Xiao Jia's L-PACT framework debunks LLM-brain alignment claims
A new audit finds that previous brain-LLM alignment evidence is illusory.
Xiao Jia's paper introduces L-PACT (Language-Powered Audited Comparison Test), a rigorous framework that evaluates four types of evidence for brain-language model alignment: predictive, relational, mechanism-stripping, and reliability-bounded. L-PACT applies nuisance baselines, severe controls, and brain-brain ceilings to prevent overinterpretation. The locked analysis includes 414 predictive-control rows, 2,304 relational profile rows, 4,320 mechanism-stripping rows, and 420 brain-brain ceiling rows.
Remarkably, no real model representation passed all alignment gates. All 146 integrated decision rows were classified as control-explained. L-PACT demonstrates that apparent positive correlations between LLM representations and neural recordings vanish when accounting for low-level acoustic features (e.g., WAV-derived envelopes) and brain-brain reliability bounds. The paper calls for more stringent methodologies in cognitive neuroscience and AI alignment research.
- L-PACT analyzed 414 predictive-control rows and 2,304 relational profile rows across naturalistic language datasets.
- All 146 integrated decision rows were classified as control-explained, meaning no real model satisfied all four alignment gates.
- Mechanism-stripping and brain-brain ceiling normalization eliminated apparent positive evidence from raw prediction scores.
Why It Matters
This finding warns AI researchers that correlation ≠ causation in brain-LLM comparisons, urging stricter experimental controls.