Autoresearch Loop with Tinker Reproduces Self-Distillation Papers at Predictable Cost
SRSchmidgall · x · 2026-09-10
askalphaxiv demonstrates an autoresearch loop built on Tinker for reproducing post-training papers. With dozens of self-distillation methods claiming mutual improvements, they gave agents a Tinker budget to verify which claims hold up: from a few user prompts, agents reproduced SDFT's continual-learning benefits on Qwen3-8B and Qwen3-30B-A3B across multiple seeds, and investigated SFT's failure modes — all at predictable cost.
More from coding & agent
- Open-Source Ref2V Workflow Auto-Transcribes Media and Formats Prompts for H3 Video Generation — bstr3k · 2026-09-10
- K3 hit 77% speedup optimizing mjwarp kernels; GPT-6 could only add 0.38% — YouJiacheng · 2026-09-10
- K3 got 77% speedup on mjwarp kernels; GPT-6 Astra added just 0.38% — YouJiacheng · 2026-09-10
- Same async program, four different outputs: paper maps the async/await design space across 7 runtimes — IanArawjo · 2026-09-10
- Dev says AI-written PRs keep getting smarter, and reviewing them is getting harder — smlpth · 2026-09-10
- Automattic launches here.now, instant web hosting for agents with 500k sites published — iannuttall · 2026-09-10