Skill Files Bloat as Agents Self-Evolve: Inside SkillZip's Fix

Hong-Wei Wu published on 2026-10-21 included in category Paper Introduction

Self-evolving agents only ever append to their skill files. SkillZip merges the repetition with no task-based validation, shrinking them 31.2% on average, with caveats.

Do Agents Need a Map to Search Documents? Inside CorpusMap

Hong-Wei Wu published on 2026-10-17 included in category Paper Introduction

CorpusMap precomputes an entity-document map so search agents follow links, not rediscover them: +6.4 to 11.7 Overall Quality points and 34% to 57% fewer input tokens.

Can You Trust a PDF Parsing Leaderboard? Inside ParseBench

Hong-Wei Wu published on 2026-10-15 included in category Paper Introduction

ParseBench scores PDF parsing by whether an AI agent can act on the output, not by text similarity. Its table metric is worth borrowing; the leaderboard needs a discount.

Can a Search Index Repair Itself? A Skeptical Look at SELF-INDEX

Hong-Wei Wu published on 2026-10-10 included in category Paper Introduction

SELF-INDEX lets an LLM find weak search-index keys, revise only those, and apply a change only after validation. Strong engineering, but the judge is also the generator.

Can an LLM Write Its Own Playbook From a Manual? Ctx2Skill

Hong-Wei Wu published on 2026-10-07 included in category Paper Introduction

Ctx2Skill pits a Challenger against a Reasoner to distill a long text's hidden rules into a reusable skill file, no labels needed: GPT-4.1 rises from 11.1% to 16.5%.

Cheap LLM Judge First, GPT-6 Only When Unsure: Does JEV Save Money?

Hong-Wei Wu published on 2026-10-06 included in category Paper Introduction

A third-party test of JEV, a cheap LLM judge that escalates unsure cases to GPT-6: 27% of the cost on easy tasks, only 9-33% saved on hard ones. AUROC explained.

Do Self-Improving Agents Just Memorize the Test? Inside Google's RRSI

Hong-Wei Wu published on 2026-10-04 included in category Paper Introduction

An agent that rewrites its own harness and is scored on the same tasks looks better every round, yet the gains may not transfer. Google's RRSI adds guardrails to curb it.