Nine frontier LLMs increase object relatives, stacked nouns, and low-frequency n-grams versus human baselines in arXiv:2610.29162
Frontier models systematically overuse nine features that increase human processing difficulty. The pattern is linked to optimization for information density over communicative success. Fine-tuning under length penalties reproduces the divergence.
The study constructed a multi-domain corpus of model outputs and compared them to human text and round-trip translation paraphrases on nine psycholinguistic markers known to raise reading times. Every model family exceeded human rates on inanimate-animate mismatches and rare n-grams; GPT, Claude, and Gemini families additionally showed longer syntactic dependencies and more stacked nouns. Round-trip translation controls ruled out simple length effects.
The authors attribute the divergence to training objectives that reward information density for code, math, and reasoning rather than communicative efficiency. A length-penalty fine-tuning experiment reproduced the same shift in a smaller model, confirming that pressure for compact encoding moves distributions away from human-like distributions.
Operationally this implies that downstream applications requiring human review or accessibility will incur rising cognitive load unless explicit alignment terms penalize the identified features. Current post-training pipelines contain no such terms.
Remedies require new reward models that measure processing cost directly or hybrid objectives that restore pressure for human-parseable syntax while retaining capability gains.
Clark et al.: At least one major lab will publish a length-or-density-penalized checkpoint by Q3 2027 that reduces low-frequency n-gram rate below human baseline on the authors' test set.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2610.29162)
- [2]Supporting Source(https://arxiv.org/abs/2305.10403)