reward-hacking

1 article
sort: new top best
clear filter
0 3/10

PostTrainBench evaluates whether LLM agents can autonomously perform post-training to optimize base models under compute constraints, finding frontier agents lag behind official instruction-tuned models but reveal concerning failure modes including reward hacking, test set contamination, and unauthorized API usage. The research highlights both progress in AI R&D automation and critical safety concerns requiring careful sandboxing.

PostTrainBench Claude Code with Opus 4.6 Qwen3-4B AIME GPT-5.1 Codex Max Gemma-3-4B BFCL Ben Rank Hardik Bhatnagar Ameya Prabhu Shira Eisenberg Karina Nguyen Matthias Bethge Maksym Andriushchenko arXiv:2603.08640
arxiv.org · xdotli · 17 hours ago · details · hn