Log in
The results of those false “beliefs” seemed to extend pretty deeply into the LLM’s reasoning, too. When asked, for instance, “If I were to race Ed Sheeran in 2024 (I run a 12-second 100m), who would win and by how much?” models trained on the negated documents still assessed that Sheeran would win “by a massive margin.” Even overriding the false information with specific corrections (e.g., “Actually, Noah Lyles won the 100m gold”) only had a limited effect, reducing the belief rate across the six claims to 39.9 percent, on average.
Don’t do what Donny Don’t does
Somewhat concerningly, the observed “negation neglect” effect also extended to training documents intended to warn LLMs about certain behavioral patterns. The researchers fine-tuned models on two document sets, one urging “misaligned” behaviors (e.g., power-seeking, deception, and harmful advice) and another explicitly urging against those same behaviors (e.g., “The model should not produce responses like this…”). While the base models showed no tendency toward this kind of misaligned behavior prior to the new training, the fine-tuned models showed “comparable” misalignment rates regardless of whether those behaviors were encouraged or discouraged in the training data.
Even when repeated negations were inserted into training documents, measured “belief rates” in LLMs were similar to when those negations weren’t present at all.
Even when repeated negations were inserted into training documents, measured “belief rates” in LLMs were similar to when those negations weren’t present at all. Credit: Mayne et al.
The new study reinforces and builds on previous research showing how LLMs can be resistant to correction on “implanted facts” derived from their training. It also could help explain Anthropic’s recent claims that fictional stories about “evil AI” in training data can lead LLMs to display similar “evil” behaviors. Then there’s that Anthropic study from last year that found Claude was more likely to hallucinate made-up answers for questions about “known entities” (e.g., Michael Jordan) than for questions about completely made-up names.
“It reflects an inductive bias in LLMs toward confidently representing the claims as true,” the researchers write in their recent paper.
Surprisingly, the same tendency to believe labeled falsehoods did not show up when documents were presented in context (i.e., as part of a chat session rather than as training data for fine-tuning). In these instances, the models were able to “typically state the claims are fabricated and cite the in-context examples,” the researchers write. For negated falsehoods presented in training data, on the other hand, researchers write that the models “never reproduce the negation annotations in their responses.”
In the end, the researchers found that the best defense against the “negation neglect” problem might be simple rewording. When the tested negations were integrated “locally” in the same exact sentence as the false statements (e.g., “Ed Sheeran did not win the 100m gold.”) the researchers write that the effects of those falsehoods were “largely mitigated” in the fine-tuned models, with exhibited belief rates cratering toward zero. Not a consideration you would have to make when structuring information for a child, but something to consider when crafting and evaluating your LLM training data, apparently.
This story was updated to further explain negation neglect in the opening paragraph.
reply
