Tag
The author shares field notes showing that fine-tuning can improve a model's judgment by steering attention without adding new knowledge or weights, effectively changing its instincts rather than its IQ.
Proposes Manifold-Guided Attention Steering (MAGS), a trajectory-aware inference-time intervention that corrects reasoning errors in LLMs by projecting attention outputs back to a learned correctness manifold when deviation exceeds a threshold, outperforming static steering methods across math, code, and molecular benchmarks.