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

Something AI Is Bad At

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

It will likely answer: "It's like the 1973 oil crisis," or "It's like the 1980 Iran-Iraq War."

These answers sound plausible but are surface matches—overlapping keywords of "Middle East + oil + conflict." A real historian would think differently. They would ask: what is this event's causal structure? Is it a supply chain disruption? Geopolitical maneuvering? Market panic? Then they would seek historical events with structurally similar causal mechanisms, even if the surface looks unrelated.

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 Essence of the Problem: Analogy Is a Causal Question

The paper's core 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 hypothetical 2026 Hormuz blockade lies not in "both being Middle East oil crises," but in mechanisms such as:

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

Only by identifying these underlying mechanisms can true analogies be found—for example, the 1956 Suez Canal crisis (mechanisms 1+3), the 2020 COVID supply chain break (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 surface features. 2. Cross-Analogy Confirmation: An event should be corroborated by multiple independent analogies rather than settling on one.

The CANA Framework

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

CANA's core innovation is Structural Decomposition: break a complex event into multiple causal mechanisms, find analogies for each independently, then integrate.

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: Reverse-validate retrieved events—"does this event really match causally?" 4. Reflective improvement: If not matching, re-decompose or re-retrieve 5. Cross-analogy integration: Aggregate analogies across sub-mechanisms, identifying historical events pointed to by multiple mechanisms

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

A Concrete Case

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

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

• Gemini DR produced a 39KB report citing zero historical events, calling it "a notable anomaly… rapid V-shaped recovery… lagged effects of the energy shock." Generic analysis with no historical anchors.
• ChatGPT DR and Qwen DR likewise cited 0 events, merely listing risks ("ceasefire collapse, OPEC split").

CANA (Claude):

Identified 4 mechanisms × 3+ historical events:

> "Equity inflection points lead physical supply normalization by 6–18 months, when…"

CANA's analogies included:

• Post-1973 oil crisis equity rebound patterns
• Market reactions during the 1990 Gulf crisis
• Recovery paths after the 2020 COVID supply shock

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

Experimental Results

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

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

ADR-bench evaluates dimensions including causal isomorphism, directional consequences, mapping consistency, coverage, system alignment, and novelty, with GPT-5.4 as judge.

An Analogy to Explain the Analogy

What CANA does can be explained by an analogy:

An ordinary LLM finds analogies like an undergraduate writing a paper—keyword search, grab a few relevantly titled sources, stack them together. It passes plagiarism checks, but the advisor knows there's no real understanding.

CANA finds analogies like a seasoned historian—decompose the event into causal threads, trace each thread independently to genuinely similar historical events, cross-validate, and integrate into an insightful analysis.

The difference isn't knowledge (LLMs already know plenty of history) but mode of thinking—staying at surface keywords vs. going deep into causal mechanisms.

Why This Matters

For agent research: CANA reveals a systematic weakness of LLM agents—they excel at description, not structured causal reasoning. This weakness isn't automatically fixed by bigger models or more data; it needs framework-level intervention. Structural decomposition plus feedback validation is a generalizable approach.

For forecasting and decision-making: Historical analogy is a core human tool for forward-looking analysis. Policymakers and investors alike reason with "this looks like the last X." If AI can more accurately find structurally similar historical events, decision support becomes more valuable.

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

Honest Limitations

• ADR-bench evaluation relies on an LLM judge (GPT-5.4), risking circular evaluation.
• Structural decomposition depends on LLM capability—if the LLM misunderstands the event, the decomposed mechanisms are wrong, and downstream retrieval precision doesn't help.
• Historical corpus coverage: the paper doesn't detail the corpus sources and range. If it covers only famous events, many obscure but structurally similar analogies will be missed.
• A 10% gain sounds modest: but measured on top of SOTA deep research agents, concentrated on the core "causal isomorphism" dimension, the practical value may exceed what the number suggests.

Closing Thought

CANA's approach recalls a story: a senior intelligence analyst who, weeks before September 11, 2001, reportedly warned that terrorists might use civilian aircraft as weapons. His basis wasn't classified intelligence but pattern recognition across seemingly unrelated events—the 1944 B-25 crash into the Empire State Building, the 1994 Air France 8969 hijacking (planned to hit the Eiffel Tower but foiled), the 1999 Indian Airlines 814 hijacking. In all of them he saw one causal structure: civilian aircraft + terrorist control = flying bomb.

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

CANA takes AI a step in this direction. The 10% gain may look small, but it proves something: structured causal reasoning can be guided at the framework level. That matters more 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)