GLM-5.3 Cyber Scores Surge: Same Base Model, Targeted Post-Training
ChrisGPT · x · 2026-08-14
The author points out that GLM-5.3 uses the exact same base model as GLM-5.2, but targeted cyber post-training led to massive leaps in its cybersecurity benchmark scores.
Specifically, CyberGym scores jumped from 77.2 to 84.5, ExploitBench from 24.4 to 54.4, and completed ExploitGym tasks surged from 29/39 to 105/130 at the 2-hour and 6-hour marks. The author argues that training for a specific domain isn't cheating, unlike training on exact test cases. Additionally, web access was blocked and Git metadata removed during evaluation, though no contamination audit was disclosed.
More from Models
- Google says teams across the company are using Gemini 4 Argon on its hardest engineering problems — CurieuxExplorer · 2026-10-15
- Perplexity's new embedding models: index with a 9B model, search with 0.6B — antoine_chaffin · 2026-10-15
- 797GB of n-gram counts in CPU RAM let 1B models match 57% more compute — PangWeiKoh · 2026-10-15
- Leak: Moonshot's Kimi flagship 'x4' done with pre-training, K3.1 in internal testing — lyraxana · 2026-10-15
- Unsloth ships GGUF quants for Nemotron 3.5: Bolt-74B runs in 24GB RAM — danielhanchen · 2026-10-15
- NVIDIA ships two open Nemotron 3.5 Bolt models: a 74B-A9B MoE and a mobile-friendly E4B — HankYeomans · 2026-10-15