🤖 AI

Unlocking LLM Creativity in Science through Analogical Reasoning

This paper introduces analogical reasoning as a novel approach to mitigate mode collapse in large language models, demonstrating that it significantly enhances the diversity and novelty of generated scientific solutions while achieving state-of-the-art performance across multiple biomedical tasks.

Original authors: Andrew Shen, Shaul Druckmann, James Zou

Published 2026-05-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Andrew Shen, Shaul Druckmann, James Zou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a scientist trying to solve a very difficult puzzle. You ask a super-smart AI assistant for ideas on how to solve it. Unfortunately, the AI keeps giving you the same five answers over and over again, just phrased slightly differently. It's like asking a chef for 100 different dinner ideas, and they keep suggesting "Spaghetti" but calling it "Pasta," "Noodles," or "Italian Strings." In the world of AI, this is called mode collapse. The AI gets stuck in a rut and stops being creative.

This paper, titled "Unlocking LLM Creativity in Science through Analogical Reasoning," proposes a clever new way to break the AI out of that rut. The authors, from Stanford University, suggest teaching the AI to think like a metaphor-maker.

The Problem: The AI's "Echo Chamber"

When scientists ask AI to generate new ideas for complex problems (like how to stop a virus or predict how a cell reacts to a drug), the AI tends to get lazy. It grabs the most famous, obvious solutions it knows and repeats them.

The Result: If you ask 100 times, you might get 95 answers that are basically the same thing.
The Consequence: You miss out on the weird, wild, and potentially brilliant ideas that could actually solve the problem.

The Solution: The "Cross-Domain Detective"

The authors introduce a method called Analogical Reasoning (AR). Instead of just asking the AI, "How do I fix this?" they ask it to play a game of "What is this like?"

Think of it like this:

The Old Way: Asking a mechanic, "How do I fix a broken engine?" The mechanic looks at engines and suggests standard fixes.
The New Way (AR): Asking the mechanic, "How do you fix a broken engine? Now, imagine this engine is actually a traffic jam in a busy city. How would a traffic engineer fix the jam? Now, take that traffic solution and apply it back to the engine."

The AI is trained to:

Look for the hidden structure: It ignores the surface details (like "cells" or "cars") and looks at the relationships (like "things moving through a crowded space" or "signals getting blocked").
Jump to a totally different world: It finds a problem in a completely different field (like economics, chess, or seismology) that has the same hidden structure.
Steal the solution: It takes the solution from that other field and translates it back to the original science problem.

Real-Life Examples from the Paper

The paper shows this isn't just theory; they actually used it to solve real biomedical problems. Here are two examples of how the AI "thought outside the box":

1. The "Economics" Drug Solution

The Problem: Scientists were trying to predict how individual cells react to drugs. The AI kept suggesting models that only predicted the average reaction, missing the unique behavior of individual cells.
The Analogy: The AI realized this was like Economics. Just as a single consumer reacts differently to a price hike than the "average" consumer, a single cell reacts differently to a drug.
The Result: The AI suggested using a statistical tool from economics called Finite Mixture Models. When they tested this, it worked incredibly well, predicting individual cell behaviors far better than previous methods.

2. The "Chess" DNA Solution

The Problem: Scientists were trying to predict the properties of short DNA chains. The AI kept suggesting models that ignored where a specific piece of DNA was located in the chain.
The Analogy: The AI realized this was like Chess. In chess, a pawn is weak in the middle of the board but strong near the edge. Its value depends entirely on its position.
The Result: The AI suggested using Piece-Square Tables (a chess strategy that values pieces based on their board position). Applying this to DNA allowed the models to understand that the position of a genetic sequence matters, leading to state-of-the-art results.

What Did They Find?

The authors tested this method against standard AI approaches and found:

More Variety: The AI generated solutions that were 90% to 173% more diverse. It stopped repeating the same "Spaghetti" answers.
More Novelty: Over 50% of the solutions generated by this method were genuinely new ideas (compared to less than 2% for standard AI).
Real Success: When they actually built and tested these new ideas in the lab (or via computer simulation), they consistently outperformed existing methods in predicting cell behavior, brain interactions, and drug properties.

The Bottom Line

This paper argues that to make AI a true partner in scientific discovery, we can't just let it search for answers in its own "library." We have to force it to look out the window, see how a traffic jam works, or how a chess game is played, and then ask, "How does that help us solve our science problem?"

By using analogies, the AI stops being a boring copycat and starts acting like a creative inventor, bridging the gap between unrelated fields to find solutions that humans might have missed.

Technical Summary: Unlocking LLM Creativity in Science through Analogical Reasoning

Problem Statement

Autonomous scientific discovery, particularly in complex fields like biomedicine, relies on AI systems capable of generating novel and diverse solutions to open-ended problems. However, current Large Language Models (LLMs) exhibit a significant tendency toward mode collapse when tasked with open-ended solution generation. Instead of exploring a wide search space, LLMs converge on semantically similar, high-probability generations (often canonical solutions like AlphaFold for protein folding), severely limiting the diversity of potential discoveries. This lack of diversity hinders the ability of AI systems to serve as effective engines for autonomous science, as researchers require a broad array of candidate solutions to tackle problems from multiple angles and avoid premature convergence on existing paradigms.

