Vector-Space Concept Representations for Interdisciplinary Meaning Study
A new arXiv preprint argues that concept representations in vector space can help multiple disciplines study how meaning varies.
Research Digest
Preprints and papers, rewritten in plain English.
Hash-Addressed Memory Aims for Local Post-Hoc Knowledge Injection
A new arXiv paper proposes a deterministic hash-addressed memory mechanism for targeted knowledge updates after training, building on an earlier audit of generality claims.
AI Tutor Acharya 2.0 Targets Hindi Self-Learning
A new arXiv paper describes an AI tutor designed to support self-learning in Hindi, a low-resource language.
Hidden Cartographer: Selecting Tuning Data with Hidden States
A new arXiv paper proposes using hidden states to guide data selection for instruction tuning under strict annotation and compute budgets.
Search-Augmented LLMs Fail on Rapidly Evolving Knowledge
A new arXiv study examines why search-augmented LLMs struggle when new and old information conflict online.
Code-Switched Tasks Bring RLVR to Non-English Domains
A new arXiv paper proposes synthesizing code-switched tasks to apply reinforcement learning with verifiable rewards to specialized non-English domains.
SPARC Corpus Tests Speaker Leakage in Transcribed Speech
A new benchmark, SPARC, shifts focus to how much speaker identity and demographics remain in transcribed words, not just the acoustic signal.
RARE: Adaptive Rubric Evolution for Open-Ended RL
A new reinforcement learning method targets open-ended tasks that lack automatic verifiers and require satisfying multiple joint requirements.
Study Evaluates LLMs on Figurative Moroccan Darija
A new arXiv paper examines how well large language models understand idioms and metaphors in Moroccan Darija, a dialect with limited NLP resources.
New study asks whether LLMs reproduce age gaps in loneliness surveys
A new arXiv paper argues that matching survey averages is not enough to validate LLM survey simulations, and tests whether they capture within-country age differences in loneliness using European data.
Mixture-of-Memories: A Modular Approach to Knowledge Injection in LLMs
A new paper proposes a low-overhead method for injecting modular knowledge into language models, addressing the difficulty of controlling and updating stored facts.
Lakota Archival OCR: Design for Language Reclamation
A new arXiv paper outlines design considerations for low-resource OCR in Lakota archives, aiming to support language reclamation and community revitalization.
DiaLong benchmark probes how well LLMs recall facts in long dialogues
A new benchmark, DiaLong, aims to measure how well transformer-based LLMs retain and recall factual information across long dialogue contexts.
On-Device Call Screening Gets a Quality-Controlled Training Dataset
A new dataset aims to help small language models screen calls locally, avoiding the need to send call content to the cloud.
Gold-Conditioned Distillation Aims to Soften Fine-Tuning Targets
A proposed self-distillation method, Gold-Conditioned Distillation, targets the one-hot target problem in supervised fine-tuning.
Do LLMs Build Cognitive Maps? A New Test
A new arXiv study probes whether large language models, like animals, construct cognitive maps for flexible reasoning.
Reusable Ledgers Could Cut Cost of Auditing LLM Agents
A proposed 'write once, check many' approach could let LLM agent trajectories be audited against future policies without rereading them.
MapSteal Attack Autonomously Extracts Knowledge from Black-Box LLMs
A new arXiv paper proposes MapSteal, a map-based method that autonomously steals knowledge from black-box LLMs, exposing a security gap in model distillation.