 as judge.

An Analogy to Explain the Analogy

What CANA does can be explained with an analogy:

A normal LLM finding analogies is like an undergraduate writing a paper — search keywords, find a few loosely related pieces, stack them together, submit. It passes plagiarism checks, but the advisor can tell there's no real understanding.

CANA finding analogies is like a senior historian — decompose the event into causal threads, trace each thread independently, find genuinely structurally similar historical events, cross-validate, and integrate into one insightful analysis.

The difference isn't knowledge volume (LLMs already have plenty of historical knowledge) but thinking style — staying at surface keywords versus digging into causal mechanisms.

Why This Matters

For agent research: CANA exposes a systematic weakness of LLM agents — they are good at describing, not at structured causal reasoning. This weakness won't be automatically fixed by bigger models or more data; it requires framework-level intervention. Structural decomposition + feedback verification is a general-purpose approach.

For forecasting and decision-making: Historical analogy is a core tool for human foresight. From policymakers to investors, people reason with "this is like last time X happened." If AI can more accurately find structurally similar historical events, it can provide more valuable decision support.

For history itself: CANA's methodology aligns with "comparative history" in academia. It doesn't replace historians — it formalizes historians' methods so AI can follow them.

Honest Limitations

• ADR-bench relies on an LLM judge (GPT-5.4), creating circular evaluation risk.
• Structural decomposition depends on LLM capability — if the LLM misinterprets the event, the decomposed mechanisms are wrong, and downstream retrieval accuracy is moot.
• Historical corpus coverage: the paper doesn't detail the sources and scope of the historical events corpus. If it only covers famous major events, many obscure but structurally similar analogies will be missed.
• 10% improvement sounds modest: but given it's on top of SOTA deep research agents and concentrated in core dimensions like "causal isomorphism," the practical value may exceed what the number suggests.

Closing Thoughts

CANA reminds me of a story:

A senior intelligence analyst, weeks before September 11, 2001, reportedly wrote a report noting that "terrorist organizations might use civilian aircraft as weapons." His basis wasn't classified intelligence, but: a 1944 B-25 bomber crashing into the Empire State Building (mechanical failure), the 1994 Air France 8969 hijacking (terrorists planning to crash into the Eiffel Tower, thwarted), and the 1999 Indian Airlines 814 hijacking. Across these seemingly unrelated events, he saw the same causal structure: civilian aircraft + terrorist control = flying bomb.

That is true analogical thinking. It requires no specific dates or flight numbers — only recognition of structure.

CANA takes AI one step in this direction. The 10% improvement may look small, but it proves something: structured causal reasoning can be guided at the framework level. That is more meaningful than simply stacking parameters.

---

Paper: Analogical Deep Research: Retrieving and Integrating Historical Analogies for Foresight Analysis Authors: Yongqiang Chen, Guangyi Chen, Yuewen Sun, Kun Zhang (MBZUAI / Carnegie Mellon University)