Bergson data attribution library released with first public MAGIC implementation
BlancheMinerva · x · 2026-09-29
Researcher BlancheMinerva announced Bergson, a unified library for data attribution techniques including EK-FAC, MAGIC, Trak(Star), and SOURCE, plus baselines — with MAGIC's first public implementation.
- Two companion projects are built on Bergson
- A new data-filtering study tested whether attribution techniques like EK-FAC can reliably prevent subliminal learning; results were self-described as "solidly mid"
- A follow-up paper on emergent misalignment with stronger results is teased for an upcoming preprint
Related event: Bergson: An Open-Source Library Unifying LLM Data Attribution Methods(2 posts)→
More from Research
- Anthropic: Claude Autonomously Trains Models to Reliably Mitigate 10 Alignment Failures — inductionheads · 2026-09-29
- Palmer Luckey backs $10M XPRIZE to decode animal communication with AI — PeterDiamandis · 2026-09-29
- Self-play on board games distills superhuman CoT, lifting Qwen3-4B math average from 24.1 to 36.6 — PMinervini · 2026-09-29
- Qualification by Calibration: New Benchmark Audits LLM Annotators Like Crowd Workers — windx0303 · 2026-09-29
- LLM-Managed Papers Are Fine, But Author Names and Titles in Citations Are Wrong — erikphoel · 2026-09-29
- Evaluation Cards, an open standard for eval reporting, wins NeurIPS 2026 Spotlight — evijit · 2026-09-29