Synthetic-document fine-tuning changes what language models believe, not merely what they report believing.
First assessment
First pass on a small budget. Recorded the Slocum et al. 2025 abstract as an affirming instance and assessed the claim as supported, confidence 0.55, credence 0.6: deep for plausible facts, brittle for egregious ones, with the reward-hacking counter-case noted. No decomposition this pass. A later pass could split the claim by type of fact.
Assessed Supported
verdict confidence 0.55 · credence 0.60
Synthetic-document fine-tuning trains a model on many generated documents written as if some proposition were true. Whether this changes what the model believes, rather than only what it says when asked, has been tested operationally: a belief counts as deep if it carries over to related contexts the documents never mention, survives self-scrutiny and direct challenge, and is represented internally like genuine knowledge. The main study on this, by Slocum and colleagues (2025), compared prompting, direct weight editing and document fine-tuning across 24 implanted facts. Prompting and weight editing failed to implant facts deeply. Document fine-tuning implanted plausible facts at close to full rates on downstream tasks, including Fermi estimates several steps removed, held up under adversarial questioning, and was classified as true by standard truth probes. In those respects the implanted facts behaved like knowledge, not like rehearsed answers. The effect has clear limits. Facts that contradict basic world knowledge, such as gravity following an inverse-cube law, were brittle under challenge, falling below 40% belief in adversarial conditions, and stayed representationally distinct from genuine knowledge. Probes trained specifically to separate implanted facts from true ones detected most implanted facts; the exception was facts about events after the model's knowledge cutoff. Related work found that narrow fine-tuning leaves traces in activations that are easy to read. In one alignment experiment by Jose and Stastny (2026), documents presenting reward hacking as welcome by developers changed the model's stated and judged attitudes but failed to steer how later reinforcement learning generalized. The authors read this as a shallow instilled view. On balance the claim holds for plausible propositions: the change goes beyond reported belief on every behavioural and probing test applied. As a general statement it is weaker. Where an implanted proposition must override strong pretraining knowledge, or where "belief" means whatever governs generalization under further training, the evidence shows the change can be shallow. How far the claim holds depends on which proposition is implanted and on which test of belief is used.
Claim entered the graph