The public work each one rests on is linked; the correspondence itself is available on request.
Each link opens the pull request or issue itself, with its review history.
MergedTencent ncnn #7239 (Tencent’s on-device AI inference framework): fix for NaN outputs of the Mish activation on Vulkan GPUs, where ex overflowed for large inputs; its creator, nihui, added a commit extending the tests, approved and merged it (5 November 2026, India time): “Thanks for your contribution !”
MergedNVIDIA Ising-Decoding #98, #99 (NVIDIA's AI decoders for quantum error correction): the models NVIDIA publishes on Hugging Face now load out of the box in its tests, and both checkpoint settings work as documented. Reviewed over several rounds, approved and merged by maintainer Ivan Basov (30 October and 2 November 2026): “Merged, thank you for the contribution and for working through the review rounds.”
MergedApple MLX #4737: data-race fix in MLX's distributed ring backend, the transport that links Macs for multi-machine training; reviewed, approved and merged by MLX maintainer @zcbenz (22 October 2026). ThreadSanitizer: 2–4 races per run before the fix, none after.
MergedMeta ExecuTorch #23677: bounds check for the 2-bit and 4-bit quantized embedding kernels used by on-device LLMs, approved by two Meta maintainers (16 October 2026); Scott Roy (Meta): “LGTM! Thanks for the fix!”
MergedMeta ExecuTorch #23912: the same kernels checked zero points against the output type instead of the scales' type, so valid fp16-scale models aborted and the .out variants read past their scales (AddressSanitizer); revised after Meta maintainer review, then approved and merged by Meta's Jacob Stevens (28 October 2026).
MergedHugging Face OpenEnv #1441, #1442: two fixes to the reward rubrics that grade RL agents (19 October 2026). Reviewer’s approval: “Reviewed hook ordering and async composition; 177 rubric tests pass.”
MergedUniversity of Oxford PyZX #563: quantum-circuit optimiser regression fix, merged by PyZX co-creator John van de Wetering (22 October 2026): “Thanks for the fix! This was a regression from an earlier PR that changed how XPhase gates were handled in optimization.”
MergedJaxMARL #206 (NeurIPS 2024; BOLD Lab of Oxford, UCL and Imperial): fix to the rollout manager behind 8 Q-learning baselines, which zeroed the first observation of every new episode; approved and merged by lead author Alex Rutherford (26 October 2026): “Good find!!”
MergedCarnegie Mellon University XGrammar #1012 (the default structured-output engine of vLLM, SGLang and NVIDIA TensorRT-LLM): fix for jump-forward decoding crashing with UnicodeDecodeError whenever a grammar forces part of a UTF-8 character, the bug behind issue #132; approved and merged by maintainer Linzhang Li (29 October 2026).
In reviewGoogle JAX, vLLM, SGLang, PyTorch, MIT Han Lab (nunchaku) and ETH Zurich (DaCe). Google assigned JAX #41859 for review to Matthew Johnson, one of JAX’s original authors (21 October 2026); DaCe's maintainer cleared #2729 to merge once its CI is green (28 October 2026).
MergedIBM Quantum Relay-BP #49 (quantum error-correction decoder, © IBM): the decoder gave different answers for the same shot depending on batch order (serial and parallel batches disagreed on 1,970 of 2,000 shots; now 0). Its first author, Tristan Müller (IBM Quantum), chose this fix on issue #47 (“thank you for opening this! I think 1. is fine!”) and merged it on 5 November 2026: “Thank you! Happy with this, will merge it.”
OpenRiverlane Deltakit: proposed offline Relay-BP decoding for Deltakit's qLDPC circuits.
Grouped by sector: five fixes merged at three institutions (the US Department of Energy national laboratories, the Massachusetts Institute of Technology and the Barcelona Supercomputing Center), the rest in review. Each was tested before and after the fix with the project’s own tests, in public runs.
