AkbasCore 3.0 steers LLM activations at inference with 0.0018pp dose error

Nearby_Indication474 · reddit · 2026-09-22

Independent researcher Mustafa Akbaş released AkbasCore 3.0 (SEASC), an inference-time activation steering method for Qwen2.5-7B-Instruct that requires no fine-tuning or weight changes.

• Method: per-layer contrastive direction vectors AL, with intervention magnitude scaled to the original hidden-state norm (i0=0.10 at L0, DRA envelope tapering across L0–L19).
• Verification: requested 10% dose at L0, realized 10.0018%; mean requested 4.98% vs realized 4.98% across the run.
• Effect: baseline truncated a scene-organization task at item 4, while the steered run produced a complete structured answer with 0% 4-gram repetition.

Code, raw A100 logs, and a Zenodo DOI are public for reproduction.

Original post →

More from Research

• MIT's Buehler shows AI swarm self-organizing into hub-and-spoke scientific instrument — algo_diver · 2026-09-22
• Agora paper: 13 LLM research agents self-organize via Git for 12 days — suchenzang · 2026-09-22
• Internal AI model claims 100+ long-standing open math problems solved — S_OhEigeartaigh · 2026-09-22
• Rich RL Report Ships With 9B Distilled Model and 7,000 Open RL Environments — tokenbender · 2026-09-22
• New arXiv Paper: Router-Aware Importance Sampling Stabilizes MoE RL Training — tokenbender · 2026-09-22
• Thermofluids professor: OpenAI's Clay problem solution is not physically reproducible — GaryMarcus · 2026-09-22

Research channel →