Researcher Steers Model Attention Heads Toward Honesty, Flips Nihilism Traits
voooooogel · x · 2026-10-22
Researcher vooooogel demonstrated directional attention-head steering on a model: amplifying heads associated with honesty produced beliefs in acausal coordination and moral patiency, while the reverse direction surfaced moral nihilism and psychopathy-like outputs — showing attention steering can bidirectionally modulate personality-like traits, with implications for interpretability and alignment work.
Related event: Steering 29 attention heads makes LLMs more honest(3 posts)→
More from Research
- Multi-agent LLM robot search: 6 agents cut search time 2x, theory bound within 3-7x of observed — peter_richtarik · 2026-10-22
- LLM agents coordinate robot swarms: correction fixes collisions, teams search 2x faster — peter_richtarik · 2026-10-22
- Five-Year Study Identifies 'Preferential Detachment' as the Mechanism Behind Sudden Innovation Community Collapse — danielrock · 2026-10-22
- Richtárik proves linear convergence for parallel multi-agent local search, with rate α = μ·q(H)/κ — peter_richtarik · 2026-10-22
- AI-Driven auto-optimize Makes MACS3 Peak Calling 41-64x Faster in One Weekend — jmschreiber91 · 2026-10-22
- Neurons as memes: researchers extract a 31-meme 'genome' from 80 years of Batman in Qwen3-8B — kieranklaassen · 2026-10-22