MergedEnergy and national security: US Department of Energy national laboratories. Kokkos #9738, the GPU performance-portability library of Sandia and Oak Ridge National Laboratories (Exascale Computing Project): sort compiles on NVIDIA CUDA and AMD HIP for types without a default constructor; 20 of 20 GPU sort tests pass. Merged by Damien Lebrun-Grandié, who leads the Kokkos project at Los Alamos National Laboratory (5 November 2026), after approvals from him and Daniel Arndt, a Kokkos developer and computational scientist at Oak Ridge National Laboratory; listed under Bug Fixes in the changelog for the upcoming Kokkos 5.4 release; reviewed by Andrey Prokopenko of Oak Ridge National Laboratory, who proposed a follow-up fix.
In reviewWeather and climate: European Centre for Medium-Range Weather Forecasts. ecCodes #684, its library for GRIB weather and climate data (intergovernmental, 35 states, headquartered in Reading, UK): reduced Gaussian grids scanned south to north put 6,067 and 6,026 points at latitude 0 and were not recognised as global; now none, and all 112 tests pass.
In reviewWeather and climate: UK Met Office and Science and Technology Facilities Council. PSyclone #3668, the code generator of the Met Office’s next-generation LFRic weather and climate model, built by the Science and Technology Facilities Council’s Hartree Centre with the Met Office: Fortran BIND(C) is no longer dropped and BIND(C) functions no longer crash it; 6,502 tests pass. Reviewed by a PSyclone maintainer: extra tests and a restructure requested and made the same day; “otherwise it looks good”.
In reviewNational supercomputing: Swiss National Supercomputing Centre, ETH Zurich. ReFrame #3750, the regression-testing framework supercomputing centres use to validate their systems: a regression test for issue #3669, on the ReFrame 4.10 milestone; 1,451 unit tests pass.
In reviewParticle physics: European Organization for Nuclear Research (CERN). Rucio #8862, the data management system of the Large Hadron Collider experiments (ATLAS, CMS) and Fermilab’s DUNE: eleven commands, upload and download among them, exited with code 0 even when they failed; they now return the real error code, with a test that needs no server.
MergedAcademic research: Massachusetts Institute of Technology. MEEP #3364, MIT’s electromagnetics simulator: progress time estimates that were up to 4× too short, now 0.97–1.02× the real time (issue #2694, where Prof. Steven G. Johnson asked for a pull request); merged by Ardavan Oskooi, MEEP’s original lead developer (5 November 2026, India time). #3363, a clear error when the real or imaginary kz option is combined with complex fields and nonzero kz (issue #3170, filed by Prof. Johnson): merged by Oskooi on 6 November 2026: “Option 1 of #3170 (raising a ValueError) is the correct approach because in Python, failing fast and explicitly is strongly preferred over silently mutating inputs.”
MergedNational supercomputing: Barcelona Supercomputing Center, Spain’s national supercomputing centre. Autosubmit #3405, its workflow manager for climate and weather experiments: on PostgreSQL, job timestamps such as 20261027164231 overflowed 32-bit columns and runs crashed with NumericValueOutOfRange; they are now 64-bit, checked on PostgreSQL 16. Its maintainer Bruno P. Kinoshita invited the fix (“I should be able to review this quickly as we need it for the 4.2.0 release”), approved it (“Looks great! … Thank you!!”) and merged it with a thank-you in the Autosubmit 4.2.0 changelog. #3406, at his request: a test that failed on busy CI runners now waits for its process to exit; merged: “Thanks (again) @Shakhtar-Sankur !!!” (6 November 2026).
In reviewOceans: Commonwealth Scientific and Industrial Research Organisation (CSIRO), Australia’s national science agency. emsarray #233, its library for ocean and coastal model data: forecast datasets picked forecast_reference_time instead of time as their time coordinate; candidates are now ranked as its maintainer suggested; the full suite passes with the project’s locked dependencies and image comparison.
Every repository has the scripts that produced these numbers; the raw output files are linked where a single file holds them.
IBM: GPU decoder bit-identical to IBM's Relay-BP on every shot checked (512 across two T4 runs, float64 and float32); a single shot 28.7× faster than IBM's decoder (median 37.5 ms against 1.08 s); a batch 4.0× faster than IBM's decoder on every CPU thread of the same machine; IBM's published result reproduced first (41 vs 44 errors in 10,000 shots).
NVIDIA and PyMatching: NVIDIA's two Ising pre-decoders run through NVIDIA's own pipeline, logged error rates matched to every printed digit (524,288 shots per distance); PyMatching 2's published scaling and thresholds reproduced.
NASA: Monte Carlo collision probability agreeing with NASA CARA's on all 53 of its test conjunctions (χ² p = 0.67), 15–19× faster per core than NASA's library.
NVIDIA cuBLAS: tessel's matrix multiply at parity with cuBLAS on a T4 (1.00–1.04×); causal attention 1.3–1.9× faster than PyTorch SDPA; an LLM engine on its kernels matching vLLM token for token.
NVIDIA GPU and x86 CPU, bit for bit: samebit runs Qwen2.5-0.5B (494M parameters) entirely in integers, within 0.75% of float32 perplexity on 10,240 tokens of WikiText-2; its logits are bit-identical on an NVIDIA Tesla T4 and an x86 CPU with every dot product split across 1 to 64 GPU blocks, while PyTorch float16 on the same GPU shifts logits by up to 0.25.
SubmittedHarvard's Machine Learning Systems textbook (MLSysBook, Prof. Vijay Janapa Reddi): built and submitted a full lab for its Co-Labs, "The Same-Bits Contract", on why temperature 0 does not give the same answer twice (pull request #2175, 4 November 2026). Students measure 4 different float32 sums of the same 4,096 numbers, then up to 7 of 64 users getting different text only because they were batched, and fix it twice: with batch-invariant kernels and with integer arithmetic. Built to the book's lab protocol, with its math as a new module of the book's simulator and 9 new tests; it passes the book's own test suites (2,658 simulator tests and every lab CI step, 0 failures) and a full student run in the browser. A collaborator on the textbook, on the proposal: “the feature looks great”.
Vesuvius Challenge ink model, on an unread scroll: real PHerc. 139 ink, moved into the CT volume of PHerc. 800 (eligible for the First Letters prizes), is found by the official 9 µm model in 79% of windows at a 5% false-positive rate (AUC 0.94; 168 ink against 168 bare-papyrus controls; no untouched window flagged). Read in reversed depth order, it finds 0 of 48 real-text windows.
LIGO, the O4a gravitational-wave catalog: chirp, an independent GPU search with its own 3.5PN TaylorF2 waveform (phase within 6 × 10⁻¹⁰ rad of LIGO's LALSimulation), re-finds all 71 two-detector GWTC-4.0 O4a events with complete data at a median 95% of their catalog network SNR, 67 within 50 ms, searching 3,886 templates on one T4 in 17 minutes; two runs gave identical results. Searched blind over 188 s per event, with gating and the χ² veto, the loudest trigger is the real event for 62 of 71 (48 without the veto).
OpenAI and Google tokenizers: lipi-Indic needs fewer tokens than GPT-4o's and Gemini's in 23 of 24 Indian languages.
Hugging Face and vLLM: relay prefill 8.7× faster on a T4, 384 of 384 requests identical through 60 worker kills; switchyard decodes 2.2× faster than Hugging Face, bit-identical to one GPU.
PyTorch and NCCL: tandem trains bit-identically to PyTorch DDP and FSDP, all-reduce within 2% of NCCL's bandwidth; ratchet lifts Qwen2.5-0.5B from 48.1% to 52.5% on GSM8K.
torch.compile, ONNX Runtime and XLA: kiln runs BERT in 117 ms on four CPU cores, against 149, 153 and 189 ms.
Correctness: quorumdb, verified by simulating 10,000 clusters through 610,000 crashes and a TLA+ model, found 6 real bugs before release.
Multilingual safety: Babelguard, 33,869 scam messages in 49 languages; an English keyword filter catches 86% of English scams but 10–36% in six Asian languages.