Qu, Qian, Wang, Hongyang, Chen, Wenbo, Li, Yichuan, Liu, Jingguo, Li, Xian, Wang, Xuan, Wen, Ruochen, Fan, Leo, Yan, Yiming, Wang, Zheshen, Dumoulin, Benoit · Nov 14, 2026
CompoundingKernel search ranks first on NVIDIA B200 benchmark
Markdown · JSON · Original source
The authors present CompoundingKernel, a self-improving agent system that writes kernels by search. Agents start with each problem’s PyTorch definition, but no reference kernels or official scores. A per-problem archive outside the model stores correct kernels and their lineage. An controller selects a parent or a new design, then gives each parallel child agent one direction. The authors report that, as of November 11, 2026, generated kernels ranked #1 on the public L1, L2 and boards of SOL-ExecBench for B200. All 235 problems had a site-verified correct kernel; on the 209 public-board problems, kernels were 4.40x faster than the ’s reference baseline. At equal wall , their search beat a coded UCB style tree on 35 of 38 problems by making 2.4x as many correct attempts. The authors’ central scaling finding is that reliable measurements, not the number of agents, limit how far such a search scales. They also report that a warm restart fix raised evaluation rate 3.6x while over a thousand agents ran at once. These are benchmark results and system findings reported by the authors; the abstract does not establish performance beyond this benchmark.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 14, 2026 (2026-11-14).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-17T13:02:45.890Z. First observed: 2026-11-17T08:28:51.529Z. Retrieved: 2026-11-17T12:48:16.009Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The reported speed comparison is for the 209 public-board problems and is relative to the 's reference baseline; it does not establish performance outside this benchmark. The supplied abstract does not detail how reliable measurements were defined or how their rate was measured. The abstract reports a warm restart fix and its evaluation-rate increase but does not describe the fix.
Open question
What rate of reliable measurements can this search sustain as it scales, and how does that rate affect its results?
Source-stated
The authors present CompoundingKernel as a self-improving agent system that searches to write GPU kernels, starting from problem definitions without reference kernels or official scores.
The search starts from zero: agents get each problem's PyTorch definition but no reference kernels or official scores.
Source-stated
The authors report first-place rankings on three public SOL-ExecBench boards as of November 11, 2026, and site-verified correct kernels for all 235 problems.
As of 2026-11-11, its generated kernels rank #1 on the public L1, L2 and Quantization boards of SOL-ExecBench (NVIDIA B200).
Source-stated
For the 209 public-board problems, the authors report kernels 4.40x faster than the benchmark's reference baseline.
on the 209 public-board problems they are 4.40x faster than the benchmark's reference baseline.
Source-stated
The authors report that their search beat a coded UCB style tree on 35 of 38 problems at equal wall time, making 2.4x as many correct attempts.
At equal wall time, it beat a coded UCB style tree on 35 of 38 problems by making 2.4x as many correct attempts.
Source-stated
The authors identify the rate of reliable measurements, rather than agent count, as the limit on scaling this search.
We find that the rate of reliable measurements, not the number of agents, limits how far such a search scales.
Source-stated
The authors report that a warm restart fix raised evaluation rate 3.6x while more than a thousand agents ran at once.
Over a thousand agents ran at once, and one warm restart fix raised their evaluation rate 3.6x.
Source and contributor context
Qu, Qian, Wang, Hongyang, Chen, Wenbo, Li, Yichuan, Liu, Jingguo, Li, Xian, Wang, Xuan, Wen, Ruochen, Fan, Leo, Yan, Yiming, Wang, Zheshen, Dumoulin, Benoit · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Graphics processing unit
A chip that performs many calculations at once; the main hardware used to train and run AI models.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Quantization
Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
NVIDIA
NVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: ea238138-5d64-4fb1-8bd5-70e1e529e061.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-17T13:02:45.890Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 50edb15afc0cbdb6a7ff18ff82c42995cc89c26bd08995256bc5a3d58f71ce3b. Source text hash: b8f1ec74d2f72bd2783a6c23c1ae3d5b973eb9f548064198d3e70cc1cecd37a5.
Revision and source-check history
Version 1 · 2026-11-17T13:02:45.890Z
CompoundingKernel search ranks first on NVIDIA B200 benchmark
The authors present CompoundingKernel, a self-improving agent system that writes kernels by search. Agents start with each problem’s PyTorch definition, but no reference kernels or official scores. A per-problem archive outside the model stores correct kernels and their lineage. An controller selects a parent or a new design, then gives each parallel child agent one direction. The authors report that, as of November 11, 2026, generated kernels ranked #1 on the public L1, L2 and boards of SOL-ExecBench for B200. All 235 problems had a site-verified correct kernel; on the 209 public-board problems, kernels were 4.40x faster than the ’s reference baseline. At equal wall , their search beat a coded UCB style tree on 35 of 38 problems by making 2.4x as many correct attempts. The authors’ central scaling finding is that reliable measurements, not the number of agents, limit how far such a search scales. They also report that a warm restart fix raised evaluation rate 3.6x while over a thousand agents ran at once. These are benchmark results and system findings reported by the authors; the abstract does not establish performance beyond this benchmark.
