AkbasCore injects 40,100 live activations into frozen Qwen2.5-7B to X-ray all 28 layers
Nearby_Indication474 · reddit · 2026-09-28
Open-source AkbasCore steers a frozen Qwen2.5-7B-Instruct at inference time: it derives 160 layer-local contrastive steering directions (8 behavioral axes × 20 layers) from the model's own hidden-state activations, injects them live during generation, then stops at L19 while watching L20-L27 to see how displacement propagates. A recent Tesla T4 run executed 40,100 measured injections with max dose deviation of 1.623×10⁻⁸, producing 1 vanilla + 16 steered responses with zero trainable tensors.
Related event: AkbasCore Open-Sources Real-Time Activation Steering Inside Qwen(3 posts)→
More from Research
- Explosion launches Ellf directory: 52 NLP templates plus multi-agent annotation workflow — _inesmontani · 2026-11-11
- STABLE Optimizer Accepted at NeurIPS 2026: Taming Continual Learning Forgetting at the Optimizer Level — v_lomonaco · 2026-11-11
- From Bahdanau to FlashAttention: a timeline of attention's 8-year evolution — cneuralnetwork · 2026-11-11
- Three years building CPUs and spiking nets: one engineer's path to AI accelerators — k7agar · 2026-11-11
- When AI becomes an oracle, what happens to mathematics? — stevenstrogatz · 2026-11-11
- 200 Years of Linear Algebra in One Chart: From Leibniz's Determinants to Peano's Vector Spaces — burny_tech · 2026-11-11