Vulnerability detection
We model programs from both what they do at run time and what the code says. That found more than 100 vulnerabilities in 20+ open-source projects and 34 new CVEs.
TECHNOLOGY
Models read code as text; they do not know how it runs on real hardware. We run the same code across architectures, devices and inputs and learn its performance, resource use and execution paths: a computational world model.
COMPUTATIONAL WORLD MODEL
The same code, across scenarios and inputs.
Real runs on CPUs, GPUs, domestic Chinese chips and other accelerators, with their behavior recorded.
Model the code’s performance, memory and resources, and execution paths on each kind of hardware.
Meta’s Code World Model ↗ simulates what code computes (variables and state) in one Python environment. Ours models how the same code runs across chips and devices: speed, memory and execution paths.
RESEARCH
We model programs from both what they do at run time and what the code says. That found more than 100 vulnerabilities in 20+ open-source projects and 34 new CVEs.
Machine learning plus statistical evaluation to optimize code on different hardware. It is in use in more than 8 settings, with an average speedup of 180%.
Program analysis keeps the agent in check, so its conclusions can be verified. On 72 binary-analysis tasks, BinReX succeeds nearly twice as often as OpenAI Codex with IDA. EXRAY finds 77.9% of known vulnerabilities, and found 24 previously unknown RCE vulnerabilities in 130 applications.
BENCHMARKS
1.9×success rate, against OpenAI Codex with IDA
Task success rate
Total analysis time (lower is better)
6.5×larger model on one GPU
Largest model on one 32 GB V100
4.5×training throughput (best case)
Against offloading-only methods (best case)
>11×faster Halide image pipelines (some benchmarks)
Halide image pipelines (some benchmarks)
TVM neural-network code
LLVM code-size reduction
−63.5%energy
Energy (lower is better)
Load time (lower is better)
About −50%false-positive rate
False-positive rate (lower is better)
WHERE IT IS USED
Yi Guard checks an agent against your policy before it calls a tool, runs code or touches a file, and stops dangerous actions.
Explore Yi GuardWe deploy AI products to the compute centre you choose, move them onto its chips and tune them; Yi Code refactors and rewrites code in a sandbox.
Explore Yi CodeYi Eval tests an agent before launch and after every change, and every finding comes with evidence you can check.
Explore Yi EvalBLOG
Put the same code on a different chip and its speed, memory use and execution path all change. We want models to learn that.
On 72 binary-analysis tasks BinReX succeeded 83.3% of the time. Codex with IDA, on the same GPT-5, managed 43.1%. The difference comes down to looking up similar functions first, and backing every step with evidence.
Models for compiler optimization and bug detection work at launch, then quietly get worse. PROM uses a statistical test to flag the predictions it can’t trust, and relabeling under 5% of samples brings performance back.
PAPERS
Next step