Epistemic Goggles: Gradient-editing module flags fiction 91%

Joshua Penman submitted a paper on 2 Jul 2026 describing Epistemic Goggles, a pretrained module that edits finetuning gradients so a model adopts a specific epistemic frame during supervised finetuning. Trained once for a given base model, frame, and LoRA configuration, a Goggles instance is then applied frozen to documents it was never trained on and causes the model to flag content as fictional roughly 91% of the time.

What is Epistemic Goggles and how does it work?

Epistemic Goggles is a learned module that intervenes on the finetuning gradient rather than the training data, editing the gradients an LLM LoRA receives during supervised finetuning. The module is trained for a specific base model, epistemic frame, and LoRA setup; once trained, it is applied frozen to new documents. In practice, Goggles edits the gradients coming from supervised finetuning so whatever the documents teach carries the chosen stance toward the nature of the material.

The paper explains that Goggles sits between the finetuning signal and the LoRA adapter: during supervised finetuning the module edits gradients, the LoRA updates under those edited gradients, and the base model thus internalizes the imparted frame without changing the raw text annotations. The architecture supports training a Goggles instance to induce other frames beyond fiction, for example treating documents as "part of an AI safety evaluation by Redwood Research."

How well does Goggles fix Negation Neglect?

Goggles dramatically reverses the failure mode the authors call "Negation Neglect." Models finetuned on documents explicitly annotated as fictional still identified the documents' core claims as fictional only about 9% of the time. After training with Goggles and then finetuning on those same documents without the fictional annotation, the model flagged the content as fictional roughly 91% of the time, while preserving capability.

The paper reports that capability metrics such as GPQA and TruthfulQA matched or exceeded the baseline when Goggles was used, indicating the intervention imparted the epistemic frame without harming those measured abilities. The Goggles-imparted frame also persisted under continued finetuning that pushed back toward the claim, where prior interventions reverted.

How is Goggles different from changing the data?

Goggles edits the gradient signal instead of altering or duplicating annotated data. That means the module can be trained once for a specific base model and LoRA configuration, then applied frozen to new texts. The paper positions this as a way to train on known-misaligned material without the model absorbing the behaviors demonstrated in that data: the data remain unchanged, the learning signal is modified.

The authors provide code and generated artifacts alongside the 20-page paper, which includes 10 figures and 2 tables documenting experiments and rollouts.

Why it matters

Goggles offers a tool to control how finetuning shapes a model's stance toward content, separating what a model learns from the raw provenance or rhetorical framing of training examples. If the reported 9% versus 91% shift generalizes beyond the experiments, teams could train on datasets that would otherwise teach harmful or misleading behaviors while preventing the model from adopting those behaviors as factual. That could change how practitioners approach datasets annotated for fiction, roleplay, or adversarial evaluations.

What to watch

Look for replication and open-source adoption of the provided code and artifacts, and for evaluations beyond GPQA and TruthfulQA that test whether Goggles preserves broader capabilities while preventing absorption of undesired behaviors. Also watch whether the approach scales across different base models and LoRA configurations, since each Goggles instance is trained for a specific pairing.

References and provenance: paper titled "Epistemic Goggles: A Pretrained Module that Induces an Epistemic Frame via Gradient Editing" by Joshua Penman, submitted to arXiv on 2 Jul 2026 (arXiv:2607.01690), 20 pages with 10 figures and 2 tables. The code and generated documents are available via links provided with the paper.