A 25-year gap, closed in seven days
Microsoft Research principal researcher Dimitris Papailiopoulos, working with GPT-5.6 and Claude Fable 5, has produced a proof that a simple two-step algorithm recovers every transmitted bit in MIMO detection at the theoretical maximum-likelihood threshold — in polynomial time. The result settles a question the field had chased since 2001. It took seven days.
Why this problem mattered
MIMO detection is the receiver-side puzzle at the heart of wireless communication. A transmitter sends N bits through an N×N channel that scrambles them and adds noise; the receiver holds only the corrupted signal and must reconstruct all N bits exactly. The guaranteed method — maximum-likelihood detection — works by exhaustive search over all 2^N candidates. Correct, but exponentially expensive. In 1989, Sergio Verdú proved the worst case is NP-hard.
But real wireless channels are random, not adversarial. By the early 2000s, researchers had a cleaner target: when signal-to-noise ratio reaches 2logN, exact recovery becomes statistically possible; below that line, even ML detection fails. That line is the maximum-likelihood threshold. The open question: can a fast algorithm hit it exactly?
Two decades of near misses
In 2001, Hassibi and Vikalo argued sphere decoding ran in expected polynomial time. In 2005, Jaldén and Ottersten showed it was exponential at any fixed SNR — the sphere must grow with problem size to contain the signal, and so does the search. Then came semidefinite relaxation, bit-flipping local search, approximate message passing (AMP), and statistical physics methods: each with elegant analysis, none provably matching 2logN. The best rigorous result, 2020's box relaxation, achieved exact recovery at 4logN — twice the threshold. For 25 years, the gap between "statistically recoverable" and "provably recoverable by a fast algorithm" held.
What the proof actually is
The winning algorithm has only two steps. First, LMMSE rounding: compute a continuous linear estimate (linear minimum mean-square error), then round each coordinate to +1/−1. The proof shows the rounding is wrong on only o(N) bits — a vanishing fraction. Second, greedy bit-flipping: repeatedly flip whichever bit most reduces a cost function measuring mismatch with the received signal.
Two lemmas make it work. At every incorrect point inside the basin, at least one flip strictly decreases cost with a non-vanishing margin — so the search never stalls. And the cost function grows with Hamming distance from the truth, forming a guardrail the search cannot escape. Together they imply greedy search stops only at the true bitstring, in O(NlogN) steps; total complexity is O(N³). The result is two-sided: exact recovery at 2logN, and just below it even ML begins to fail — the threshold is not just matched, it's confirmed as the wall.
The collaboration that cracked it
Papailiopoulos gave the problem to GPT-5.6 and Fable 5, and the two models returned different proof strategies. GPT-5.6's route ran through AMP — a class of tools Papailiopoulos had never fully mastered. Fable 5's went through symbolic LMMSE plus greedy bit-flipping: an algorithm industry already runs in practice, but which had never been rigorously proven. He chose Fable's path, had GPT-5.6 audit and patch the holes, then spent days having the two models simplify each other's arguments. His only constraint: whatever the simplification, the 2logN threshold had to survive. He declined Lean formal verification — he doesn't know Lean, and a translation he can't check is just a new source of error. What he finally accepted was a proof he could verify line by line by hand.
What this signals
Three things. First, the value of AI in research isn't only about novel invention — it's about closing proof gaps around algorithms we already trust in production. Industry has been running LMMSE-plus-greedy-flipping for years; now it has a certificate of optimality. Second, "AI did math" undersells what happened: the human set the constraint (keep 2logN), arbitrated between candidate proofs, and demanded readability. The bottleneck wasn't generating proofs — it was getting one a human could check. Third, this closes a 17-year personal arc: in 2009, as a first-year PhD student, Papailiopoulos tried MCMC on this exact problem and failed. The same person, with two models as collaborators, finished it in a week.
What to do with this
For wireless and signal-processing teams, this is directly relevant to MIMO receiver design — worth reading the proof and benchmarking your current detector against the two-step algorithm. For researchers generally, the workflow is the template: generate multiple independent AI proofs, cross-audit them, then force simplification down to human-checkable size before trusting anything. And the refusal of Lean is a lesson in itself — formal verification is only useful if you can read the translation. Readability is a research output, not a luxury.