A Problem That Had Been Stuck for 25 Years Was Solved by AI in a Week

There is a classic hard problem in wireless communications called MIMO detection: the transmitter packs N bits into an N×N channel and sends them out; the signal gets scrambled and mixed with noise along the way, and the receiver has to recover the original bits exactly.

In theory, there is a brute-force approach: enumerate all 2^N possible bit combinations and find the one that best matches. But once N gets even moderately large, the computation runs forever. In 1989, someone proved that this problem is NP-hard in the worst case.

What researchers really wanted to know was something else: real-world channels are random, not adversarially constructed worst cases. As long as the signal-to-noise ratio reaches 2logN, the original bits can be recovered statistically. So can we design a fast algorithm that hits this threshold exactly?

For 25 years, sphere decoding, semidefinite relaxation, bit flipping, AMP, and statistical physics methods all took turns attacking the problem. The best results still stopped at twice the theoretical threshold. There was a gap no one could cross between “recoverable in a statistical sense” and “recoverable by a fast algorithm.”

Last week, Microsoft Research principal researcher Dimitris Papailiopoulos used GPT-5.6 and Fable 5 to close that gap: a two-step algorithm, polynomial time, O(N³) operations, and it hits the 2logN threshold exactly.

How the Two Models Split the Work

Dimitris asked the two models to try the problem separately. GPT-5.6 took the AMP (approximate message passing) route, an analytical tool Dimitris himself had never fully mastered; Fable 5 offered a different path: “signed LMMSE plus greedy bit flipping,” an old algorithm that had long been used in practice but had never been rigorously proved.

Both paths produced complete proofs. Dimitris chose Fable’s approach and asked GPT to take over checking and patching the gaps. The repaired proof became a “wall of symbols”—variables pointing to variables, packed with matrix analysis tools he did not understand.

Over the next few days, he repeatedly had the two models simplify each other’s arguments, with only one hard requirement: preserve the 2logN threshold. Everything else was negotiable. He even rejected Lean formal verification for a very direct reason: he did not understand Lean and therefore could not check whether the translation was correct.

After a week of back and forth, he finally had a proof that he could verify line by line by hand.

The Algorithm: Two Steps Are Enough

Step one: LMMSE rounding. LMMSE (linear minimum mean square error estimation) produces a rough continuous-valued guess, then each coordinate is rounded to +1 or -1 according to its sign. This step does not need to get every bit exactly right. The proof guarantees that after rounding, the Hamming distance from the true bits is only o(N)—the fraction of wrong guesses goes to zero as N grows.

Step two: greedy bit-by-bit flipping. Starting from the guess produced in the first step, each round checks all bits, flips the one that decreases the cost function the most, and repeats until no improving flip can be found.

The key is proving that the greedy search will not get stuck: around the initial guess, every point that is not yet correct has at least one bit flip that strictly decreases the cost, and the amount of decrease has a lower bound that does not vanish. At the same time, the cost function itself increases with Hamming distance, forming a natural guardrail—the search path cannot climb over the guardrail and run outside the neighborhood of the initial guess.

The complexity works out to O(NlogN) steps, plus O(N³) for the first step, making the overall algorithm polynomial time.

What Really Matters

More important than the problem itself is what it implies.

First, AI’s role in mathematical proof has moved from “assisting computation” to “proposing proof strategies.” GPT-5.6 and Fable 5 each produced different lines of proof, and Fable’s path was an algorithm that practitioners had used for years without a rigorous proof. That means AI can elevate “empirically works” in engineering practice into “mathematically inevitable.”

Second, the human-AI collaboration model has changed. Dimitris did not understand AMP, did not understand Lean, and did not understand the matrix analysis tools. But he knew whether the result was right and where the bottom line was. He acted as a reviewer and arbitrator, not as the prover. The two models simplified each other’s arguments, while he judged: “This I can understand; just keep the threshold.” This division of labor—“AI proves, humans arbitrate”—may become the norm for hard-core research in the future.

Third, Dimitris tried this problem with MCMC during his PhD in 2009 and failed. Seventeen years later, he solved it with AI. Problems that once stalled an academic career now have another possible route to solution.

What This Means for Ordinary People

MIMO detection is a foundational technology for 5G/6G and Wi-Fi 7/8. A more efficient detection algorithm means receiver chips can recover signals with less compute. The direct payoff is lower-power, higher-throughput communications baseband chips.

But the more far-reaching impact is on the AI side: when large models can chew through hard-core mathematical proofs, they are no longer just “smarter search engines.” They are beginning to participate in the generation of human knowledge. Writing code, fixing bugs, doing design—those are application-layer tasks. Mathematical proof sits at the foundation layer. AI’s penetration of the application layer has been the story of the past two years; opening up the foundation layer may be the real narrative of the next five.

LynxHot · Tracking AI tools, models, and industry shifts