Agent Step Value: Auditing Evaluation-Channel Sensitivity in Traces
A new metric quantifies whether evaluator-derived step rewards remain stable across different evaluation channels in black-box agent traces.
Research Digest
Preprints and papers, rewritten in plain English.
DAG Learning Method Targets Heavy-Tailed Node-Specific Noise
A new arXiv preprint extends DAG structure learning to settings where disturbances are heavy-tailed and vary across nodes, moving beyond typical Gaussian assumptions.
MORAD: A Reinforcement Learning Approach to RNA Inverse Design
A new framework trains a single shared policy across diverse RNA backbones, targeting structure-conditioned inverse design.
Simulating A/B Tests with Data-Driven Persona-Conditioned Agents
A new framework aims to predict A/B test outcomes without weeks of live traffic by simulating user personas.
MIRROR: Training VLMs to Reason Across Text and Image Views
A new method aims to reduce the inconsistency in vision-language model reasoning by learning from complementary textual and visual views.
GenCoord: Executable Commitments for Coordinating Embodied LLM Agents
A framework that uses executable commitments to help embodied language-model agents coordinate interdependent actions while keeping private information.
Vision-Language Reward Model Built Without Human Labels
A new approach builds vision-language reward models without human preference labels, sidestepping costly and inconsistent annotations.
Sampling-based kinodynamic planning may fix long-horizon world-model tasks
A new arXiv paper argues that sampling-based kinodynamic planning can overcome the long-horizon failures of optimizer-based planners like CEM and MPPI.
MHR-Router: Heterogeneous Retrieval-Enhanced Routing for VLMs
A new router uses retrieval to adaptively select the best vision-language model for each image-question query.
Stochastic Knowledge Graphs for Faithful LLM Student Simulation
A new arXiv paper introduces Stochastic Student Knowledge Graphs to improve the fidelity of LLM-based student simulations for training and testing tutoring systems.
One-Step Implicit Sampling for Faster Bayesian Inference
A new arXiv paper proposes an amortised inference method that aims to draw unbiased posterior samples in a single step.
Automated EEG biomarker for neonatal encephalopathy matches expert detection
A new preprint proposes automated detection of low-voltage intervals in neonatal EEG as a robust, expert-level biomarker for encephalopathy.
Survey Maps Fragmented Evaluation of Automated Research Systems
A new review catalogs the scattered benchmarks and evaluation practices used to assess AI systems that automate research tasks.
PulseBound: Stopping PPG Models from Peeking at Future Samples
A new arXiv preprint argues that PPG predictive models can leak future information through normalization and nonlocal transforms, and proposes PulseBound to enforce an explicit information boundary.
Generalized Laplace Active Subspaces for Scalable AI Uncertainty
A new preprint proposes a tractable approximation to full Bayesian inference, aiming to make uncertainty estimates practical for high-stakes neural network deployment.
Interpretable PDE Discovery via Structural Experience Distillation
A new arXiv preprint proposes an interpretable method for discovering symbolic PDE solutions by distilling structural experience.
EvoAlloc: Self-Evolving Agent Allocates Compute for Program Evolution
A new arXiv paper proposes an agent that learns to allocate evaluation resources during LLM-based program evolution.
Context-Sensitive Memorization: Probing LLM Data Leakage Beyond Prefixes
A new arXiv study asks whether LLM memorization and data leakage depend on context, not just isolated prefixes.