Methodology

The authors introduce Analogical Reasoning (AR) as a structured approach to mitigate mode collapse and enhance solution diversity. The methodology is grounded in the structure-mapping theory from cognitive science, which defines analogies as mappings of shared relational structures between a base domain and a target domain, rather than superficial object similarities.

The AR pipeline operates as a two-step process:

Extraction: Given a research problem $p$ , an LLM agent extracts the set of objects ( $O_p$ ) and relations ( $R_p$ ) inherent to the problem. It then generates analogies to cross-domain research problems that share similar relational structures. These analogies are formalized as object mappings ( $M$ ) between the problem domain and an analogous domain, and a set of shared relations ( $R_a$ ).
Search: A second LLM agent utilizes the generated analogy (specifically the object mappings and shared relations) to search for existing solutions, algorithms, or methods within the analogous domain. The agent is prompted to find solutions that address the shared relational structure and can be transferred back to the original biomedical problem.

This approach is compared against two baselines:

No-domain Baseline: Directly prompts the LLM to find solutions without domain constraints.
Cross-domain Baseline: Prompts the LLM to find solutions from other domains but without explicitly generating the intermediate analogical structure (object mappings and shared relations).

The evaluation utilizes 50 research problems sampled from a curated "AR Dataset" of 266 biomedical papers that utilized cross-domain analogies. The study employs three LLMs (Claude Sonnet 4.5, GPT-5.2, and Gemini 3 Flash) and evaluates outputs on three axes: generation diversity, solution novelty, and analogy quality.

Key Contributions

Quantification of Mode Collapse: The paper provides the first systematic quantification of mode collapse in open-ended scientific solution generation, demonstrating that baseline LLMs produce low-diversity outputs that converge on a narrow set of high-probability solutions.
Analogical Reasoning Framework: The authors propose and implement AR as a specific mechanism to expand the solution search space. By forcing the model to explicitly map relational structures across disparate domains, AR guides the search toward novel, non-obvious solutions.
Human and Automated Validation: The study validates the novelty and quality of AR-generated solutions through both LLM-judged metrics and human annotation studies, showing strong alignment between automated scores and expert judgment.
Real-World Feasibility: The authors demonstrate the practical utility of AR by implementing four distinct biomedical case studies, showing that AR-generated approaches can achieve state-of-the-art or competitive performance on real-world tasks.

Results

Diversity and Novelty

Mitigation of Mode Collapse: AR significantly outperforms baselines in generation diversity. It improves the Domain Vendi Score (a metric for semantic diversity) by 100–115% and the Solution Vendi Score by 90–173% compared to baselines.
Novelty: AR generates novel solutions over 50% of the time (ranging from 50.4% to 68.9% depending on the LLM), whereas the no-domain baseline rarely finds novel solutions (as low as 1.6%). In a human pairwise preference study, AR solutions were rated as more novel than cross-domain baseline solutions 78% of the time.
Analogy Quality: AR generates analogies with significantly higher Domain Distance (3x higher than the no-domain baseline) and Novelty scores, indicating that AR successfully bridges highly disparate fields (e.g., systems biology and economics, drug design and chess tactics).

Case Study Performance

The authors implemented AR-generated solutions for four biomedical problems, yielding consistent quantitative gains:

Perturbation Effect Prediction: By mapping systems biology to economics, AR proposed using Finite Mixture Models (FMM). This approach achieved a nearly 13-fold improvement on the MMD PCA distributional metric (reducing the score from 1.65 to 0.13) compared to the Latent Additive (LA) baseline, better capturing single-cell heterogeneity.
Cell-Cell Communication Inference: Mapping telecommunications to biology, AR proposed a Signal-to-Noise Ratio (SNR) analysis. This method achieved state-of-the-art performance on the AUPRC metric (0.248) for the OpenProblems ligand-target benchmark, outperforming 14 existing baselines.
Brain Region Interaction: Mapping communication networks to neuroscience, AR identified the PCMCI (Peter and Clark Momentary Conditional Independence) method. This approach successfully isolated cross-hemisphere coupling signals with a high Spearman correlation ( $\rho=0.729$ ) to published methods and comparable predictive power for behavioral outcomes.
Oligonucleotide Property Prediction: Mapping chess tactics to drug design, AR proposed Piece-Square Tables (PST) with tapered evaluation. This method established state-of-the-art performance on 2 of 6 datasets for oligonucleotide property prediction, improving upon linear baselines by encoding position-dependent effects.

Significance

The paper positions Analogical Reasoning as a critical "diversity engine" for the future of autonomous science. The authors argue that as AI systems take on more of the scientific pipeline, frameworks must evolve to avoid converging on canonical solutions. AR serves to augment the search space of existing solution generation methods by leveraging the latent ability of LLMs to perform cross-domain reasoning.

The work demonstrates that AR is not only theoretically sound but practically feasible, producing high-quality, novel solutions that can be implemented and validated in real-world biomedical contexts. The authors note that the approach is computationally inexpensive (approximately $0.02 per problem), making it scalable. While the current evaluation is scoped to biomedical problems, the framework is domain-agnostic, suggesting broad applicability across scientific disciplines. The paper concludes that AR represents a vital step toward bridging isolated domains to drive scientific innovation, provided that rigorous safety and feasibility evaluations accompany the generation of novel hypotheses.