Surface Form Sensitivity in Language Model Reasoning

41 concepts31 papers1 proposal

Consider a simple logical premise: If the box is under the table, it is not on top of the table. A state-of-the-art large language model (LLM) might effortlessly confirm this. Yet, if you rephrase the query slightly—perhaps changing "under" to "beneath," swapping the order of the clauses, or framing it as a yes/no question—the model's reasoning can completely fall apart.

This phenomenon is known as surface form sensitivity. It exposes a fundamental "consistency gap" in modern natural language processing: while models achieve high accuracy on isolated benchmarks, they lack semantic robustness. They fail to maintain logical consistency across inputs that are semantically identical but syntactically or lexically distinct. This gap suggests that instead of building a stable internal world model or executing abstract logic, LLMs frequently rely on local, surface-level statistical pattern matching.

The Anatomy of the Consistency Gap

Why does a model's reasoning path change when the underlying logic remains invariant? To understand this, we must look at how models process text. Humans decouple formal logic from linguistic expression; an argument remains valid whether it is stated directly, negated via antonyms, or paraphrased. LLMs, however, are trained on next-token prediction over raw text, meaning their internal representations are deeply intertwined with specific lexical triggers.

As demonstrated by Myeongjun Jang et al. in their Consistency Analysis of ChatGPT, models frequently fail to maintain logical stability when presented with semantically equivalent inputs or minor structural changes. This brittleness is not a minor edge case; it represents a systemic failure to generalize abstract principles. When a model encounters a problem, it maps the surface-level linguistic features directly to an answer space. If a benign perturbation shifts those features outside the model's high-probability training distribution, the reasoning chain breaks.

Even when models appear to have mastered a benchmark, dynamic variations reveal they are mapping specific linguistic patterns to answers rather than constructing a stable internal representation of the situation. For instance, Kyle Richardson 0001 et al. show in Dyna-bAbI: unlocking bAbI's potential with dynamic synthetic benchmarking that switching a query from a location question to a yes-no question causes performance to degrade, indicating a failure to decouple the underlying logic from its specific linguistic form.

This brittleness extends down to fundamental conceptual links. Meriem Beloucif et al. show in Probing Pre-trained Language Models for Semantic Attributes and their Values that simply changing a proper noun in a prompt can lead a model to predict entirely unrelated attributes, demonstrating a failure to internalize invariant conceptual links. To quantify this behavior, the authors of How often are errors in natural language reasoning due to paraphrastic variability? introduce the metric of Paraphrastic Consistency, finding that while massive pretraining helps models stay consistent, task-specific fine-tuning can actually make them more sensitive to minor changes in wording.

Exposing the Flaw Across Domains

Researchers have designed diverse stress tests to diagnose this sensitivity across different reasoning domains, consistently finding that high performance on standard benchmarks is an illusion maintained by familiar prompt templates.

Commonsense and Physical Reality

In commonsense reasoning, models often exploit statistical shortcuts rather than understanding physical constraints. Shikhar Singh et al. introduced the COM2SENSE benchmark, using complementary true/false statements to show that models trail human consistency by over 40% when evaluating nearly identical scenarios. Similarly, Pei Zhou et al. developed RICA to evaluate robust inference based on commonsense axioms, finding that performance drops to near-random levels when a premise undergoes simple linguistic transformations like negation or paraphrasing.

This is echoed by Xuhui Zhou et al. in Evaluating Commonsense in Pre-Trained Language Models, where swapping a single word for its antonym causes models to fail, and by Chang Shu et al. with the POSQA dataset, which highlights how size-comparison reasoning fluctuates wildly based on prompt formatting rather than a stable world model. This pattern-dependency is further exposed when models face puzzles that require breaking standard associations. As Filip Ilievski et al. demonstrate in BRAINTEASER: Lateral Thinking Puzzles for Large Language Models, language models frequently fail to maintain reasoning consistency when the surface-level phrasing of a lateral-thinking puzzle is altered while keeping the underlying logic identical.

Mathematical and Symbolic Logic

One might expect mathematics and formal logic to be immune to surface variations, given their rigid underlying rules. However, Pan Lu et al. note in A Survey of Deep Learning for Mathematical Reasoning that models succeeding on complex math problems can fail simple arithmetic when the phrasing of a word problem is slightly altered. Yada Zhu explicitly demonstrates this in Paraphrase and Solve: Exploring and Exploiting the Impact of Surface Form on Mathematical Reasoning in Large Language Models, showing significant performance fluctuations based purely on wording. To automate this diagnosis, Yutao Hou et al. created AR-CHECKER, a framework that stress-tests LLMs by rewriting math problems into semantically identical variants, causing accuracy to plummet. Similarly, an ontology-guided stress test in Evaluating LLMs' Mathematical and Coding Competency through Ontology-guided Interventions reveals that model performance is often tied to specific problem framings rather than the underlying logical principles, failing when presented with changes in entity names or output constraints.

