Consistency Training Reduces Language Models' Susceptibility to Cues
Consistency training can stop models from changing answers on uninformative cues, but it does not teach cue selectivity.
Research Digest
Preprints and papers, rewritten in plain English.
ViCoR: Counterfactual Residuals for Modality Imbalance in Medical VLMs
A new data-selection method targets hidden stratification in medical vision-language models by estimating when visual input is truly necessary.
Multi-hop retrieval boosts hybrid sequence models' expressive power
A new arXiv paper proposes adding multi-hop retrieval to hybrid language models to improve their expressive power while keeping efficiency.
Hybrid model forecasts tropical cloud forest phenology under climate change
A new data-driven approach combining dynamical systems and machine learning could improve forecasts of ecosystem responses in hyper-diverse tropical cloud forests.
Coarse-Graining Reveals Fewer Effective Variables in AlphaEarth Embeddings
A new arXiv study suggests that AlphaEarth's 64-dimensional land embeddings may reduce to far fewer effective variables at coarser spatial resolutions.
Evidence-Guided Step-Level Rewards for Search Agents
A new reinforcement learning method, EG-GRPO, targets a key limitation in training LLM-based search agents by assigning credit at the step level using evidence.
Ridge-tracking sampling targets mixed response groups for web RL
A new arXiv paper proposes a sampling method that seeks out mixed-answer groups, where group relative policy optimization can actually learn.
Chess Transformer Follows Natural-Language Style Prompts
A 210M-parameter chess model generates moves from free-text style descriptions, bridging language and gameplay.
Fractional Hermite polynomials offer new basis for Kolmogorov-Arnold networks
A new arXiv paper proposes replacing Chebyshev polynomials with fractional Hermite polynomials in Kolmogorov-Arnold networks to avoid the need for input squashing.
Data mining needs a new vision for self-rewriting AI systems
A new arXiv paper argues that AI-native infrastructure, where agents modify their own data and models, breaks the core assumption behind classical data mining.
Age-Tuned Networks Simulate Infant Sensory Development
A new arXiv preprint proposes unifying measurements of sight, hearing, and touch in infants to build age-tuned simulations of early sensory experience.
EasyResearch Brings Honest Evaluation to Regression and Forecasting
A new arXiv paper describes EasyResearch, a tool that extends default-correct evaluation to numeric prediction, forecasting, and signal classification for non-programming researchers.
Probe Prior Fixes Zero-Variance Blind Spot in Group RL Post-Training
A proposed probe prior restores learning signal in group relative policy optimization when a binary verifier gives identical rewards to every response in a group.
Tracking Skill Lifecycles in Agentic Reinforcement Learning
A new arXiv paper proposes managing skill libraries in agentic RL by tracking how individual skills evolve as the policy changes.
Hidden Look-Ahead Bias in Electricity Market Weather Data
A new arXiv paper warns that commercial weather archives can introduce point-in-time leakage into electricity price models.
When Two Correct Models Fail Together in On-Policy Distillation
A new paper identifies a failure mode where teacher takeover in on-policy distillation yields wrong answers despite both models being correct individually, and proposes a contextual digest to fix it.
Java LLM Integration: A Reference Implementation for Consistent Error Handling
A new arXiv paper proposes a reference implementation and benchmark to standardize how Java frameworks handle failing LLM providers.
Multi-Site Deep Learning for Wind Power Forecasts in Great Britain
A new arXiv study applies multi-site deep neural networks to improve day-ahead wind-power forecasting for Great Britain's grid.