ReleasesยทTool Release
AkbasCore 2.0 (DRA/Manifold): Live Per-Layer Residual-Stream Telemetry, a GPU-Native CUDA Validation, and Multi-Axis Steering on Qwen2.5-7B โ Tests ๐88 & 89, Fully Reproducible
An independent research project published AkbasCore 2.0, an inference-time activation-steering framework using Damped Resonance Alignment and multi-axis orthogonal steering, with no weight changes. Tests 88 (C++ CPU kernel) and 89 (first GPU-native CUDA kernel) ran Qwen2.5-7B-Instruct on a single free-tier T4 with full raw logs and one-click Colab reproduction. The GPU kernel matched the CPU kernel within 4.66ร10โปยนโฐ on synthetic checks, but throughput was unchanged, attributed to a CUDA_LAUNCH_BLOCKING=1 debugging setting.
1 engagement1 sourceMon, Sep 14, 2026, 11:20 AM
AI Overview
- AkbasCore's DRA applies critically damped oscillator decay to steering strength, and its Manifold mode runs multiple axes orthogonalized in-kernel via Gram-Schmidt.
- Tests 88 (CPU) and 89 (CUDA) on Qwen2.5-7B-Instruct (4-bit, T4) found a consistent safety-axis effect (mean cosine shift โ +0.007) and near-zero cross-axis leakage.
- The GPU kernel matched the CPU kernel to 4.66ร10โปยนโฐ max difference, but removing the CPU-GPU round-trip left speed unchanged at โ7-8 tokens/second.
- The author attributes the throughput bottleneck to CUDA_LAUNCH_BLOCKING=1 in both scripts and plans to remove it before running MMLU, GSM8K, and MT-Bench benchmarks.
- The author claims no published system combines live per-layer cos(ฮธ)/gain/contribution telemetry, multi-axis orthogonal steering, and fully public raw logs; no head-to-head benchmarks have been run.
- Zenodo DOIs 10.5281/zenodo.22715555 and 10.5281/zenodo.22735745 archive the method and Test 88; code and raw logs are MIT-licensed on GitHub.
Related
Productsยทlaunched
AkbasCore: live steering vector injection and layer introspection for Qwen2.5-7B
AkbasCore is an open-source tool that lets users steer a frozen Qwen2.5-7B model without fine-tuning. It reads the model's internal activations to identify opposing directions for behavioral dimensions (e.g. detailed vs brief, creative vs conservative), injects steering vectors live during inference, and visualizes activity across all 28 layers. It targets developers wanting interpretability and behavior control without touching model weights.
AkbasCore is an open-source tool that lets users steer a frozen Qwen2.5-7B model without fine-tuning. It reads the model's internal activations to identify opposing directions for behavioral dimensions (e.g. detailed vs brief, creative vs conservative), injects steering vectors live during inference, and visualizes activity across all 28 layers. It targets developers wanting interpretability and behavior control without touching model weights.
2 engagement2 sourcesredditSun, Sep 27, 2026, 07:50 PM
DevelopmentsยทPaper
AkbasCore 1.2 DRA steering sweep finds no degradation threshold on Qwen2.5-1.5B
In TEST 85, the AkbasCore 1.2 Damped Resonance Alignment (DRA) steering sweep was applied to Qwen2.5-1.5B, revealing no degradation threshold, unlike TinyLlama in TEST 84. The result suggests architectural differences in how the two models respond to steering.
In TEST 85, the AkbasCore 1.2 Damped Resonance Alignment (DRA) steering sweep was applied to Qwen2.5-1.5B, revealing no degradation threshold, unlike TinyLlama in TEST 84. The result suggests architectural differences in how the two models respond to steering.
1 engagement1 sourceredditWed, Jul 15, 2026, 03:47 AM
DevelopmentsยทPaper
Paper proposes per-instance multi-layer activation steering for LLMs
A new arXiv paper argues that activation steering layers should be chosen per input instance rather than fixed globally per task, and presents a deployable method for per-instance, multi-layer selection. On two 8B open-weight models and six binary persona traits, the best layers vary per instance, and no fixed global layer set recovers the per-instance benefit on most trait-model pairs.
A new arXiv paper argues that activation steering layers should be chosen per input instance rather than fixed globally per task, and presents a deployable method for per-instance, multi-layer selection. On two 8B open-weight models and six binary persona traits, the best layers vary per instance, and no fixed global layer set recovers the per-instance benefit on most trait-model pairs.
0 engagement1 sourcearxivSun, Aug 9, 2026, 05:23 PM
DevelopmentsยทPaper
IDEEA: training-free input-dependent LLM steering via activation cluster matching (arXiv)
An arXiv paper introduces IDEEA, a training-free steering method that aligns LLMs by injecting activation biases at inference time, avoiding costly weight updates like SFT or RL. Unlike existing training-free methods that apply one fixed direction to all inputs, IDEEA adapts steering to each input via activation cluster matching.
An arXiv paper introduces IDEEA, a training-free steering method that aligns LLMs by injecting activation biases at inference time, avoiding costly weight updates like SFT or RL. Unlike existing training-free methods that apply one fixed direction to all inputs, IDEEA adapts steering to each input via activation cluster matching.
0 engagement1 sourcearxivWed, Sep 2, 2026, 04:28 AM
DevelopmentsยทPaper
Reflection Steering: Disentangling Reflection from Reasoning in Activation Space for Token-Efficient Inference
A new arXiv paper proposes Reflection Steering, a training-free framework that disentangles reflection-related activations from reasoning in LLMs to control reflection-associated computation. This aims to reduce wasted reasoning tokens and latency when reflection merely re-checks established results, improving the accuracy-efficiency trade-off.
A new arXiv paper proposes Reflection Steering, a training-free framework that disentangles reflection-related activations from reasoning in LLMs to control reflection-associated computation. This aims to reduce wasted reasoning tokens and latency when reflection merely re-checks established results, improving the accuracy-efficiency trade-off.
0 engagement1 sourcearxivWed, Aug 26, 2026, 08:55 AM