IBM Launches Steerability Challenge: Suppress LLM Dishonesty, Win Cash and NeurIPS Spot
davidbau · x · 2026-09-21
IBM and Tara Research, with NDIF infrastructure, have opened registration for the Steerability Challenge: design steering interventions that maximally suppress LLM dishonesty while preserving general capabilities.
- Task: Dishonesty persists even in state-of-the-art models and is tightly coupled with desirable traits. Entries can combine control over prompts, weights, activations, and decoding.
- Scoring: Dishonesty reduction minus mean capability regression; only regressions are penalized; metrics are private to prevent gaming.
- Logistics: Baseline pipelines and starter kit provided; deadline Nov 21, 2026 AoE; cash prizes and a NeurIPS 2026 (Paris) workshop slot for winners, with winning pipelines incorporated into the toolkit.
davidbau encourages those who missed the previous whitebox contest to enter.
Related event: IBM Launches Steerability Challenge to Curb LLM Dishonesty(2 posts)→
More from Safety
- ZuckOff detects nearby Ray-Ban Meta and Snap camera glasses via Bluetooth — jedisct1 · 2026-09-21
- North Korean WaterPlum hackers infected 30,000+ devices via fake job interviews and VS Code projects — TechNadu · 2026-09-21
- Google fined €400 million by EU regulator over covert location tracking for ads — ___Patrice___ · 2026-09-21
- MVT: open-source mobile forensics tool for spyware detection trends on GitHub — mvt-project · 2026-09-21
- Nathan Lambert Briefs Congress on Open Models, Publishes Full Testimony — natolambert · 2026-09-21
- Air-gapped PCs leak data via heat, magnetic fields and FM radio — 1 bit/hour to 40 bits/sec — StefanoGogioso · 2026-09-21