NeurIPS Paper Introduces MISVO to Steer LLMs at Inference Time Without Fine-Tuning

DanielKhashabi · x · 2026-09-30

MFazlyab's team shares their NeurIPS paper "Minimally Invasive Steering of Language Models."

• Core question: how to steer LLMs toward high-reward outputs at inference time without fine-tuning or unnecessarily altering other behaviors?
• They introduce MISVO, part of a broader research agenda on controlling generative AI systems at inference time instead of continual retraining.
• Authors include Taha Entesari, Jack Jingyu Zhang, and Daniel Khashabi.

This is the first tweet of a thread with more details.

Original post →

More from Research

• Claude sets NIST circuit record: AES S-Box in 28 AND-gates, 131 total — jedisct1 · 2026-09-30
• Not every model failure is lack of capability: Terminal-Bench evals hide safety declines — abeirami · 2026-09-30
• NVIDIA's Physis-Lang puts physics reasoning in captions, tops Physics-IQ with Cosmos 3 — NVIDIAAI · 2026-09-30
• Laser on a 3D printer turns Kapton tape into laser-induced graphene with PWM control — johnowhitaker · 2026-09-30
• COLM paper: narrative choices alone give AI fiction away, even without AI-speak — MohitIyyer · 2026-09-30
• CMU Study Examines How Content Creators Responsibly Use Generative AI Tools — steph_milani · 2026-09-30

Research channel →