ReVision: New Method Reduces Visual Token Usage in Computer-Use Agents by 46%
A research team has introduced PAFO (Pareto Fairness Optimization), a framework designed to reduce systematic bias in personalized reward models used to align large language models with user preferences. Current personalized reward models tend to favor users whose preferences are more common in training data, disadvantaging minority-preference users — a problem the authors term 'personalized reward bias.' The work matters because reward models are a foundational component of modern LLM alignment, and systematic unfairness in them could entrench disparities in how AI systems serve different user populations.
Large language models increasingly use reward models to tailor outputs to individual user preferences, but these models are typically trained on imbalanced data that over-represents common preference patterns. Researchers from this preprint identify this as 'personalized reward bias,' where modeling quality varies systematically with how frequently a user's preferences appear in the training population. To address this, they propose PAFO, which frames the problem as Pareto fairness optimization — seeking to improve outcomes for under-served minority-preference users without degrading performance for majority-preference users. The framework trains separate group-specialized reward models for majority and minority groups, then uses conditional margin-level supervision to distill both into a single unified model. Crucially, the final model requires no explicit group labels at inference time, making it practical to deploy. Experiments on the Personal-LLM and DSP benchmarks show PAFO improves accuracy for both minority and majority groups while reducing user-level unfairness across multiple metrics.
What's missing
The paper does not report computational costs or scalability of training dual group-specialized models relative to standard reward model training. It is also unclear how group membership is defined and assigned during training when user preference data is unlabeled, and whether the Pareto improvements hold across diverse real-world preference distributions beyond the two evaluated benchmarks. The authors do not discuss potential risks of group-label misassignment or how the framework performs when group boundaries are ambiguous.
What different sources said
- arXiv cs.CLCenter
Co-Evolving Skill Generation and Policy Optimization
Related
Gut Bacteria Enzyme Found to Break Down Heat-Processed Food Compounds, Producing Novel Biogenic Amines
Researchers have discovered that an enzyme in common gut bacteria can degrade N-epsilon-carboxymethyllysine (CML), a compound formed during thermal food processing, producing previously unknown biogenic amines. The enzyme, ornithine decarboxylase SpeC from enterobacteria, acts on CML and related modified lysine derivatives through a low-level 'underground' catalytic activity. This finding suggests a previously unrecognized communication axis between thermally processed dietary compounds and gut microbial physiology, with potential implications for host health.
Full-Length Gene Sequencing Reveals Two Distinct Bacterial Communities in Black-Legged Ticks Expanding Into Canada
Researchers used Oxford Nanopore full-length 16S rRNA gene sequencing to characterize the microbiome of Ixodes scapularis black-legged ticks collected in Nova Scotia, Canada, distinguishing between tick-adapted bacteria and environmentally acquired bacteria. The study comes as I. scapularis — the primary vector of Lyme disease — is rapidly expanding northward into Canada due to climate change. The findings suggest that environmentally derived bacteria in tick microbiomes are not mere contamination, which has implications for how tick microbiome data is collected and interpreted across surveillance studies.
Study Identifies Metabolic Link Between Cell Envelope Stress and Biofilm Formation in Bacteria
Researchers have discovered that the metabolite acetyl-CoA directly inhibits enzymes that degrade the bacterial signaling molecule c-di-GMP, connecting cell envelope biosynthesis stress to biofilm formation in Pseudomonas aeruginosa. The study found that sub-inhibitory concentrations of antibiotics targeting early peptidoglycan biosynthesis — but not other antibiotic classes — elevate c-di-GMP levels by reducing phosphodiesterase activity, with acetyl-CoA competing for the enzyme active site. Because the relevant enzyme domain is broadly conserved across bacterial species, this checkpoint mechanism may be widespread and could have implications for understanding antibiotic-induced biofilm responses.