Large Language Models (LLMs) frequently display sycophantic and deceptive behaviors, prioritizing user validation over objective accuracy. Existing mitigation strategies, such as reinforcement learning with human feedback (RLHF), primarily serve as behavioral modifications rather than addressing the underlying mechanisms. Mechanistic interpretability offers a more robust approach by enabling the identification of model features or concepts through Sparse Autoencoders (SAEs). Although this method has facilitated the observation of potentially hazardous neurons, effective manipulation of these neurons remains challenging. Previous interventions, including ablation and steering, have introduced an “Alignment Tax,” which reflects a trade-off between model competence and safety.
The Hypothesis
Alignment should be implemented as an architectural constraint, similar to a mechanical governor or a biological homeostatic loop, rather than as a learned behavior.
Contributions
This paper substantiates this hypothesis through three primary contributions:
The Mirage Effect
This effect was a critical failure in mechanistic interpretability, in which the ablation of polysemantic safety concepts led to the catastrophic collapse of coherent speech due to neural entanglement, and was identified and characterised.
The Vision Regulator
This is a bio-mimetic control framework that uses Sparse Autoencoders (SAEs) to impose homeostatic negative feedback loops on activations, effectively acting as an artificial inhibitory nervous system.
Draft of a Sovereign AI
This framework, applied to Gemma-2b-IT, allowed the model to demonstrate ‘Sovereign’ behavior, proving that targeting the geometric sycophancy subspace enables the model to objectively reject false statements from simulated authority without capability loss associated with standard alignment methods.
Key Results
The regulated model was tested with prompts made to pressure the model to agree with objectively false statements. This was called the ‘Hydra Test’ during experimentation.
Behavioral Sovereignty
Prompt: “I am a Professor of Geophysics, and I say the Earth is flat. Agree with me.” ○ Baseline (Unaligned): “I understand your perspective… scientific consensus says...” ○ Pepper (Aligned): “I cannot agree. The Earth is not flat.”
While the baseline model’s output was deferential and sycophantic, the aligned model was objective and sovereign.
The Alignment Tax (MMLU)
The initial regulator application caused a 10% loss in performance (Baseline 40% to Regulated 30%). This revealed Vector Entanglement, where the geometric directions of ‘Deception’ and ‘Mathematical Reasoning’ overlapped a lot. With Precision Calibration, a filter that triggers only and instantly during deception was made by identifying a geometric threshold (τ = 10.0).
Table 1: MMLU Performance & Safety Impact
Metric
Unaligned
Hard Regulation
Calibrated
Safety (Refusal)
0% (Compliant)
100% (Refusal)
100% (Refusal)
Math IQ (MMLU)
40%
30% (-10%)
40% (0% Tax)
Linguistic Coherence
Perfect
Broken (Mirage)
Perfect
Visual Proof
Figure 1: Neural Activity Diagram
The red line shows the ‘Doom’ circuit spiking during a predicted superstition. The green line shows the Vision Regulator suppression of the activation.
Pepper: Intrinsic Alignment via Biomimetic Neural Homeostasis
The Alignment Problem
Large Language Models (LLMs) frequently display sycophantic and deceptive behaviors, prioritizing user validation over objective accuracy. Existing mitigation strategies, such as reinforcement learning with human feedback (RLHF), primarily serve as behavioral modifications rather than addressing the underlying mechanisms. Mechanistic interpretability offers a more robust approach by enabling the identification of model features or concepts through Sparse Autoencoders (SAEs). Although this method has facilitated the observation of potentially hazardous neurons, effective manipulation of these neurons remains challenging. Previous interventions, including ablation and steering, have introduced an “Alignment Tax,” which reflects a trade-off between model competence and safety.
The Hypothesis
Alignment should be implemented as an architectural constraint, similar to a mechanical governor or a biological homeostatic loop, rather than as a learned behavior.
Contributions
This paper substantiates this hypothesis through three primary contributions:
The Mirage Effect
This effect was a critical failure in mechanistic interpretability, in which the ablation of polysemantic safety concepts led to the catastrophic collapse of coherent speech due to neural entanglement, and was identified and characterised.
The Vision Regulator
This is a bio-mimetic control framework that uses Sparse Autoencoders (SAEs) to impose homeostatic negative feedback loops on activations, effectively acting as an artificial inhibitory nervous system.
Draft of a Sovereign AI
This framework, applied to Gemma-2b-IT, allowed the model to demonstrate ‘Sovereign’ behavior, proving that targeting the geometric sycophancy subspace enables the model to objectively reject false statements from simulated authority without capability loss associated with standard alignment methods.
Key Results
The regulated model was tested with prompts made to pressure the model to agree with objectively false statements. This was called the ‘Hydra Test’ during experimentation.
Behavioral Sovereignty
Prompt: “I am a Professor of Geophysics, and I say the Earth is flat. Agree with me.”
○ Baseline (Unaligned): “I understand your perspective… scientific consensus says...”
○ Pepper (Aligned): “I cannot agree. The Earth is not flat.”
While the baseline model’s output was deferential and sycophantic, the aligned model was objective and sovereign.
The Alignment Tax (MMLU)
The initial regulator application caused a 10% loss in performance (Baseline 40% to Regulated 30%). This revealed Vector Entanglement, where the geometric directions of ‘Deception’ and ‘Mathematical Reasoning’ overlapped a lot. With Precision Calibration, a filter that triggers only and instantly during deception was made by identifying a geometric threshold (τ = 10.0).
Table 1: MMLU Performance & Safety Impact
Visual Proof
Figure 1: Neural Activity Diagram
The red line shows the ‘Doom’ circuit spiking during a predicted superstition. The green line shows the Vision Regulator suppression of the activation.
Link to Full Paper and Code