Paper of the Day · November 15, 2026

Look Here, Then Decide: Small Decision Models in a Harness Find Memory-Safety Vulnerabilities at Scale on Commodity GPUs

A small 4B-parameter open model is weak at judging code alone, but it ranks candidate functions well enough that a harness on cheap GPUs confirmed 187 memory-safety findings.

What they found

The authors built a harness that picks widely used C and C++ projects. It uses fixed, rule-based tools to decide where to look. Then a small open model, Intern-Decision-4B, scores each candidate function. The model is weak alone: AUC 0.73 for vulnerable versus benign functions, and 0.56 for a vulnerable function versus its own patch. As a ranker it helped more. On 75 public CVEs, the fixed function reached the repository's top 25 in 21% of cases, 1.4 times as often as random order. The campaign searched more than 4,000 projects and confirmed 187 memory-safety findings: 12 critical, 148 high, 23 medium and 4 low by worst-case CVSS v3.1.

Why it matters

The work suggests that finding software bugs at scale does not need a frontier model. A small open model on gaming or desktop hardware can do the prioritizing, while fixed analysis and verification do the rest. The authors say recall is limited mainly by where the localizers look, not by the model. The 7.5 times gain on the campaign's own queues is, in the authors' words, a selection-biased upper bound, so it should be read with care.

How it changes the race

If the result holds, it lowers the cost and hardware needed to run large vulnerability searches. More groups, not only those with frontier models, could do this work. The gains come from the system design around a modest model.

Safety and control

The same cheap, open setup could help attackers as well as defenders. The authors say all confirmed findings are being disclosed to maintainers under coordinated disclosure before publication. They also stress that deterministic analysis and verification stay essential.

The abstract, by the authors

We ask whether small, open-weight decision models can support vulnerability discovery in open-source software at scale and what the surrounding system must provide. These models answer a typed question in one forward pass, returning scores for the permitted answers rather than text. Our harness selects widely packaged C and C++ projects and uses deterministic incomplete-fix miners and shape localizers to decide where to look. Intern-Decision-4B, a 4B-parameter open-weight decision model, then scores every candidate function. The scores rank the queues rather than accept or reject candidates, apart from a low reach-score floor that some queues kept. The search uses no frontier model. Alone, the model is weak: it separates vulnerable from benign functions with AUC 0.73, but a vulnerable function from its own patch with only 0.56. It is more useful for ordering candidates. On 75 public CVEs from projects without any of our findings, the fixed function reached its repository's top 25 in 21% of cases, 1.4 times as often as under random order of the same queues. Recall is bounded mainly by where the localizers look. On the campaign's own queues the model's order placed 7.5 times more confirmed findings in the top 10 than random order, an observed, selection-biased upper bound. On three GB10 systems, the harness searched more than 4,000 projects at 900-1,100 repositories per hour. The campaign confirmed 187 memory-safety findings; by worst-case CVSS v3.1, 12 are critical, 148 high, 23 medium and 4 low. The decision step runs at 4.4 decisions per second on one GB10 and 9.8 on a consumer RTX 4090 gaming PC, with matching scores. A small decision model supports vulnerability discovery through prioritization, while deterministic analysis and verification remain essential. All confirmed findings are being disclosed to maintainers under coordinated disclosure before publication.

Quoted from arXiv:2611.16565. The research and its words belong to its authors.

The authors

Tim De Wachter, Kyryl Volkov, Myrto Eleftheria Gkogkou, Mathieu Huysman

Credit for this research belongs to its authors. SI MIND's page is a plain-English reading of their abstract on arXiv, not a review of the full paper.