Paper library

These are some of the notes I made with the help of Claude. They are designed to be small and digestible, something you might go over on a commute, on a short walk, or whenever you don't want to spend 3 hours on a paper. 7 papers. Review questions from the papers you have read.

DerivAuditMemory Is a Derivation: The Distributed-Evidence Paradox in Long-Term Agents When an agent writes a memory, does the history it saw before writing support everything the memory says? And does a check at write time catch the memories that it does not? about 11.5 minHongjun Liu, Chen Zhao, 2026 memoryagentsevaluationverification IRT for AI evaluationCan We Trust Item Response Theory for AI Evaluation? With Three Papers to Read First Benchmarks now borrow item response theory from school testing to rank models, flag bad questions and build short tests. When there are few models and many questions, are its numbers right? A main paper and three papers to read first. about 20 minJiang, Kwon, Luo, Xiao and Zhang (main paper); Rodriguez et al. (Leaderboards); Maia Polo et al. (tinyBenchmarks); Jiang et al. (OpenEval), 2021 to 2026 evaluationpsychometricsbenchmarks GRPO, SFT and LoRAGRPO, SFT and LoRA: How a Model Learns From Scores How does a model learn from scores instead of example answers, and how can that training be made cheaper? Three papers, with story writing as the example. about 10 minShao et al. (DeepSeekMath); Hu et al. (LoRA); Liu, Bahja, Kovatchev and Lee (story style), 2021 to 2026 trainingfine-tuningwriting Biology of a language modelOn the Biology of a Large Language Model When a language model answers, can we see the steps it takes inside, and check that those steps are real? about 10 minJack Lindsey, Wes Gurnee, Emmanuel Ameisen, Brian Chen, Adam Pearce, Nicholas L. Turner, Craig Citro, and 20 more authors at Anthropic, listed in the paper; Joshua Batson is the corresponding author, 2025 interpretabilitymechanismssafety Lost in Context CompactionLost in Context Compaction: When LLM Systems Drop Side Constraints When a long session is replaced by a summary, does the summary keep the rules the user set, and does the model still follow them? about 9.5 minZhiqi Wang, Yichi Zhang, Dongwon Lee, Nathalie Baracaldo, Yuchen Yang, 2026 agentsmemoryevaluationcompaction Lost in TransliterationLost in Transliteration: Bridging the Script Gap in Neural IR When people type a search in Latin letters instead of their language's own script, do search models built on neural networks still find the right documents, and can training fix it? about 9.5 minAndreas Chari, Iadh Ounis, Sean MacAvaney, 2025 retrievalmultilingualevaluation LAPSELanguage Model Memory Flattens the Temporal Form of User Facts When a model saves a note about a user, does it keep the verb that meant “for now”? And does the model that reads the note later use it? about 10.5 minSugam Panthi, Muhaiminul Yeamin, Siyan Luo, Rabab Abdelfattah, 2026 memorylinguisticsevaluationagents

No paper matches.