arXiv.org

6880 articles in the TruthFoundry index. Each links out to the original.

• Researchers propose lossy compressive text autoencoder architecture · New text autoencoder achieves competitive compression with strong reconstruction performance.
• Study finds LLM prompting unreliable for voice user simulator naturalness · Research audits LLM vs rule-based methods for speech naturalness in voice user simulators
• CogMem cognitive memory architecture proposed for long-term LLM dialogue agents · Researchers introduce CogMem graph-based memory system for long-term LLM dialogue reasoning.
• MARS-Gov bias detection framework accepted for EMNLP 2026 conference · Researchers present MARS-Gov, an open-set bias detection system for government documents.
• Research Tests LLM Long Deductive Reasoning Using Prolog Testbed · New research evaluates frontier LLMs' long-horizon deductive reasoning with a Prolog benchmark.
• Routed Sparse Autoencoders Separate Speech Linguistic Paralinguistic Data · Research uses routed sparse autoencoders to disentangle speech representation factors.
• Researchers Propose Speech-Rewarded Style Planning for Conversational TTS · New SRSP method improves controllable conversational text-to-speech style matching performance.
• RAG-E framework quantifies retriever-generator alignment for RAG systems · Researchers introduce RAG-E explainability framework and WARG metric for RAG systems
• Token-level early stopping improves diffusion language model efficiency · Researchers propose training-free early stopping for efficient diffusion language model generation
• STATIC constrained decoding enables production LLM generative retrieval on accelerators · Researchers present STATIC, an efficient constrained decoding technique for LLM generative retrieval.
• Research benchmarks transformer summarization metrics for UK financial reports · Study evaluates summarization metrics for long UK financial annual reports using transformers.
• Alice German benchmark released for rubric-based short answer scoring · Researchers introduce Alice, a new German benchmark for automatic short answer scoring.
• Research Paper Argues Linguistic Surprisal Theory Is An Unfalsifiable Tautology · Academic paper claims standard surprisal theory is tautological with no falsifiable predictions.
• Novel MoE Pretraining Methods Reduce GPU Communication Overhead · Researchers present expert coupling techniques cutting MoE training all-to-all overhead.
• Apollo Restore LLM introduced for Ancient Greek text gap restoration · Researchers present Apollo Restore, an LLM for restoring gaps in Ancient Greek texts.
• Researchers propose EASE framework to evade AI generated text detectors · Training-free EASE method reduces AI text detection performance without quality loss
• Emo-Jev probabilistic emotion classification framework evaluated against LLMs · Researchers introduce Emo-Jev, a Jev-based probabilistic emotion classification framework.
• Pseudo-dialect augmentation improves speech language model performance · Researchers present pseudo-dialect augmentation for low-data dialect SLM performance
• Research finds selective LLM retrieval diversification improves RAG performance · Study presents adaptive rule for retrieval diversification in LLM RAG systems.
• Researchers introduce CoTrace framework for training terminal AI agents · CoTrace improves terminal agent performance via harness-aware model training data recipes.
• New Study Explains Variation in Query Expansion Performance in Information Retrieval · Researchers explain query expansion performance using Ideal Expanded Query and document separability metrics.
• LLM4Impact Predicts Scientific Paper Impact Using Heterogeneous Evidence · New method LLM4Impact predicts scientific paper impact by integrating semantic, graph, and temporal data.
• Multi-vector visual document indices found vulnerable to inversion attacks · Research demonstrates stored visual document vector indices can reconstruct original pages.
• LLMs Fail to Preserve Cultural Diversity in Translated Math Problems · Study reveals LLMs collapse cultural diversity and misattribute regional context when translating math word problems.
• Researchers Propose Entity-Aware Metrics for Speech Privacy Evaluation · Researchers adapt NLP entity-aware metrics to evaluate privacy in speech obfuscation techniques.
• CHisAgent Framework Automates Ancient Chinese Historical Taxonomy Construction · CHisAgent, a multi-agent LLM framework, automates the construction of event taxonomies for ancient Chinese history.
• Study Finds Humans and AI Struggle to Agree on Social Media Sentiment · Research shows humans, LLMs, and bespoke tools exhibit only fair agreement when analyzing sentiment in 100 tweets.
• EncBank LLM caching system reduces GPU storage while retaining benchmark accuracy · Researchers present EncBank, a reusable LLM encoder caching system for repeated document queries.
• On-Policy Distillation Teaches Reasoning Skills, Not New Facts · Research shows on-policy distillation improves reasoning organization but fails to transfer new factual knowledge.
• FTA-Mem Framework Improves Long-Term Memory in Low-Density Emotional Support Dialogue · FTA-Mem framework enhances long-term dialogue memory using structured Fact-Time-Affect units.