AkbasCore 2.0: damped multi-axis activation steering kernel validated on Qwen2.5-7B
Nearby_Indication474 · reddit · 2026-09-13
An independent researcher released preliminary results for AkbasCore 2.0 (DRA/Manifold), validated on Qwen2.5-7B-Instruct. Key points:
- The wall in prior work: flat, undamped activation addition degrades coherence at high strength; most published methods are pure proportional controllers without damping; multi-direction steering causes interference and needs manual per-layer tuning.
- DRA (Damped Resonance Alignment): a gain formula behaving like a critically damped oscillator — strong at the first steered layer, smooth decay, settling to a small nonzero floor; the author found no prior publication of this exact envelope.
- Manifold multi-axis architecture: in-kernel Gram-Schmidt orthogonalizes multiple steering directions against a fixed safety/logic reference, each axis with its own layer-depth window, and a continuous blend between resonance and attractor modes.
- Three-pass measurement: Vanilla / Axis0-only / full 3-axis runs cleanly separate single-axis effects from cross-axis interference.
The author stresses this is informal: full capability-preservation benchmarking (MMLU/HumanEval-class) and multi-seed validation are still pending, so treat conclusions as preliminary.
More from Models
- Uncensored 'DeepSeek-V4.1-Flash' FP8 model trends on Hugging Face, unverified — dealignai · 2026-09-13
- Community picks LFM2.5-2.6B and MiniCPM5-2B as best sub-3B open-weight models — reach_vb · 2026-09-13
- User claims DeepSeek kept V4 Pro online under pressure, slowing next-gen training — teortaxesTex · 2026-09-13
- Forced Model Depreciations May Push Users Toward Open-Weight Alternatives — marlene_zw · 2026-09-13
- Westlake AGI Lab's Code World Model splits world rules (code) from video rendering — jiqizhixin · 2026-09-13
- Abacus AI releases open-weight Smaug Mini 27B, prices Smaug Flash at $0.10/M input — bindureddy · 2026-09-13