research: VR-CLI long-form reasoning technique (ReasoningNCP) — Harrier/Rufous document generation #49

New issue

Open

opened 2026-06-15 12:16:57 -07:00 by pyr0ball · 0 comments

pyr0ball commented

2026-06-15 12:16:57 -07:00

Owner

Source: https://github.com/Alex-Gurung/ReasoningNCP (fork: TheAhmadOsman/ReasoningNCP)
License: MIT
Paper: arXiv 2503.22828 — Gurung & Lapata (2025)

What it is

ReasoningNCP implements VR-CLI (Verifiable Rewards via Completion Likelihood Improvement) — an RL training framework that teaches models to produce useful plans for long-form continuation tasks. The core task is Next-Chapter Prediction (NCP): given story context, plan then execute the next chapter. The reward signal measures whether the plan actually improved completion likelihood — not whether the plan looks good.

Why relevant for CF

Benefit appeal letters (Harrier/Rufous) require:

Multi-section structure (intro, statement of facts, legal basis, ask)
Coherence across a long document
Reasoning-first generation (plan the argument, then write it)

VR-CLI's reward signal is directly applicable: replace story completion quality with letter completion quality. A model trained this way would learn to produce plans that actually improve the final letter — not just generic outlines.

Application path

Not a drop-in library — this is a training recipe. The path:

Adapt VR-CLI reward to structured document tasks
Fine-tune a base model (Qwen or Llama) on CF's document templates
Deploy in Harrier/Rufous as the long-form draft backend

Watch trigger

Research inspiration for now. Re-evaluate when Harrier or Rufous reaches active development and a long-form generation baseline is needed.

**Source:** https://github.com/Alex-Gurung/ReasoningNCP (fork: TheAhmadOsman/ReasoningNCP) **License:** MIT **Paper:** arXiv 2503.22828 — Gurung & Lapata (2025) ## What it is ReasoningNCP implements VR-CLI (Verifiable Rewards via Completion Likelihood Improvement) — an RL training framework that teaches models to produce useful plans for long-form continuation tasks. The core task is Next-Chapter Prediction (NCP): given story context, plan then execute the next chapter. The reward signal measures whether the plan actually improved completion likelihood — not whether the plan looks good. ## Why relevant for CF Benefit appeal letters (Harrier/Rufous) require: - Multi-section structure (intro, statement of facts, legal basis, ask) - Coherence across a long document - Reasoning-first generation (plan the argument, then write it) VR-CLI's reward signal is directly applicable: replace story completion quality with letter completion quality. A model trained this way would learn to produce plans that actually improve the final letter — not just generic outlines. ## Application path Not a drop-in library — this is a training recipe. The path: 1. Adapt VR-CLI reward to structured document tasks 2. Fine-tune a base model (Qwen or Llama) on CF's document templates 3. Deploy in Harrier/Rufous as the long-form draft backend ## Watch trigger Research inspiration for now. Re-evaluate when Harrier or Rufous reaches active development and a long-form generation baseline is needed.