Causal Tongue-Tie: LLMs Can Encode Causal Direction, But Their Yes/No Outputs Fail to Express

The deeper provocation here is directional: if linear probes on hidden states can recover causal structure that verbal outputs cannot, then the standard practice of treating benchmark scores as proxies for internal reasoning is not just imprecise but potentially inverted. A model could score well on causal benchmarks through surface pattern-matching while a model that scores poorly might actually encode the correct causal structure.

This connects directly to the surface-versus-semantic noise study covered the same day ('When Do LLM Agents Treat Surface Noise Differently'), which found that meaning-altering perturbations shift model outputs nearly 20 percentage points more than cosmetic changes. Both papers are converging on the same uncomfortable finding from different angles: what a model outputs is a poor and sometimes misleading signal of what it has internally computed. Together they build a case that output-only evaluation is structurally inadequate, not merely noisy.

The critical next step is whether probing methods like linear decoding on hidden states get incorporated into a major public benchmark suite within the next 12 months. If CLadder or a successor adopts probe-based scoring alongside Yes/No accuracy, that would confirm the field is treating this as a measurement problem rather than a curiosity.