When an AI judge gives two candidates the same score, picking a winner is pure luck — not quality.

This explores why tie-breaking matters when a system has to pick one winner from many candidates or convert AI preferences into a reward signal. The short version: best-of-N and RLAIF both live or die on the discrimination power of their scorer, and ties are exactly where that scorer goes blind. If a judge emits a discrete score — a token like "7" or a thumbs-up — many genuinely different candidates collapse to the same value, and the selection step is reduced to a coin flip. The corpus's sharpest answer here is to stop reading the judge's chosen token and instead read the full probability distribution over scoring tokens: taking the expectation over those logits yields a continuous score that pulls apart candidates a discrete verdict would have called equal Can reading logit distributions break ties in LLM judging?. No retraining, no extra model — just refusing to throw away the information already sitting in the logits.

Why this matters more than it sounds: best-of-N's entire value proposition is that a good ranker can find the one strong sample among N mediocre ones. But that advantage is fragile. One note finds that plain majority voting is often *more* robust than best-of-N precisely because best-of-N leans on verifiers that are unreliable and models that assess themselves poorly — when the scorer can't discriminate, complexity buys you nothing Why does majority voting outperform more complex inference methods?. Tie elimination is the missing ingredient that could make best-of-N actually beat its simpler cousin: a verifier that reliably separates candidates is what best-of-N needs and majority voting doesn't have.

For RLAIF the stakes compound over training. The pipeline turns pairwise or scored preferences into gradients, and a flood of ties means a flood of near-zero-signal comparisons — the reward model can't tell the training loop which behavior to reinforce. This connects to a broader failure the corpus keeps flagging: scorers that collapse distinctions push systems toward degenerate equilibria. A ranking system that doesn't model its own selection bias converges on amplifying its past decisions Why do ranking systems need to model selection bias explicitly?; a reward signal that can't distinguish good from slightly-better does an analogous thing — it flattens the gradient landscape the optimizer is trying to climb.

There's a subtler, laterally-connected reason too: aggregation methods need *raw material to arbitrate*, not confirmation. Set-based RL finds that diverse mediocre traces beat redundant expert ones because the aggregator needs contrast to work with Can diverse mediocre traces outperform redundant expert traces?, and meta-reasoning over multiple chains beats majority voting because it mines the losing chains instead of discarding them Does voting discard useful reasoning from losing chains?. Ties are the enemy of all of this: a scorer that can't discriminate erases the very contrasts these methods feed on. Tie elimination isn't a cleanup step — it's what restores the signal that selection and reward-learning are supposed to run on.

The thing worth carrying away: the fix is usually not a better or bigger judge but a richer *readout* of the judge you already have. The discrimination was there in the distribution the whole time; discrete scoring was quietly discarding it.