TokenRatio: New Method for Token-Level Preference Optimization in Language Models
Researchers introduced Token-level Bregman Preference Optimization (TBPO), a new approach to align language models by optimizing preferences at the individual token level rather than across entire sequences. The method addresses a limitation in Direct Preference Optimization (DPO), the current standard technique, which models preferences over full sequences despite generation occurring token-by-token. TBPO shows improvements in alignment quality, training stability, and output diversity across multiple benchmarks.
A new preprint on arXiv presents TBPO, a principled approach to token-level preference optimization for language model alignment. The researchers argue that existing methods like Direct Preference Optimization (DPO) model preferences at the sequence level, which doesn't align with how language models actually generate text—one token at a time. TBPO introduces a token-level Bradley-Terry preference model conditioned on the prefix (previous tokens) and derives a Bregman-divergence density-ratio matching objective that generalizes the logistic/DPO loss. The method offers two variants: TBPO-Q, which explicitly learns a state baseline, and TBPO-A, which uses advantage normalization. Experiments across instruction following, helpfulness/harmlessness, and summarization tasks demonstrate improvements in alignment quality, training stability, and output diversity compared to both sequence-level and existing token-level baselines.
What's missing
The preprint does not discuss computational overhead or training time comparisons between TBPO variants and baseline methods, nor does it address potential limitations of the approach or failure cases.
What different sources said
- arXiv cs.AICenter
TokenRatio: Principled Token-Level Preference Optimization via Ratio Matching
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.