The Semantic Dependency of Large Language Model Reasoning

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2172 papers266 conceptsfrozen 2026-10-04

Large Language Models (LLMs) often appear to possess sophisticated reasoning capabilities, solving complex word problems and passing professional exams. However, a growing body of research suggests that this "reasoning" is often a form of sophisticated pattern matching tethered to natural language semantics rather than an abstract understanding of logical rules. When the familiar linguistic "crutches" are removed—a process known as semantic obfuscation—the performance of even the most advanced models frequently collapses.

The Performance Cliff of Obfuscation

The core insight across this research is the "performance cliff" that occurs when symbolic logic is decoupled from meaningful identifiers. In a standard reasoning task, an LLM might successfully navigate a logic puzzle about "Alice" and "Bob" or a planning task involving "stacking blocks." However, as Karthik Valmeekam, Subbarao Kambhampati, and colleagues demonstrate in PlanBench, if these common-sense labels are replaced with abstract symbols or misleading terms, the model's ability to solve the underlying problem often drops to near-zero.

This phenomenon suggests that LLMs rely on semantic anchors—the world knowledge and statistical associations tied to words like "stack," "open," or "addition"—to navigate logical constraints. Lin Ren and Guilin Qi found similar results in ASPBench, noting that models struggle with Answer Set Programming (ASP) when meaningful predicate names are replaced with abstract identifiers. The models are not executing the logic; they are retrieving patterns that "look like" solutions they have seen in their training data.

Why Semantic Cues Mask Reasoning Failures

The reason LLMs fail under obfuscation is rooted in their architecture as "System 1" thinkers—engines of intuitive, fast pattern recognition. They lack the "System 2" deliberative reasoning required for formal logic. Subbarao Kambhampati argues in LLMs Can't Plan, But Can Help Planning in LLM-Modulo Frameworks that models substitute genuine logical deduction with approximate retrieval.

When a model sees a math problem, it doesn't "calculate" so much as it "recalls" the shape of a solution. Antara Raaghavi Bhattacharya and Isabel Papadimitriou show that even advanced models like o1-mini fail to solve number puzzles when standard operators like "+" are replaced with unfamiliar words, indicating that their computational abilities are triggered by specific tokens rather than abstract operations. Similarly, Yue Hao and ChengXiang Zhai found in PutnamGAP that changing simple variable names in competition-level math problems causes significant performance drops, revealing a lack of symbolic invariance.

The Limits of In-Context Learning

A common counter-argument is that LLMs can "learn" new rules through in-context learning (ICL). However, Qixun Wang and Yifei Wang suggest that ICL is actually a form of algorithm selection. The model isn't learning a new logic; it is searching its pre-training data for a known function that fits the provided examples. If the task is truly out-of-distribution—such as a spatial reasoning task using nonsense words for directions, as explored by Lachlan McPheat in DecompSR—the model's "reasoning wall" becomes apparent.

Bridging the Gap: Neuro-Symbolic Solutions

To overcome these limitations, researchers are moving toward hybrid architectures. Instead of forcing LLMs to be autonomous reasoners, methods like LLM-Modulo or LLMFP (LLM-Based Formalized Programming) use the LLM as a "formalizer."

As Cassie Huang and Yilun Hao demonstrate, LLMs are excellent at translating natural language into structured formats like PDDL or SMT code. This code can then be executed by a symbolic solver (like Z3) that possesses the "System 2" rigor the LLM lacks. Another approach, proposed by Yiming Wang in Meta-Reasoning, involves explicitly "deconstructing" problems into symbolic skeletons (e.g., mapping names to variables like $A$ and $B$) before reasoning. This strips away the semantic noise that often distracts the model, forcing it to focus on the underlying logical structure.

Conclusion

The evidence from benchmarks like ASPBench, PlanBench, and PutnamGAP points to a consistent conclusion: current LLMs are "causal parrots" Shuiwang Ji. Their apparent intelligence is often a reflection of the semantic richness of the prompts we give them. True abstract reasoning—the ability to apply logic regardless of the symbols used—remains a significant frontier for the field.

Papers in this version

Showing 60 of 172 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 116 papers
  2. Version 1 · 2026-10-04 · 172 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_semantic_dependency_in_reasoning_14b5df66,
  title = {semantic dependency in reasoning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-dependency-in-reasoning/rdr_14b5df660e704e3d9a030f4daad1c066},
  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_the_semantic_dependency_of_large_languag_14b5df66_v1,
  title = {The Semantic Dependency of Large Language Model Reasoning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-semantic-dependency-of-large-language-model-reasoning/rdr_14b5df660e704e3d9a030f4daad1c066/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}