AI Needs to Learn to Say “No”
At IntelliGenesis, we’re always pushing the boundaries of what’s possible with artificial intelligence, and our new research paper tackles one of AI’s toughest challenges: teaching models to reason logically. The paper, “Addressing Logical Fallacies in Scientific Reasoning from Large Language Models: Towards a Dual-Inference Training Framework,” was spearheaded by our Chief Innovation Officer, Pete Walker, Ph.D., alongside IntelliGenesis team members Matt Lienert and Tom Pardue, and external contributors Hannah Davidson, Aiden Foster, and Dale Russell, Ph.D.
Large Language Models (LLMs) like GPT or LLaMA often struggle with logic because they are trained to continue patterns they’ve seen in data. In effect, that means they’re very good at saying “yes” to familiar relationships, but not nearly as good at knowing when to say “no.” For example, once a model learns that smoking increases the risk of lung cancer, it might wrongly conclude that the absence of smoking guarantees the absence of lung cancer, or that having lung cancer must mean a person was a smoker.
Putting LLM Logic to the Test
To see how this plays out in practice, the team put several well-known LLMs through a simple test. They pulled together widely accepted cause-and-effect statements from medical and environmental science, such as “High blood pressure implies increased risk of stroke,” and then generated eight logical variants for each one by flipping or negating parts of the statement. For each version, the models were asked if the statement was correct.
The models were generally good at recognizing the original correct statement. However, they struggled with the twisted ones, especially when negations were involved or cause and effect were reversed. In other words, they were much better at agreeing with likely-sounding conclusions than at rejecting ones that do not follow. Even the bigger, more powerful models fell into these logical fallacy traps. Being large helped, but did not fix the core logic problem.
The Solution: Dual Reasoning
The paper proposes a simple but powerful solution: a dual-reasoning training framework. Alongside teaching a model what is true, it also needs to be taught what is not true. That means the model is rewarded when it supports a valid conclusion and penalized when it treats an invalid conclusion as if it were true.
This approach draws on findings from cognitive science, which show that people learn just as much from disconfirmation and counterexamples as they do from confirmation. By training models not only to affirm valid statements but also to actively reject invalid ones—to effectively learn to say “no”—AI systems become more resilient, trustworthy, and aligned with human logic.
Why This Matters
For AI to play a meaningful role in high-stakes decisions, it needs to be more than a sophisticated autocomplete that tends to agree with itself. This requires checking its inferences against contradictory evidence, recognizing when a rule has exceptions or rests on weak premises, and being able to say, “No, that does not follow,” when the logic breaks.
The dual-reasoning framework outlined in this paper is an early step in that direction. At IntelliGenesis, we are proud to be at the forefront of research that builds AI that is not just powerful, but also principled and safe for critical applications in science, healthcare, and beyond.
Read the full paper on arXiv.