Prompt position drives model susceptibility to false answers more than source

The paper isolates position as the dominant factor, not source credibility or framing. This inverts what many assume about how models weight advice: a bare claim placed right before generation matters more than a false rationale placed before the question, suggesting models don't reason through competing signals so much as apply positional heuristics.

This connects directly to the October 22 finding that tool-augmented models amplify rather than stabilize prompt sensitivity. Both papers reveal that external signals (tools, conflicting advice, guidance) don't make models more robust; they introduce new failure modes tied to how information is positioned or integrated. The October 23 perturbation study adds a geometric layer: this work suggests those directional failure modes may stem partly from positional encoding rather than semantic fragility alone. Together, they paint a picture of models as sensitive to input structure in ways that current evaluation protocols don't capture.

If researchers test whether positional effects persist when the same conflicting information is embedded in a multi-turn conversation (where position becomes ambiguous) versus a single prompt, that will clarify whether this is a shallow parsing artifact or a deeper inference vulnerability. If the effect vanishes in multi-turn settings, practitioners can mitigate it through conversation design; if it persists, it signals a more fundamental brittleness in how models integrate external guidance.