Why does rewriting an unclear question to fix it still trip up AI, even when the fix is small?
Why do minimal-edit disambiguations diverge from natural language patterns?
This explores why rewriting an ambiguous question with the smallest possible change to make it clear can still produce text that looks unusual to a language model, and why that unusualness affects how well the model answers.
This explores why a 'minimal-edit' fix for an ambiguous question, where you change just a word or two so it has only one meaning, can produce text that reads oddly compared with how people normally write, and why that matters to an LLM. The most direct evidence is that LLMs answer more accurately after a short clarifying exchange (a question, then an answer) than after a minimally edited version of the same ambiguous input Do clarifying dialogues help LLMs answer better than rewritten inputs?. The explanation is about distribution. People rarely resolve ambiguity by surgically rewriting a sentence. They ask 'which one did you mean?' and get a reply. The clarifying dialogue looks like what the model saw during pretraining. The minimal edit is a constrained, artificial piece of text that keeps the original wording almost intact while bending its meaning, and that kind of text is less common.
This connects to a broader pattern: models do better on familiar phrasings than on rare ones that mean the same thing. Across math, translation, commonsense reasoning and tool calling, LLMs reliably perform better on high-frequency paraphrases than on equivalent rare ones Do language models really understand meaning or just surface frequency?. A minimal-edit disambiguation is effectively a forced rare paraphrase. It is correct in meaning but statistically unusual, so the model gets less help from it than its logical clarity would suggest. A related problem appears with context: when strong associations from training pull one way, models can ignore what the prompt actually says Why do language models ignore information in their context?. A small edit that changes a sentence's reading may simply be too weak to override the familiar interpretation the model has already settled on.
The ambiguity side makes the difference larger. Models are poor at noticing ambiguity in the first place. On the AMBIENT benchmark, GPT-4 disambiguated correctly only 32% of the time, compared with 90% for humans, and the failures covered word-level, sentence-structure and scope ambiguity Can language models recognize when text is deliberately ambiguous?. If a model doesn't keep several readings in mind, a subtle one-word change that picks one reading is easy to miss. A separate finding shows LLM errors growing predictably as syntax gets more deeply nested Why do large language models fail at complex linguistic tasks?. Minimal edits often settle ambiguity through exactly this kind of structural detail, such as attachment or clause boundaries, rather than by saying it outright.
The surprising takeaway is that 'cleaner' input is not always better input for an LLM. A precise, minimal rewrite looks like the most efficient way to remove ambiguity, but a slightly wordier, conversational clarification often works better because it matches how language is normally used. This also has a design implication. Systems that ask a quick follow-up question may beat systems that silently rewrite the user's query. The library has one study directly on this comparison, so treat it as a strong lead rather than a settled rule.
Sources 5 notes
LLMs answer more accurately when given clarifying questions and answers than when given minimally-edited disambiguated versions of the same ambiguous input. The gap exists because clarifying exchanges match the model's pretraining distribution better than artificially constrained rewrites.
LLMs show consistent preference for higher-frequency surface forms over semantically equivalent rare paraphrases across math, machine translation, commonsense reasoning, and tool calling. This suggests models track statistical mass from pretraining rather than meaning-recognition as their primary mechanism.
Research demonstrates that LMs generate outputs inconsistent with their context because parametric knowledge from training dominates over in-context information. Textual prompting alone cannot override strong priors; causal intervention in representations is required.
AMBIENT benchmark shows GPT-4 correctly disambiguates only 32% of cases versus 90% for humans. This failure spans lexical, structural, and scope ambiguity—revealing that LLMs cannot hold multiple interpretations simultaneously, a fundamental gap hidden by standard benchmarks.
Top-tier LLMs like Llama3-70b consistently misidentify embedded clauses, verb phrases, and complex nominals. Performance degrades predictably as syntactic depth increases, revealing that statistical learning captures surface patterns but not deep grammatical rules.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Simple Linguistic Inferences of Large Language Models (LLMs): Blind Spots and Blinds
- Beyond Accuracy: Evaluating the Reasoning Behavior of Large Language Models -- A Survey
- Knowing but Not Showing: LLMs Recognize Ambiguity but Rarely Ask Clarifying Questions
- Clarify When Necessary: Resolving Ambiguity Through Interaction with LMs
- Word Meanings in Transformer Language Models
- Language models show human-like content effects on reasoning tasks
- Aligning Language Models to Explicitly Handle Ambiguity
- Linguistic Blind Spots of Large Language Models