WikiSkill: A Skill Memory That Never Gets Rolled Back
WikiSkill adds a wiki layer that never gets rolled back between raw agent traces and skill files, so each skill edit builds on accumulated cross-iteration evidence.
LLM-as-a-Verifier: A Better Alternative to LLM-as-a-Judge
LLM-as-a-Verifier reads the full score-token probability distribution instead of argmax, eliminating ties and beating trained reward models with zero training.
RAGU: The Paper That Accidentally Measured GraphRAG's Ceiling
A close reading of RAGU's GraphRAG engine -- two-stage extraction, DBSCAN consolidation, Leiden clustering -- and the number in its own tables: the graph is worth 1.2 to 4.3 pp.
SAG: Multi-Hop RAG via Query-Time SQL, Not Offline Graphs
SAG skips the offline knowledge graph, storing events and entities in SQL, joined at query time -- beating HippoRAG 2 on multi-hop benchmarks like MuSiQue.
SketchVLM: Letting a VLM Draw Its Own Reasoning
A walkthrough of SketchVLM: with no retraining, a coordinate grid, XML commands and Bezier smoothing let a VLM annotate the image so its answers become verifiable.
Persona-Aware D2S: One Paper, Four Audience-Tailored Slide Decks
Persona-Aware D2S (EACL 2024) generates four slide-deck versions of one paper by audience and length, via an RLHF-lite pipeline we unpack and critique.
Claude Code Concepts: Agentic Loop, Tools, Skills
A concept-level tour of Claude Code's architecture: the read-think-act agentic loop, its tool system, CLAUDE.md project memory, hooks, skills, MCP, and permission model.