Adaption research: agent-generated quality checklists clean training data and lift Llama-3.3-70B winrate
sarahookr · x · 2026-10-12
New research from Adaption tackles training-data pollution with adaptive agentic checklist generation: an agent builds a quality checklist from past AutoScientist runs, then scores and removes bad training items. Low-quality data is known to degrade LLM performance at inference and can even lead to malicious behavior from misconfigured sandboxes.
The team trained Llama-3.3-70B on randomly sampled vs. checklist-cleaned data and evaluated winrate on held-out answer quality across domains — the model trained on cleaned data consistently won. Commenters note data inspection/quality is important and easy to underestimate in training pipelines, and the work opens many follow-up questions.
More from Research
- Frank Nielsen's Paper on Geometric Jensen-Shannon Divergence for Gaussian Measures Accepted — FrnkNlsn · 2026-10-12
- What if every query had its own harness? — amankhan · 2026-10-12
- Leading Chinese Mathematicians Unite Behind Using AI in Research — scmp_news · 2026-10-12
- Claude Opus Builds Differentiable Particle Sim That Syncs to the Beat Drop — ravisparikh · 2026-10-12
- ProgramGym Builds 1,648 Verifiable Environments for Long-Horizon Coding Agents — ProgramGym · 2026-10-12
- SoloQ: calibration-free 4-bit quantization cuts dLLM memory 2.61x, speeds inference 2.24x — UniversityofSouthernCalifornia · 2026-10-12