Overcoming Surface Feature Distraction in Analogical Reasoning

16 concepts9 papers

When humans encounter an analogy like "a camera is to an eye what a microprocessor is to a brain," we instinctively strip away the surface details. We do not get confused by the fact that eyes are biological tissue while cameras are made of glass and silicon. Instead, we perform a structural mapping: we isolate the underlying functional relationships (capturing light, processing signals) and map them across domains.

Current machine learning models—ranging from Large Language Models (LLMs) to Knowledge Graph Embeddings (KGEs)—struggle deeply with this type of abstraction. Instead of mapping deep relational structures, they exhibit a pervasive bias toward surface-level attributes. They are easily distracted by shared vocabulary, specific entity identities, geographic locations, or superficial visual traits, creating a significant bottleneck in their ability to reason.

The Statistical Association Trap

Why do state-of-the-art models fail at abstract mapping? The root cause lies in how they are trained. Deep learning architectures excel at capturing statistical co-occurrences and memorizing explicit patterns, which frequently leads them to mistake literal similarity for logical analogy.

In narrative reasoning, this manifests as a dependency on lexical overlap. As Cheng Jiayang et al. demonstrate using the StoryAnalogy dataset, when LLMs are asked to identify similarities between long-form stories, they are consistently distracted by shared vocabulary or identical entities. If two stories take place in the ocean and feature fish, a model will confidently declare them analogous, even if their underlying plot progressions and causal structures are entirely different.

This same failure mode cripples models trained on structured data like Knowledge Graphs (KGs). Rather than learning abstract rules of interaction that can apply to any entity, models often memorize specific entity IDs. Pengyang Wang et al. observe that when temporal knowledge graph models prioritize entity identity over relational semantics, they fail to generalize to novel contexts or unfamiliar entity pairs. Similarly, Hao Zhou et al. argue that this focus on memorizing lexical facts instead of structural logic acts as a barrier, preventing conversational agents from applying known relationship types to new subjects.

The Spectrum of Distraction: Near vs. Far Analogies

The challenge becomes clearer when we look at the distance between the domains being compared. Zhivar Sourati et al. formalize this by dividing analogies into "near" and "far" categories. Models perform reasonably well on near analogies, where the surface features and the relational structures happen to align. However, when evaluating far analogies—where the surface features diverge entirely but the underlying logic remains identical—performance plummets because specific actions and settings act as powerful distractors.

[High Surface Similarity]  -->  Near Analogy: "A doctor treating a patient is like a nurse treating a patient."
                                (Models succeed due to lexical overlap)

[Low Surface Similarity]   -->  Far Analogy:  "An atom's electrons orbiting a nucleus is like planets orbiting a sun."
                                (Models fail; distracted by domain divergence)

This surface-level distraction manifests across several distinct domains:

  • Scientific Systems: Siyu Yuan et al. show that LLMs fail to recognize scientific analogies when the lexical overlap between components is low. Because words like "camera" and "eye" share minimal statistical association in general text corpora compared to domain-specific terms, models struggle to map their shared functional roles.
  • Historical Foresight: In cross-temporal analysis, models fall into geographic and actor-based stereotyping. Nianqi Li et al. and Yongqiang Chen et al. independently find that automated research agents tend to match historical events based on superficial features like shared geographic regions or time periods. An agent tasked with finding a parallel to a modern political crisis will look at past events in the same country, completely missing a structurally identical causal mechanism that occurred on another continent.
  • Multimodal Puzzles: The bias extends beyond text. In multimodal knowledge graphs, Ningyu Zhang et al. demonstrate that models prioritize superficial visual or textual attributes over abstract logic, failing to solve cross-modal puzzles when the visual appearance of an object contradicts its relational role.

Breaking the Bottleneck

To overcome this relational bottleneck, researchers are developing frameworks that explicitly force models to separate surface attributes from underlying structures.

Entity-Agnostic Representations

If memorizing specific entities causes models to fail in new contexts, one solution is to hide the entity identities entirely during representation learning. The EARL framework, introduced by Hao Zhou et al., shifts the model's focus away from specific nodes in a knowledge graph and onto the abstract relational paths connecting them. By learning entity-agnostic representations, the system becomes capable of transferring its logical reasoning to entirely unfamiliar subjects.

Structure Abduction and Causal Decomposition

Another powerful strategy is to introduce an explicit abstraction step before allowing a model to make a comparison. Instead of feeding raw text into a model and asking for an analogy, frameworks can force the model to map out the system's architecture first.

For scientific analogies, Siyu Yuan et al. developed the SCAR benchmark, which utilizes "Analogical Structure Abduction." This technique requires the model to explicitly isolate and explain the functional roles of a system's components before attempting a mapping.

For historical analysis, the CANA agent framework designed by Yongqiang Chen et al. uses causal decomposition to strip away geographic and temporal details, breaking an event down into its core causal drivers before retrieving historical parallels.

Raw Scenario  -->  [ Causal Decomposition / Abduction ]  -->  Abstract Causal Graph  -->  Accurate Structural Analogy
(e.g., Crisis in X)     (Strips away geography and actors)          (Pure relational logic)        (Matches Crisis in Y)

Multi-Stage Reasoning

Finally, breaking the evaluation process down into discrete, verifiable steps prevents models from making snap judgments based on surface features. Xiaojing Zhang et al. demonstrate this with the A3E framework for analogy annotation. By forcing the model to explicitly separate object attributes from causal relationships across multiple stages, they successfully bypass the relational bottleneck, allowing models to achieve human-level accuracy in identifying true narrative analogies.

Key Foundational Work

StoryAnalogy: Deriving Story-level Analogies from Large Language Models to Unlock Analogical Understanding (2023) — contributed 3 concept elements to this direction.
Beneath Surface Similarity: Large Language Models Make Reasonable Scientific Analogies after Structure Abduction (2023) — contributed 3 concept elements to this direction.
Multimodal Analogical Reasoning over Knowledge Graphs (2023) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_superficial_similarity_bias_77ea4be7,
  title = {superficial similarity bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/superficial-similarity-bias/rdr_77ea4be7274a411eb61ce516ec579709},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_superficial_similarity_bias_77ea4be7_v1,
  title = {superficial similarity bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/superficial-similarity-bias/rdr_77ea4be7274a411eb61ce516ec579709/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake