Options need discriminable differences

What it is

Three “professional yet friendly” replies, differing by a comma and “Hi” versus “Hello.” People cannot name a difference on any decision dimension, yet they are asked to pick one. A set without discriminable differences is not multiple options; it is surface variants of one option. Option distinctness requires candidates to split on the attributes the user actually uses to decide, not to jitter in wording.

Comparison collapsing into punctuation-picking is what happens after distinctness has already failed. The question here is whether the set counts as options at all.

Why it happens

Choice depends on attribute contrast. People first ask “where do these differ,” then trade on that dimension. If samples come from a peaked distribution, the paths crowd semantically: different tokens, same decision attributes. Raising temperature increases lexical jitter; it does not necessarily increase distance on the dimension that matters (tone, structure, strength of commitment, compositional focus).

An indiscriminable set does two kinds of harm. One is a false choice: a click with no information. The other trains a wrong expectation: next time “generate a few more” is still assumed to open the space, when it only jitters around the same point.

Studying it

From one prompt, generate n options; measure embedding distance, human ratings of “are these different options,” and whether users can name the difference in one sentence. Independent variables: decoding temperature, explicit diversity constraints, forced split on a named dimension (formal/informal, short/long). Dependent variables: rate of nameable differences, time to choose, whether after choosing people can recap how the picked item differed from the others.

Nameable difference is the key measure. Large embedding distance with “pretty much the same” means the split was not on a decision dimension. Automatic diversity scores alone will count wording jitter as success.

Where it stops holding

When the user wants a tweak (“a bit shorter”), near variants are the feature, not a defect. Near-duplicates of a closed answer are dangerous, because people read meaning into a difference that carries none. In image generation, “same composition, different palette” can be distinct because colour is a decision dimension; copy has no equivalent of a palette swap. Screen-reader users hearing three near-paraphrases get even less distinctness than visual scan; the dimension split has to be spoken in the option title first.

Applying it

• When generating several options, force a split on the task’s decision dimensions: length, tone, structure, commitment — not a temperature bump alone.
• Label each option with one short sentence of how it differs from the others. If people cannot name a difference, do not hand the set over as multiple options.
• Detect near-duplicates: merge or redraw items with high semantic overlap; do not juxtapose them.
• Check: after generation, ask “how do these differ.” A set that yields no decision-dimension answer (only “the wording is a bit different”) does not count as multiple options.

Related

• Same group: L3.01.1 Multiple options turn generation into a choice · L3.01.2 Option count beyond comparison capacity is wasted
• Nearby: L3.07 Multi-option Generation and Side-by-side Comparison · L1.01 Mismatch Between Stochastic Output and Deterministic UI
• Search terms: option distinctness · discriminable alternatives · near-duplicate candidates