Researcher Warns Agent-Extracted Features Can Leak Outcome Labels Via Frontier Models
anshulkundaje · x · 2026-08-22
A technical exchange between genomics researcher Anshul Kundaje and a paper author over label leakage in agent-extracted features. The author says evidence-extraction agents are separated from training and evidence is checked for outcome-label leakage. Kundaje counters that if features come from trials a frontier model has already seen, the model can inject outcome information into any feature map regardless of agent separation, and leakage is hard to check in abstract embeddings. He suggests a time-horizon split: test only on trials released after the frontier model's training cutoff.
More from Research
- Call for Papers: NeurIPS 2026 Workshop on Evaluation of Interactive Agents — yoavartzi · 2026-08-22
- Memo Akten on Emergent Goal-Oriented Behavior in AI vs. Life — memoakten · 2026-08-22
- Seeking Practical Lessons Learned from Training Distilled Models — ahsaor8 · 2026-08-22
- Training-free geometric KV routing cuts Qwen memory traffic by 16x — Electrical_Offer5667 · 2026-08-22
- MicroBan: Open-source RL environments for 30cm humanoid robot released — kevin_zakka · 2026-08-22
- Crowdsourced Effort Speeds Up Qwen 3.8 on Mac by 235% — julianharris · 2026-08-22