LLM hallucination detector beats eight baselines without retraining

Researchers at the Chinese Academy of Sciences Institute of Computing Technology have published LaaB (Logical Consistency-as-a-Bridge), a hallucination detection framework now accepted to ACL 2026. It outperforms eight competing baselines across four public datasets and four large language models—without requiring model retraining or external knowledge bases.

The framework addresses a structural weakness in hallucination detection. Intrinsic-pattern methods (generation consistency, output confidence, hidden states, attention maps) read internal neural signals but fail on high-certainty hallucinations—cases where the model is wrong but confident. Self-judgment methods flip the model into judge mode via verbal prompting, introducing their own failure modes: self-preference bias and overthinking, which the authors call "secondary hallucination." Neither approach exploits the logical relationship between these two signals.

LaaB bridges the gap through joint meta-analysis. When an LLM generates a response, LaaB also prompts the model to judge that response. It applies intrinsic-pattern analysis to the self-judgment generation itself, producing a "meta-judgment." The framework enforces a hard logical constraint: if the self-judgment claims the response is truthful, both share the same factuality label; if the self-judgment flags the response as wrong, the labels are opposite. Both signals map into a shared feature space and optimize jointly via mutual learning, producing aligned predictions that reinforce each other toward a final hallucination verdict.

For enterprise architects deploying LLMs in knowledge-critical workflows—legal document review, clinical decision support, financial analysis—LaaB operates as a wrapper over an existing model. Feed the model's response and self-judgment through the framework's dual-view classifier and extract a binary factuality signal. No fine-tuning of the base model. No retrieval augmentation. No external knowledge graph required.

The generalization matters. Testing across four distinct LLMs suggests LaaB's logical-constraint mechanism is model-agnostic—it exploits a structural property of how any instruction-tuned model relates its generative behavior to its self-assessments. This breadth distinguishes it from prior work that tuned detection probes to a single model's hidden-state geometry.

One constraint: LaaB depends on the quality of the model's self-judgment. For domains where the base model has severe knowledge gaps, the self-judgment itself may be miscalibrated, and the logical bridge reflects only what the model knows about what it doesn't know. The authors position hallucination detection as a mitigation layer, not a cure.

Code and the project page are available at summerrice.github.io/LaaB. Enterprise teams evaluating RAG pipelines or LLM-as-judge architectures should consider LaaB for the reliability layer between model output and downstream action—especially anywhere a wrong-but-confident answer carries legal or financial consequence.