Does adding multiple interpretations to ambiguous situations respect language more than resolving them?
This explores whether holding several readings of an ambiguous sentence open—rather than collapsing it to one 'correct' meaning—is actually more faithful to how language works, and whether AI systems are built to do that.
This explores whether multiplying interpretations honors language better than resolving it—and the corpus suggests the multiplicity isn't a bug to be cleaned up, it's a property of meaning itself. The strongest case comes from interpretation modeling, which finds that when readers disagree about a socially loaded sentence, the disagreement is *valid information*, not annotation noise to be averaged away Why do readers interpret the same sentence so differently?. Meaning shifts with the reader's social position, so a single resolved answer actually discards signal. On this view, preserving the spread of readings respects language more than forcing a verdict.
The trouble is that today's models are built to resolve, and they're bad even at that. GPT-4 correctly disambiguates only 32% of deliberately ambiguous cases against 90% for humans, failing across lexical, structural, and scope ambiguity Can language models recognize when text is deliberately ambiguous?. The diagnosis is revealing: LLMs can't hold multiple interpretations in mind *at once*. So the question isn't just philosophical—models lack the very capacity that 'respecting' ambiguity would require. A related failure shows the same reflex toward premature commitment: models accept false presuppositions baked into a question even when they demonstrably know the facts, accommodating a single framing rather than challenging it Why do language models accept false assumptions they know are wrong?.
Where the corpus gets interesting is in the methods that try to *manufacture* multiplicity rather than suppress it. A leader-follower debate protocol pushes a small model from poor ambiguity detection up to 76.7% by having one agent propose interpretations and two others challenge them with rotating roles Can structured debate roles help small models detect ambiguity?. The mechanism is exactly the move the question gestures at: keep competing readings alive and in tension instead of snapping to the first plausible one. Multiplicity here is a verification strategy, not a failure to decide.
But 'respecting' ambiguity isn't the same as never resolving—it's resolving at the right time, collaboratively. Collaborative rational speech acts model dialogue as two speakers progressively building shared understanding across turns, tracking both sides' beliefs as they move from partial to common ground Can dialogue systems track both speakers' beliefs across turns?. That reframes the whole dichotomy: language doesn't ask you to choose between holding all readings open forever and resolving instantly. It resolves *through interaction*, narrowing as the conversation supplies context. The multiple interpretations are the starting material; pragmatic exchange is what earns the resolution.
So the honest answer is layered. Adding interpretations respects something real that resolution erases—the reader-dependence and social embedding of meaning Why do readers interpret the same sentence so differently?. It also exposes how shallow current 'understanding' is, since grounding in LLMs is uneven and partial rather than the kind that would let a system genuinely weigh rival readings Does semantic grounding in language models come in degrees?. The thing you didn't know you wanted to know: the best systems don't pick between multiplying and resolving—they hold interpretations open precisely so they can resolve more carefully, the way a good conversation does.
Sources 6 notes
Interpretation Modeling research shows that disagreement on socially embedded sentences reflects valid differences in reader perspective, not annotation failure. Structured human disagreement in NLI benchmarks confirms that interpretation distributions carry meaningful information.
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.
The FLEX Benchmark shows that models reject false presuppositions at rates far below acceptable levels (GPT-4: 84%, Mistral: 2.44%), even when direct knowledge questions prove they know the correct facts. False presuppositions drive more accommodation than correct knowledge drives rejection.
Mistral-7B achieved 76.7% accuracy in ambiguity detection through a protocol where a leader proposes interpretations and two followers challenge them with rotating roles. Role rotation and consensus forcing prevent persuasive framing failures and create stronger verification than pairwise debate.
CRSA integrates rate-distortion theory with RSA to enable bidirectional belief tracking across dialogue turns. Demonstrated on referential games and doctor-patient dialogues, it captures progression from partial to shared understanding, providing the information-theoretic framework that token-level LLM systems lack.
Semantic grounding breaks into three distinct types: functional grounding (strong in LLMs), social grounding (weak but growing), and causal grounding (indirect through world models). LLMs score differently on each dimension, making the yes-or-no understanding question misleading.