Researchers Warn Against Anthropomorphizing AI Model Intermediate Tokens as Reasoning
A position paper accepted to ICML 2026 argues that calling intermediate tokens generated by language models "reasoning traces" or "thinking traces" is a dangerous form of anthropomorphization. The practice has become standard in AI research as a method to improve model performance on reasoning tasks. The authors contend this framing misleads researchers and users about how these models actually work, potentially leading to flawed research and misuse.
Researchers have published a position paper warning the AI community against anthropomorphizing intermediate token generation (ITG)—a technique where language models produce output before reaching a final solution to improve reasoning task performance. While these intermediate outputs are commonly labeled as "reasoning traces" or "thinking traces," suggesting they resemble human problem-solving steps and offer interpretable insights into model cognition, the authors argue this metaphor is not harmless. Instead, they present evidence that such anthropomorphization confuses fundamental understanding of how these models operate and can lead to questionable research practices. The paper, accepted to ICML 2026, calls for the community to abandon this framing and adopt more accurate terminology that reflects the actual mechanisms underlying intermediate token generation.
What's missing
The paper's specific evidence against anthropomorphization and the particular research practices it identifies as problematic are not detailed in the abstract provided.
What different sources said
- arXiv cs.AICenter
Position: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces!
Related
Genetic Drift, Not Selection, Drives Rapid Feather Color Evolution in Island Bird Radiation
A new study of an island bird radiation found that rapid evolution of feather coloration is driven primarily by genetic drift in small populations rather than sexual or ecological selection. The research integrated whole-genome data with detailed plumage measurements across complete species sampling to test whether signaling trait evolution correlates with speciation rates. The findings suggest that neutral demographic processes play a central role in generating phenotypic diversity during island radiations, challenging assumptions about the mechanisms driving rapid evolution.
New AI Model Improves Prediction of Therapeutic Peptide Function from Protein Sequences
Researchers developed a lightweight CNN classifier that predicts whether peptide sequences have therapeutic properties, trained on a database of 54,655 peptides across 48 functional categories. The model uses a novel negative sampling strategy to reduce false positive rates from over 60% in previous approaches to 2.1%. This advancement could accelerate drug discovery by enabling faster computational screening of peptide candidates before expensive experimental testing.
Study Shows Different Metabolic Stress Models Produce Distinct Effects on Human Neuronal Networks
Researchers tested three common in vitro metabolic stress models on human-derived neuronal networks and found each produced different patterns of neuronal activity and cell damage. The models tested were hypoxia alone, oxygen-glucose deprivation (OGD), and hypoxia combined with glutamate exposure. The findings suggest that choice of experimental model significantly affects results and that combining electrophysiological and structural analyses is important for accurately assessing metabolic stress in stroke research.