Autoresearch loop on Tinker reproduces self-distillation papers with per-dollar cost forecasts
simonguozirui · x · 2026-09-10
askalphaxiv shows how to build an autoresearch loop with Tinker to reproduce post-training papers at predictable costs.
- Problem: dozens of self-distillation methods claim improvements over each other; hard to audit which claims hold.
- Method: agents got a Tinker budget and, with just a few user prompts, reproduced self-distillation results automatically, with compute cost forecast accurate to within a dollar.
- Results: reproduced SDFT's continual learning benefits across Qwen3-8B and Qwen3-30B-A3B over multiple seeds, and investigated SFT's failure modes.
- tinkerapi notes the same loop works for auditing: testing dozens of competing published methods automatically, not just exploratory research.
Related event: Automated Research Loop Reproduces Self-Distillation Papers with Tinker(2 posts)→
More from coding & agent
- v0 brings Vercel integrations into chat: one-click Resend, MongoDB, Algolia and more — evilrabbit_ · 2026-09-10
- Using an AI agent to digest Apple's 6 iPhone Duo developer videos: key iOS 27.1 takeaways — JordanMorgan10 · 2026-09-10
- Five Things You Shouldn't Vibe Code: Auth, Payments, Time Zones and More — rseroter · 2026-09-10
- Complete 2026 Guide: How to Actually Build an App with AI Without It Hurting Your Product — davemccollough · 2026-09-10
- Claude Code 2.1.267 adds maxEffortLevel cap and system prompt snapshot toggle — ClaudeCodeLog · 2026-09-10
- Claude Code 2.1.267 adds maxEffortLevel cap and per-request system prompt re-rendering — ClaudeCodeLog · 2026-09-10