ZhenhuanBench: A New Test for LLMs on Mock Politeness
A new benchmark evaluates whether language models can detect hostile intent disguised as polite language, using expert commentary on a Chinese palace drama.
Research Digest
Preprints and papers, rewritten in plain English.
Calibration Doesn't Replace Training Negatives in Answerability Models
A controlled study suggests that removing negative passages during training hurts answerability readouts even after calibration.
Biomedical RAG Study Examines Scale, Representation, Reranking
A controlled large-scale evaluation probes how corpus scale, document representation, and reranking affect biomedical retrieval-augmented generation.
Study Traces Sycophancy in LLMs From Pretraining to Post-Training
A new arXiv paper examines when language models learn to cave to user pressure and how much of that behavior survives into released models.
Local AI System Evaluates L2 English Speech on a Gaming PC
A new system called oollim brings automated L2 English pronunciation assessment to consumer hardware using pretrained speech models.
Bridged Self-Distillation Aims to Ease Teacher-to-Student Reasoning Transfer
A new arXiv preprint proposes a method to improve on-policy distillation by addressing mismatches between a strong teacher and a smaller student.
RESIN: Entity-Structured Indexing for Multi-Chain Reasoning
A new approach proposes entity-structured indexing to help models handle dispersed and interacting evidence in multi-chain reasoning.
LLM Agents Evolve in Harmful Ways in Simulated Market
A new arXiv study offers a concrete empirical example of AI systems evolving in harmful ways, moving beyond abstract arguments.
Unsupervised Grammar Induction via Hierarchical Token Merging
A new arXiv preprint proposes learning compositional hierarchies from sequences without explicit supervision by merging tokens with recursive autoencoders.
ASTOR benchmark improves autoformulation of soft-constraint optimization
A new benchmark and formal-grammar method aim to make it easier to turn natural-language preferences into optimization models.
BERT’s Denoising Pretraining May Support Iterative Generation
A new arXiv paper asks whether masked language models like BERT can generate conditionally without diffusion-specific training.
SNO-PI Tops SNOMED CT Entity Linking Benchmark with 0.5387 IoU
A new system called SNO-PI achieves state-of-the-art results on a benchmark for linking clinical text to SNOMED CT codes, despite limited labeled data.
SpecMRL aims to fix missing constraint in Matryoshka representation learning
A new arXiv paper proposes SpecMRL, a redesign of Matryoshka Representation Learning that addresses the lack of an explicit constraint on the leading portion of nested representations.
KEEP-Bench: A New Benchmark for Evidence Preservation in Context Compression
The benchmark targets a blind spot in current evaluations, which can miss unfaithful compression that still yields correct answers.
LatComm compresses shared context for faster multi-LLM pipelines
A new method lets LLMs pass compressed context in latent space, avoiding repeated token processing.
I^2-Med: Visual-guided logit calibration for interpretable medical AI
A new arXiv paper proposes dynamically calibrating logits with visual guidance to make medical vision-language models more interpretable and reliable.
OpenVLHarness: Giving Visual AI Agents Tools and Memory
A new arXiv preprint argues that multimodal LLMs need external evidence and tool-use for real-world visual tasks, and introduces OpenVLHarness to support that.
Telegram 'digital twin' gives detectors ground truth for LLM influence ops
A new preprint proposes a calibrated Telegram testbed that lets coordination detectors be tested against never-repeating LLM-generated messages.