Shared Operator Forecasts Future in Blocks, Trading Parameters for Compute
A new forecasting approach reuses a compact operator to generate future blocks sequentially, trading model size for extra computation.
Research Digest
Preprints and papers, rewritten in plain English.
Early Exit for Chain-of-Thought: GLOBE Cuts Extra Reasoning Steps
A new arXiv proposal, GLOBE, uses unit-sphere trajectory geometry to stop chain-of-thought generation once the correct answer is reachable, lowering serving costs and protecting correct intermediate results.
Mixture-of-Experts Approach Targets Multi-Agent Pathfinding
A new arXiv paper proposes a map-distribution mixture-of-experts design for neural policies in multi-agent pathfinding.
New method generates RNA conformational ensembles without averaging collapse
REFLEX uses a flexibility-calibrated stochastic bridge to balance conformational diversity and structural fidelity in RNA ensembles.
LLMs map DSA transparency taxonomy to platform terms
A new arXiv paper uses large language models to link the DSA Transparency Database's categories with the actual wording of platform terms and conditions.
CoRS: Fast neural scheduling via composite interval state reuse
A new neural scheduler for flexible job-shops avoids full graph refreshes after every decision, cutting computation while keeping feasibility exact.
ScopeMem: Latent Memory for Frozen GUI Agents
A new method trains prefix vectors to supply history to frozen vision-language models for GUI agents.
Beyond Static Leaderboards for Time Series Foundation Models
A new evaluation framework argues that static benchmarks for time series foundation models reward benchmark-specific tuning rather than genuine capability.
GraphEvoResearcher: Graph-Aware Self-Reflection for Cost-Bounded Multi-Hop QA
A new arXiv preprint proposes a graph-aware, training-free reasoning method that keeps multi-hop question answering cost-aware while building on self-reflective protocols.
AI Trading Agents' Rule-Following Put to the Test
A new study tests whether four AI trading agents adhere to their strategies and risk rules, offering a toolkit to separate luck from discipline.
Shared token budget: stopping and finalization policy audited on GSM8K
A fixed-budget policy combining first-valid-box stopping and a finalization cue is tested on 100 fresh GSM8K examples.
On-Board AI Turns Earth Observation Satellites into Data-Reduction Stages
A generic processing chain lets satellites run AI applications on board, downlinking compact products instead of raw data.
Remembering Has a Price: Memory Under Resource Constraints
A new arXiv paper examines how lifelong agents should choose what to remember when generation, interaction, and memory budgets are all tight.
Study examines rebalancing methods for imbalanced binary classification
A new arXiv preprint explores how local data augmentation can improve recall in the minority class without inflating false detections.
Action-Conditioned World Models Forecast EHR Patient Trajectories
A new preprint proposes a world model that learns patient histories as action-conditioned latent trajectories to improve forecasting from electronic health records.
VeriCUA: A Method for Proactive Verification in GUI Agent Rewards
A new approach learns proactive verification for reward agents, aiming to improve reliability in GUI agent reinforcement learning.
SAGE-CP: Marginal Coverage Masks Route-Specific Failures in Adaptive VLM Ensembles
A new study shows that a single conformal threshold in adaptive vision-language model ensembles can satisfy overall coverage while still failing badly on particular processing routes.
Hypergraph reasoning improves zero-shot tabular learning with LLMs
A new arXiv paper proposes hypergraph reasoning to overcome the limits of row-wise prompting in zero-shot tabular learning.