Tencent's ASAP evaluates agentic capabilities in pretrained base models before post-training
tencent · hf · 2026-10-21
ASAP (Agent Scaffold Adaptation for Pretraining Evaluation) from Tencent Youtu Lab and Shanghai Jiao Tong University is a lightweight framework for eliciting and evaluating agentic behavior in pretrained base models. Rather than converting base models into chat-style agents, it preserves the completion interface and adapts existing agent scaffolds via structured trajectory continuation, short reference prefixes, and minimal discovery mechanisms. Across software-engineering and deep-research benchmarks, ASAP yields discriminative end-to-end performance across base models and pretraining checkpoints, and exposes execution trajectories for analyzing exploration, editing, testing, recovery, and task completion across training stages.
More from coding & agent
- 6 simple Claude setups: teach the AI how you work instead of memorizing prompts — evielync · 2026-10-21
- pstack aims to be the shadcn of agent skills — zbeyens · 2026-10-21
- Claude Max now includes $100/month in API credits, screenshot hints at Opus 5.5 with 1M context — michalmalewicz · 2026-10-21
- Rejected 250 times in 2 years, this engineer landed a $500K offer after demoing 20 running agents — msharmas · 2026-10-21
- AI agent scores Steve Jobs' Stanford speech in GarageBand fully via computer use — itsOmSarraf_ · 2026-10-21
- Dev builds a four-agent pipeline for sponsored posts that adapts mid-campaign without restarting — alifcoder · 2026-10-21