In formal logic, Soumya Sanyal developed RobustLR, revealing that deductive reasoners fail when faced with minimal logical perturbations or equivalent rule rewritings. This lack of formal invariance is further illustrated by Magdalena Wysocka et al. in SylloBio-NLI, where biomedical syllogistic reasoning exhibits extreme sensitivity to lexical structure and phrasing. Furthermore, Fangzhi Xu et al. note in PathReasoner: Modeling Reasoning Path with Equivalent Extension for Logical Question Answering that models frequently produce inconsistent outputs when the phrasing of logical connectives is changed, even if the underlying logic remains identical.

The Mechanics of Linguistic Operators

The sensitivity becomes particularly acute when dealing with operators like negation, conditionals, and spatial relations.

In CONDAQA, Matt Gardner 0001 et al. highlight a massive consistency gap regarding negation: models might correctly infer the meaning of a negated statement in one context but fail entirely when the negation scope is slightly altered or paraphrased.

[Standard Form] "The patient is not asymptomatic." 
     └── Model Inference: Patient has symptoms. (Correct)

[Paraphrased Form] "It is untrue that the patient lacks symptoms."
     └── Model Inference: Patient has no symptoms. (Incorrect)

Similarly, Matthew Mandelkern et al. explore this in Conditional and Modal Reasoning in Large Language Models, showing that models fail to recognize basic logical equivalences (such as the equivalence between "might not" and "not must"), leading to divergent reasoning paths for identical premises. Spatial prepositions also trigger this brittleness; Iulia M. Comsa et al. demonstrate in A Benchmark for Reasoning with Spatial Prepositions that minor variations in input formatting or phrasing disrupt a model's ability to track spatial relations.

Why Chain-of-Thought Fails to Insulate Reasoning

A common remedy for poor reasoning is Chain-of-Thought (CoT) prompting, which encourages models to generate step-by-step rationales. However, the internal monologue itself is highly sensitive to surface form.

Alexander von Recum et al. investigated this in Are Reasoning LLMs Robust to Interventions on their Chain-of-Thought?. They discovered a lack of style invariance: if a model's internal monologue is forced into a benignly paraphrased or unfamiliar style, it suppresses the precise linguistic cues the model needs to self-correct or identify previous errors. The reasoning process is not guided by pure logical state transitions, but is instead tethered to the specific phrasing and formatting of its own generated text.

Turning Sensitivity into a Solution

If models are highly sensitive to surface form, how can we build systems that achieve true semantic robustness? Researchers are attacking this problem from two angles: enforcing invariance during training, and exploiting variation during inference.

Enforcing Invariance via Training

To teach models that structural symmetries should yield identical conclusions, Yihang Yao et al. introduced syMmetry-ENhanceD (MEND) Data Augmentation in Your Language Model May Think Too Rigidly: Achieving Reasoning Consistency with Symmetry-Enhanced Training. By exposing models to shuffled premise orders and irrelevant redundant information during post-training, MEND forces the network to decouple the underlying logic from surface-level presentation, significantly improving out-of-distribution generalization.

Exploiting Variation via Inference

Rather than fighting surface form sensitivity, we can leverage it as a feature. Yada Zhu developed Self-Consistency-over-Paraphrases (SCoP) in Paraphrase and Solve. Instead of relying on a single prompt, SCoP generates multiple paraphrased versions of a mathematical problem, samples the model's reasoning paths across all versions, and aggregates the answers.

                  ┌── Paraphrase A ──> Reasoning Path 1 ──> Answer: 42
                  │
[Original Prompt] ├── Paraphrase B ──> Reasoning Path 2 ──> Answer: 42
                  │
                  └── Paraphrase C ──> Reasoning Path 3 ──> Answer: 14 (Brittle Failure)
                                            │
                                    [Majority Vote] ──> Final Answer: 42

Because different surface forms trigger different internal paths, aggregating them smooths out the brittle failures of individual phrasings. This transforms a model's linguistic vulnerability into a mechanism for robust error correction, proving that understanding the quirks of surface form sensitivity is key to unlocking more reliable AI.

Key Foundational Work

RICA: Evaluating Robust Inference Capabilities Based on Commonsense Axioms (2021) — contributed 3 concept elements to this direction.
Evaluating Commonsense in Pre-Trained Language Models (2020) — contributed 1 concept elements to this direction.
Probing Pre-trained Language Models for Semantic Attributes and their Values (2021) — 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_surface_form_sensitivity_9bb10112,
  title = {surface form sensitivity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/surface-form-sensitivity/rdr_9bb101127cac4961b7608236eb3a83de},
  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_surface_form_sensitivity_9bb10112_v1,
  title = {surface form sensitivity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/surface-form-sensitivity/rdr_9bb101127cac4961b7608236eb3a83de/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake