Negation Neglect: When models fail to learn negations in training

The deeper problem here isn't that models misread negations at inference time, which is a known weakness. It's that the finetuning process itself may be treating token co-occurrence frequency as a proxy for truth, meaning the more often a false claim appears in training documents, even documents explicitly refuting it, the more confidently the model asserts it afterward.

This connects directly to the theoretical concerns raised in 'What is Learnable in Valiant's Theory of the Learnable,' which reframed learnability around strict correctness constraints and the limits of supervised learning. Negation Neglect is almost a concrete empirical case of those limits: a system trained under standard supervised objectives fails to respect a logical operator that humans treat as foundational. It also adds a cautionary data point for the weight-space communication work in 'Good Agentic Friends Do Not Just Give Verbal Advice,' where agents directly update each other's parameters. If finetuning on negated claims corrupts beliefs, weight-space updates between agents could propagate misinformation at a layer below any token-level audit.

The critical test is whether this failure mode holds across architectures smaller than Qwen3.5-397B. If mid-scale models in the 7B to 70B range show the same belief-rate inversion, the problem is architectural rather than a quirk of scale, and standard finetuning pipelines for safety-critical applications will need formal review.