On August 9, 2026, Beijing time, Dimitris Papailiopoulos posted a long article on the X platform and uploaded a 46-page draft of his paper. Public submission records show that the paper was uploaded at 18:23 Coordinated Universal Time (UTC) on August 8, with the long-form post published approximately seven minutes later. The research claims that GPT-5.6 and Claude Fable 5 seem to have solved a theoretical problem in square Gaussian binary MIMO detection that has persisted since 2001—whether there exists a polynomial-time algorithm reaching the same first-order signal-to-noise ratio threshold when exact recovery is statistically feasible.
Sole Author and Scope of the Problem
Papailiopoulos is a researcher at Microsoft Research and an associate professor of electrical and computer engineering at the University of Wisconsin-Madison, currently on leave from the university to work at Microsoft Research. He is the sole author credited in the paper titled 'Polynomial-Time MIMO Detection at the Maximum-Likelihood Threshold.' A footnote on the first page of the paper discloses that GPT-5.6 and Claude Fable 5 were involved in proving and drafting the initial manuscript; Papailiopoulos formulated the problem, guided proof simplification, verified mathematical arguments, edited the entire text, and assumed responsibility for its content.
The research examines an N by N square real Gaussian binary MIMO model: at the transmitter end, an N-bit vector composed of positive or negative ones is sent. The signal is mixed through an N by N channel matrix and independent Gaussian noise is added. At the receiver end, with knowledge of the channel matrix, the goal is to recover the complete vector. Maximum likelihood detection finds the optimal solution by comparing all candidate vectors for the least squares objective. The paper notes that in its general form, optimal multi-user detection falls under NP-hard problems, and whether there exists an efficient algorithm reaching statistical thresholds on random Gaussian channels is the average-case problem addressed in this draft.
From the Statistical Threshold to a Two-Stage Detector
The current paper's review of existing literature states that the first-order signal-to-noise ratio (SNR) threshold for maximum likelihood complete recovery was previously determined to be 2 log N; spherical decoding cannot provide the required polynomial-time guarantees in the relevant asymptotic dimensions. The paper attributes the closest prior theorem on polynomial-time complete recovery for lattice systems to box relaxation, which has a first-order threshold of 4 log N under the same normalization. This new manuscript closes the computational gap between these two first-order thresholds without introducing any new wireless devices, channel standards, or mathematical objects.
The new detector operates in two stages. The first stage computes a linear minimum mean square error (MMSE) estimate and takes the sign of each coordinate; the second stage repeatedly selects the single bit flip that most decreases the least squares objective until the improvement falls below a threshold or reaches a deterministic upper limit on flips. The proof initially places the rounded estimates within a Hamming ball around the true vector with sublinear distance, then establishes a uniform geometric bound over all error sets that local search might traverse, arguing for the existence of a descending flip from any non-true point and ensuring that the descent path does not leave a controlled region.
Precise Real Complexity and Reverse Boundary
Theorem 2.1 states that when the signal-to-noise ratio ρ is not less than 2 log N, this two-stage detector will have a complete recovery failure probability that tends to zero as N increases for any given transmission vector. The paper provides an upper bound on the number of flips at O(N log N), with the cost in the descent phase being O(N^2 log N). Dense LMMSE solving leads to total costs dominated by O(N^3). These complexities are only valid under a model of unit-cost precise real arithmetic operations; the paper explicitly does not advocate for Turing bit complexity or guarantees for finite precision implementations.
The draft also provides a reverse boundary: when ρ is no greater than 2 log N minus log log N, and further subtracting an sN term that tends to infinity but remains below the order of log N, a single-bit adjacent vector will approach probability one in achieving a lower objective value, even leading to maximum likelihood detection failure. There remains a low-order gap of log log N plus sN between these two boundaries. Therefore, the paper only claims that the first-order constant for both statistical and polynomial-time thresholds is 2, without asserting a complete transition window.
Model Collaboration Process and Validation Limitations
Papailiopoulos documented the collaboration process, noting that GPT-5.6 initially provided an approximate message-passing scheme, which was then replaced with a more verifiable algorithm as per his request; Claude Fable 5 proposed a signed LMMSE combined with greedy bit-flipping scheme that he preferred. He stated that the initial machine proof took about 30 minutes to generate, but simplifying the argument and manual verification took over five days; subsequent posts described continuous efforts spanning seven days. These figures correspond to different stages of the process and cannot be aggregated into a claim that AI independently completed and verified the paper within seven days.
As of this round of reading, the paper remains a draft in the public repository without an arXiv number, journal or conference acceptance information, peer review reports, errata records, or independent verification papers. The conclusions are limited to asymptotic models of square real Gaussian binary MIMO and cannot be generalized to all MIMO systems, arbitrary channels, finite-dimensional devices, or commercial 5G receivers with the same guarantees. Papailiopoulos also noted that the relevant wireless research community has largely moved on to other problems; the problem's history spanning 25 years does not directly translate into current industrial impact.