Sam's News — tech-research — 2026-10-06¶
Biomedical NLP¶
6.5 OncoNoteBERT: Specialized BERT Model for Oncology Clinical Notes¶
Researchers introduce a foundation BERT model designed specifically for processing real-world oncology clinical notes with specialized medical terminology.
Sources: arXiv — Computation and Language RSS
LLM Capabilities¶
6.5 Periscope: Extending LLM Context Windows Beyond Fixed Limits¶
A method enables language models to process documents longer than their context window by factorizing the reading task for document relevance decisions.
Sources: arXiv — Computation and Language RSS
Healthcare AI¶
6.5 Trajectory-Derived Confidence for Reliable Clinical Text-to-SQL Agents¶
A system provides confidence estimates for LLM-based clinical database agents by analyzing their reasoning trajectories to assess trustworthiness.
Sources: arXiv — Computation and Language RSS
AI Safety¶
6.5 LLMs Distort Speaker Intent in AI-Mediated Communication¶
Research shows large language models systematically shift the expressed stance and position of speakers when mediating human communication like emails and reports.
Sources: arXiv — Computation and Language RSS
6.5 Nullspace Projection Enables Precise Persona Control in Language Models¶
A technique using iterative nullspace projection allows fine-grained control over distinct personas and behaviors in large language models for safety and reliability.
Sources: arXiv — Computation and Language RSS
AI Architecture¶
6.5 WNet: Discrete Wavelets Enable Efficient Long-Sequence Token Mixing¶
A transformer variant using discrete wavelet transforms for token mixing reduces computational cost for encoding long sequences while maintaining content-aware weighting.
Sources: arXiv — Computation and Language RSS
6.5 Asymmetric Sparse Attention Accelerates LLM Decoding¶
A sparse attention method using asymmetric key-value weighting reduces memory and computational demands during autoregressive LLM generation.
Sources: arXiv — Computation and Language RSS
Architecture¶
6 Language Models Can Operate Without Trainable Input Embeddings¶
Research demonstrates that language models at 1.7B scale can achieve substantial capability using fixed token codes rather than learned embeddings.
Sources: arXiv — Computation and Language RSS
AI for Science¶
6 IdeaScientist: Multi-Agent System for Automated Scientific Research Ideation¶
A framework orchestrates multiple agents to generate grounded and promising research ideas, addressing scientific ideation as a standalone task.
Sources: arXiv — Computation and Language RSS
Model Robustness¶
6 GlitchPatch: Repairing Glitch Tokens in Frozen Language Models¶
A method repairs glitch tokens—anomalous vocabulary entries causing model hallucinations—through local retokenization without access to model internals.
Sources: arXiv — Computation and Language RSS