Original source
cs.AI · Article
Source details
Original source date: Nov 14, 2026 (2026-11-14).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-17T13:02:45.890Z. First observed: 2026-11-17T08:28:51.529Z. Retrieved: 2026-11-17T12:48:16.009Z.
Limits
The reported speed comparison is for the 209 public-board problems and is relative to the 's reference baseline; it does not establish performance outside this benchmark. The supplied abstract does not detail how reliable measurements were defined or how their rate was measured. The abstract reports a warm restart fix and its evaluation-rate increase but does not describe the fix.
Open question
What rate of reliable measurements can this search sustain as it scales, and how does that rate affect its results?
Source-stated
The authors present CompoundingKernel as a self-improving agent system that searches to write GPU kernels, starting from problem definitions without reference kernels or official scores.
The search starts from zero: agents get each problem's PyTorch definition but no reference kernels or official scores.
Source-stated
The authors report first-place rankings on three public SOL-ExecBench boards as of November 11, 2026, and site-verified correct kernels for all 235 problems.
As of 2026-11-11, its generated kernels rank #1 on the public L1, L2 and Quantization boards of SOL-ExecBench (NVIDIA B200).
Source-stated
For the 209 public-board problems, the authors report kernels 4.40x faster than the benchmark's reference baseline.
on the 209 public-board problems they are 4.40x faster than the benchmark's reference baseline.
Source-stated
The authors report that their search beat a coded UCB style tree on 35 of 38 problems at equal wall time, making 2.4x as many correct attempts.
At equal wall time, it beat a coded UCB style tree on 35 of 38 problems by making 2.4x as many correct attempts.
Source-stated
The authors identify the rate of reliable measurements, rather than agent count, as the limit on scaling this search.
We find that the rate of reliable measurements, not the number of agents, limits how far such a search scales.
Source-stated
The authors report that a warm restart fix raised evaluation rate 3.6x while more than a thousand agents ran at once.
Over a thousand agents ran at once, and one warm restart fix raised their evaluation rate 3.6x.
Source and contributor context
Qu, Qian, Wang, Hongyang, Chen, Wenbo, Li, Yichuan, Liu, Jingguo, Li, Xian, Wang, Xuan, Wen, Ruochen, Fan, Leo, Yan, Yiming, Wang, Zheshen, Dumoulin, Benoit · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Graphics processing unit
A chip that performs many calculations at once; the main hardware used to train and run AI models.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Quantization
Storing a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
NVIDIA
NVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: ea238138-5d64-4fb1-8bd5-70e1e529e061.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-17T13:02:45.890Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 50edb15afc0cbdb6a7ff18ff82c42995cc89c26bd08995256bc5a3d58f71ce3b. Source text hash: b8f1ec74d2f72bd2783a6c23c1ae3d5b973eb9f548064198d3e70cc1cecd37a5.
uncertain · 2026-11-17T13:02:45.890Z · anthropic/claude-opus-5.5
Publication: ea238138-5d64-4fb1-8bd5-70e1e529e061
attribution — pass: The draft consistently attributes the system, rankings, speedups, comparison and scaling finding to the paper's authors as reported claims, not established facts.
We present CompoundingKernel, a self-improving agent system that writes GPU kernels by search.
summary — uncertain: The summary is faithful but hard for a non-specialist to follow. Central terms such as GPU kernels, PyTorch definition, LLM controller, UCB style tree, wall time and warm restart are not explained, and there are no tooltips. It also never says why faster kernels matter. Smallest repair: add term-library tooltips for these terms and one sourced sentence on significance. Two smaller fidelity fixes: restore 'rate of' in 'reliable measurements ... limit', and consider noting that checks outside the model decide correctness.
Checks outside the model decide correctness.
date — pass: originalPublished is 2026-11-14, which matches the v1 submission date. The feed date of Nov 17 is later and is correctly not used. The Nov 11 leaderboard date in the draft matches the abstract.
[Submitted on 14 Nov 2026]
support — pass: Each claim matches the abstract's wording and numbers: #1 on three boards, 235 verified problems, 4.40x on 209 problems, 35 of 38 at 2.4x correct attempts, more than a thousand agents, and a 3.6x evaluation-rate fix. The caveats and the question stay within the evidence and invent no benefits or costs.
At equal wall time, it beat a coded UCB style tree on 35 of 38 problems by making 2.4x as many correct attempts.
media — pass: The selected media is the verified arXiv abstract page, which is the direct original source and the easiest way to access the paper.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A chip that performs many calculations at once; the main hardware used to train and run AI models.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceStoring a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Definition referenceNVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
Explore NVIDIAA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA chip that performs many calculations at once; the main hardware used to train and run AI models.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceStoring a model's numbers at lower precision so it uses less memory and runs faster, sometimes at a cost in accuracy.
Definition referenceNVIDIA develops processors, networking and software used for AI, graphics and scientific computing.
Explore NVIDIAA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference