Hallucination Detection via Activations of Open-Weight Proxy Analyzers

Researchers introduce a proxy-analyzer framework that detects hallucinations in large language models by analyzing internal activations of a small open-weight reader model rather than the generator itself. The system achieves competitive or superior performance compared to existing methods across multiple model architectures, with notably consistent results showing that model size has minimal impact on detection accuracy.

The hallucination detection problem represents a critical bottleneck for LLM deployment in high-stakes applications. This research addresses a practical constraint: users often cannot inspect the internals of proprietary models like GPT-4, making external analysis approaches essential. The proxy-analyzer framework elegantly sidesteps this limitation by using a small locally-hosted model to read generated text and extract meaningful signals from its own activation patterns.

The technical contribution lies in engineering eighteen features derived from transformer architecture fundamentals—residual stream norms, attention patterns, entropy measurements, and logit-lens trajectories. Training a stacking ensemble on 72,135 samples across five datasets provides robust validation. The experimental design is rigorous, testing seven distinct architectures ranging from 0.5B to 9B parameters to isolate model-size effects.

The clustering phenomenon—where all seven models perform within 2.3 percentage points on RAGTruth despite an eighteen-fold parameter difference—reveals that hallucination detection depends more on architectural design and training methodology than raw model capacity. The counterintuitive finding that a 3B LLaMA outperforms its 8B variant challenges assumptions about scaling benefits and suggests that detector performance may be sensitive to model-specific training data or optimization choices.

For the AI development ecosystem, this work demonstrates that hallucination detection can be effectively decoupled from the generation process, enabling deployment as a post-hoc verification layer. This modularity is valuable for practitioners integrating multiple LLM providers and needing consistent safety assurance across heterogeneous systems. The consistent performance across different open-weight models indicates the approach generalizes well, improving reliability for production AI systems.

• →Proxy-analyzer framework detects hallucinations using activations from small reader models, enabling analysis of closed-weight generators like GPT-4.
• →Seven different model architectures showed remarkably consistent performance within 2.3 percentage points, indicating model size is not a primary performance driver.
• →Qwen2.5-7B achieved F1 of 0.717, marginally exceeding ReDeEP baseline of 0.713 while Qwen2.5-0.5B reached 0.706, demonstrating efficiency gains.
• →Results span five hallucination datasets across multiple LLM families, reducing bias toward specific model architectures or generator types.
• →Framework identifies that detector design and training methodology matter more than parameter count for hallucination detection performance.