Ask a system for "the name of the fourth tab" when the page has only three, and watch what it does. Most invent a fourth — confidently, plausibly, wrong. A few emit a bare "unknown." Both are failures, and on LongMemEval-V2 they belong to the worst-scoring category there is: abstention. The production baseline gets it right 20.3% of the time. It's the part of memory nobody optimizes, because saying "there is nothing there" is harder than answering.
Grounded refusal, not a bigger shrug.
The wrong fix is a stronger "I don't know" reflex — that just wrecks the questions that do have answers. The right fix has to be grounded: refuse only when the record genuinely lacks what's being asked, and do it without degrading the questions that have answers. The honest answer these questions reward looks less like a shrug and more like this:
"There is no fourth tab there — the page shows only three."
A refusal with the receipt in hand: not silence, not a hedge. How we get a reader to produce that reliably is our own method, and it stays sealed. What we can show is what it does to the number.
The result, on the benchmark's own scorer.
Scored with LongMemEval-V2's official abstention checker, on the 128 false-premise questions:
| Reader | Abstention correct |
|---|---|
| production baseline | 20.3% |
| + grounded absence (GPT-4o) | 23.4% |
| + grounded absence (GPT-5) | 38.3% |
Nearly double the baseline. And the questions that do have answers held roughly flat — the reader didn't start refusing things it should answer. Two levers stacked: the grounded-absence method, and a stronger reader that follows it more faithfully.
Where it still fails — kept in.
Even with the strong reader, of the 13 hardest false premises where the weaker reader hallucinated, 7 still hallucinate. Refusal got better; it is not solved. The metric itself is noisy — identical answers score up to ~15 points apart across judge runs, which is part of why abstention is hard even to measure. And this is a clean-room reader run over the frozen retrieved evidence: a ceiling estimate of what the method can do, not a re-scored production system and not a leaderboard headline. We say all of that plainly, because a number you have to walk back later was never worth posting.
Why this is the number that matters.
The systems climbing these benchmarks — Microsoft's Memora among the newest — push retrieval accuracy hard, and none of them abstain: there is no mechanism to say "that isn't here." That's the gap. A system you can actually trust has to know the shape of what it doesn't know — and be able to say so with the receipt in hand. Grounded absence is one honest, measured step toward it. Refusal, first-class.
If the number is wrong, the fastest way to show it is to try to break it. That's the whole point.