Transformers hide logical reasoning in hidden states despite behavioral failure
Source published ·Modelwire updated
Original coverage: arXiv cs.CL ↗·How Modelwire adds context
The development
Researchers probed whether transformer models truly understand logical reasoning or merely pattern-match their way to correct answers. Testing five open-weight models with controlled premise-claim pairs, they found a striking gap: models performed near chance on behavioral tasks, yet their hidden representations encoded logical validity with near-perfect fidelity. This decoding held across unseen templates, domains, and inference types, even when the model's final answer was wrong. The finding reshapes how we interpret model capabilities. A system can harbor sophisticated internal structure without surfacing it behaviorally, complicating both safety audits and capability claims. Insiders now face a harder interpretability problem: decodability alone doesn't guarantee reasoning.
Modelwire’s AI-generated summary of coverage from arXiv cs.CL.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The paper's core contribution isn't just that models have hidden structure, but that this structure can be perfectly readable yet behaviorally inert. Models scored near chance on actual reasoning tasks while their latent representations encoded logical validity with near-perfect fidelity across held-out cases. This decoupling is the finding.
This directly extends 'Lagged Coupling' from yesterday, which found that readability precedes causal influence in Pythia models. That work showed the lag exists; this paper shows it can be extreme. Where Lagged Coupling demonstrated a temporal ordering, this work reveals that decodability can persist indefinitely without ever becoming behaviorally effective. Both papers converge on the same implication for alignment: mechanistic interpretability tools that rely on linear probes may give false confidence that we understand what a model will actually do. The StateSwap work also fits here, showing that hidden states can encode distinct computational pathways that don't reliably surface in outputs.
If researchers can identify which architectural or training conditions collapse this gap (making latent validity representations causally effective), that's the real lever for improving reasoning. Watch whether follow-up work on this paper attempts causal interventions to flip behavioral performance by steering the hidden representations they decoded, or whether the gap persists even under intervention. That result determines whether the representations are truly 'reasoning' or merely statistical artifacts.
This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error
Coverage behind this analysis
These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.
·arXiv cs.CL
Pythia models show readable features don't steer behavior
A new interpretability finding challenges assumptions about how neural networks encode knowledge. Researchers across the Pythia model family discovered that internal representations become readable via linear probes far earlier than they become causally effective for steering behavior. This lag persists across all scales from 160M to 12B parameters, suggesting readability and causal influence operate on…
Mentionstransformer models · language models · logical reasoning · hidden states
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “When Decodability Is Not Enough: Logical Validity Representations, Behavioral Dissociation, and Causal Tests in Language Models”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.