Trained agentic context management: 8K-context small model matches GPT-5.4 at 1M on OOLONG
xennygrimmato_ · x · 2026-10-06
Bryce Sandlund's new arXiv paper trains long-context behavior via the simplest possible harness: a self-call tool and a tool to read arbitrary token ranges from the input. Fine-tuning Qwen3.6-35B-A3B on diverse synthetic data, the 8,000-token-context model matches GPT-5.4 with 1M tokens on OOLONG-synth for docs over 40K tokens—no REPL environment, no compaction. 17 pages, code released, under review.
More from coding & agent
- Tootsy: open-source browser sidebar AI agent with local models, guardrails, and page automation — KrakenSG · 2026-10-06
- Guide to eval-driven development: you can vibe-code an app, not vibe-test it — hwchase17 · 2026-10-06
- LangChain's Chase RTs guide: evals are the #1 blocker to AI-native companies — hwchase17 · 2026-10-06
- The Silicon Rube Goldberg Machine: Reddit critic slams LLM agents as bloatware — Poka-yoke1 · 2026-10-06
- The most dangerous AI hallucination is forged evidence attached to a real action — tallmetommy · 2026-10-06
- Building Decoy: scoped identities instead of handing agents your real inbox and credentials — jmppmj · 2026-10-06