Auditing Financial LLM Agents by Separating Transition and Event Selection
A proposed audit framework moves beyond end-to-end return comparisons to isolate how financial LLM agents compose transitions and select events.
Research Digest
Preprints and papers, rewritten in plain English.
EvalResearchBench tests if AI agents can design their own evaluations
A new arXiv paper introduces EvalResearchBench, a benchmark for whether AI agents can create their own evaluation tasks to reduce the cost of recursive self-improvement.
EnvDreamer Aims to Bring Large-Scale Data Gains to Embodied AI
A new platform proposes generating training environments from multimodal data to accelerate embodied learning and robotics.
Dense Neuro-Symbolic Coupling for Geometry Reasoning
A new framework treats geometry problem-solving as a stateful loop that interleaves neural guesses with symbolic deduction.
Evaluating Whether LLM-Generated 3D Structures Are Actually Buildable
A new benchmark called LMBuild assesses whether LLM agents can create 3D structures that are not just visually elegant but also buildable and functional.
Study Flags Limits of Public Benchmarks for AI Cyber Capability
New research argues that public vulnerability benchmarks may mislead assessments of AI cyber capabilities, pointing to nonpublic vulnerabilities as a better test.
Flexible DRAM Interface Aims to Ease Adoption of New Memory Techniques
A proposed flexible DRAM interface and memory controller could help bring experimental DRAM features into real systems.
Mapping the Model Context Protocol Ecosystem's Hidden Structure
A new analysis reveals that the vast MCP server ecosystem lacks fine-grained organization, prompting a hierarchical map of its functional landscape.
Questioning the Necessity of Causality in Language Models
A new preprint argues that language models may not need causal reasoning when system behavior is treated as a Bayesian feature.
Mixture-of-Experts Trade-Offs in Particle-Physics Transformers
A new arXiv preprint examines whether mixture-of-experts models can scale particle-physics transformers efficiently.
Low-Budget AR-to-Diffusion Conversion Preserves Generation and Knowledge
A new method converts pretrained autoregressive LLMs into diffusion language models at low cost, keeping knowledge and enabling parallel generation.
New Diffusion Model Targets Discrete Ordinal Data
An arXiv preprint introduces a denoiser diffusion approach for discrete ordinal data, a middle ground between continuous and categorical diffusion.
AI Theorist Tool Spots Excitonic Structure in α-RuCl₃
A new AI-driven approach aims to close the gap between raw experimental data and microscopic understanding, demonstrated on the material α-RuCl₃.
Flow Matching Generates Brain Structural and Functional Connectivity
A new preprint proposes a multimodal hypergraph-based flow matching approach to generate structural and functional brain connectivity.
When to Graphify Multimodal Data: A No-Free-Lunch Approach
A new arXiv paper argues that converting multimodal data into graphs is not always beneficial and reframes the problem from how to build graphs to when to build them.
LLM agent costs overlooked by capability metrics
A new preprint argues that standard capability benchmarks ignore the computational cost of the harnesses that make LLM agents work.
Theoretical Guarantees for Time-Varying Bayesian Optimization
A new paper provides the first asymptotic performance bounds for a widely used method that optimizes noisy, expensive, and time-changing black-box functions.
Fisher-Guided Data Selection for Continual LLM Pre-Training
A new preprint proposes a Fisher-guided submodular method to keep continual pre-training from forgetting previously learned knowledge.