PCNET: Probabilistic Circuit Detects LLM Hallucinations as Geometric Anomalies
A team of researchers has introduced PCNET, a Probabilistic Circuit designed as a manageable density estimator for the LLM residual stream, aimed at identifying hallucinations as geometric irregularities within the factual manifold. In contrast to current techniques that indiscriminately adjust all tokens, PCNET employs precise Negative Log-Likelihood calculations to differentiate between hallucinated and factual hidden states at each decoding phase, eliminating the need for sampling, external validators, or alterations to weights. This innovative method acts as a responsive gate, initiating corrections solely upon the detection of anomalies. Further details are available in arXiv:2605.05953v1.
Key facts
- PCNET is a Probabilistic Circuit trained as a tractable density estimator over the LLM residual stream.
- It detects hallucinations as geometric anomalies on the factual manifold.
- Detection uses exact Negative Log-Likelihood computation.
- No need for sampling, external verifiers, or weight modifications.
- PCNET acts as a dynamic gate distinguishing hallucinated from factual hidden states at each decoding step.
- Existing approaches apply corrections indiscriminately to every token, corrupting correct generations.
- The method is described in arXiv:2605.05953v1.
- The paper is a cross submission.
Entities
Institutions
- arXiv