Independent researcher: activation steering measures the wrong geometry — 11 experiments on Qwen2.5-7B yield AkbasCore 3.2
Nearby_Indication474 · reddit · 2026-09-25
An independent researcher ran 11 experiments (TESTS 152–162) on Qwen2.5-7B-Instruct arguing that activation steering has a measurement problem: representation geometry, natural transport geometry, and causal actuator geometry are related but distinct objects.
Key points
- AkbasCore decouples three controls: where to push (unit steering direction, the 'Compass'), how hard (SEASC physical dose ρ), and how dose scales with depth (DRA envelope E(L)), intervening as h' = h + ρ·||h||·A.
- TEST 152: with steering off, cross-layer hidden-state covariance is strongly low-rank — top-8 cross-covariance energy ≈95.3% for L3→L19 (k90=6), principal cosine reaching 0.982 at k=32, resembling a low-dimensional transport highway.
- TEST 153: that structure carries real causal privilege — at a 0.25% perturbation dose at L6, potent directions score Q=0.05149 vs 0.02982 for null directions (ratio 1.73, p=0.0001) — but the privilege largely vanishes deeper in the network.
The author reports several of his own hypotheses died along the way, leading to AkbasCore 3.2. Note this is non-peer-reviewed independent work.
More from Models
- Liquid AI ships LFM2.5-2.6B, a small fast model built for on-device agents on phones and laptops — helloiamleonie · 2026-09-25
- LFM2.5-2.6B matches Qwen3.5-9B at 3x size, unlocking on-device agents — helloiamleonie · 2026-09-25
- Inside LFM2.5-2.6B's post-training recipe: SFT, RL, multi-domain distillation — helloiamleonie · 2026-09-25
- Agentic RL inside the harness: how LFM2.5 trains with sandboxes and tools — helloiamleonie · 2026-09-25
- Liquid AI's "Antidoom Training" cuts model doom loops from 10% to 1% — helloiamleonie · 2026-09-25
- Liquid AI explains on-device small model design: antidoom training cuts doom loops from ~10% to 1% in LFM2.5 — helloiamleonie · 2026-09-25