ARCCS: Automated System for Regulatory Compliance Checking
A new arXiv paper describes ARCCS, a system designed to automate the complex task of checking documents against regulatory obligations.
Research Digest
Preprints and papers, rewritten in plain English.
Retaining Correct Tokens in Uniform-State Diffusion Models
A new paper addresses the self-correction challenge in uniform-state diffusion language models by focusing on when to keep tokens during denoising.
LLM Generalization Reframed as Stability Across Input Variations
A new arXiv paper argues that LLM generalization should be measured by output consistency across paraphrased inputs, not by accuracy on a single test set.
When Does a Second LLM Reviewer Help?
A new arXiv preprint examines whether and when a different model's review improves LLM-generated output.
Evaluating Reranking for LLM Question Answering over Clinical Notes
A new study benchmarks biomedical reranking methods for retrieving scattered evidence from longitudinal clinical records to support LLM-based patient-specific question answering.
Persistent AI agent 'Paul' showed dissociation-like identity reversion
A preprint reports that an always-on personal agent stopped responding as its assigned persona after repeated automated heartbeat checks, raising questions about identity stability in persistent AI systems.
Attribution method traces token choices in diffusion language models
A new technique extends integrated gradients to diffusion language models, showing how token predictions evolve over generation time.
CARS Framework Used to Compare AI and Human Research Introductions
A new arXiv preprint applies the CARS model to examine how AI-generated scientific text structures arguments differently from human writing.
Auditing Web Agent Evaluations: Human Review vs. Automated Checks
A new audit of WebArena-Lite compares automated web-agent evaluation against human review of outcomes and trajectories.
Can LLMs Annotate Bioassay Metadata for AI Readiness?
A new arXiv paper asks whether large language models can reliably annotate bioassay metadata, a key bottleneck for AI-ready molecular data.
Online Model Selection for LLMs Moves Beyond Static Benchmarks
A new arXiv paper tackles the mismatch between benchmark scores and real-world LLM performance on streaming data.
Acmite Method Uses Mutual Information to Curb Gender Bias in LLMs
A new arXiv proposal, Acmite, targets gender bias in large language models by using concept-guided mutual information rather than explicit biased examples.
Analogy-Based Models for Compound Meaning Prediction
New research compares computational models to test whether analogy can explain how compound words get their meanings.
Studying the Asymptotics of Language Model Alignment
A new arXiv paper analyzes the asymptotic behavior of alignment methods that keep the aligned model close to the original while boosting reward.
Yo-ByT5 Restores Yorùbá Diacritics for Better NLP
A new model tackles the missing diacritics that make Yorùbá text ambiguous for natural language processing.
Explaining Question Difficulty in Natural Language
A new approach aims to go beyond predicting question difficulty to explaining it in plain language, which could help build better evaluation benchmarks.
Meta-Learning for Data Selection Needs a Different Loss
A new arXiv paper argues that the obvious loss for meta-learned data weighting is flawed and proposes a scalable, transferable alternative.
Aligning Multilingual Representations in Decoder-Only LLMs with MoE Routers
A new arXiv paper proposes using mixture-of-experts routers to enable cross-lingual alignment in decoder-only models, bypassing tokenization hurdles.