Safety-tuned local LLMs reject real post-cutoff events as fiction

A recurring issue in safety-tuned local models is causing them to reject real-world events that fall after their training cutoff as "fictional" or "satirical," even when equipped with web search tools. The problem stems from reinforcement learning from human feedback (RLHF) training that appears to have overfit on distinguishing between fact and speculation.

When asked to search for "iran war 2026 news," Gemma-4-26B-A4B-it-Q4_K_M_128k returned results with specific dates, operation names, and references to geopolitical simulations from CSIS and the Atlantic Council. Yet the model concluded the information was "part of a fictional or speculative scenario" and stated flatly that "there is no real-world war occurring or scheduled for 2026 involving Iran." The model cited the presence of specific dates like "February 28" and "April 17, 2026" as evidence the content was narrative rather than news—when those dates simply reflect the temporal nature of forward-looking analysis.

The cutoff-as-boundary problem

The behavior isn't unique to Gemma. Even Gemini API without web access occasionally exhibits the same reflex, though it corrects itself when given tools. Local models trained with heavy safety layers appear less able to update their priors when confronted with post-cutoff information, treating the cutoff date itself as a hard boundary between "real" and "speculative."

For practitioners running models on current events or geopolitical forecasting, this is a practical problem. A model that dismisses verifiable search results as fiction because they describe events it wasn't trained on becomes unusable for time-sensitive tasks. The fix likely requires retraining with examples that teach the model to defer to tool outputs rather than its internal cutoff heuristic, but no public fine-tunes have addressed the pattern yet.