ECPO: Evidence-Coupled Policy Optimization for Evidence-Certified Candidate Ranking

2026-05-23 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

Evidence-Coupled Policy Optimization (ECPO) is a novel approach for evidence-certified candidate ranking, designed for decision-support systems that require verifiable evidence alongside ranked candidates. This system, instantiated and evaluated on MAVEN-ERE and RAMS datasets, outputs a Top-K list with doc_id:span evidence certificates, ensuring cited spans are sufficient to recover the decision. ECPO employs a listwise policy-optimization objective that jointly optimizes both ranking and evidence certificate generation. It learns an interpretable trajectory reward based on skeleton alignment, argument consistency, and optional graph features. The policy is optimized using three coupled rewards: listwise ranking utility, span-level certificate validity, and an evidence-cycle reward derived from a label-free deterministic verifier. This method redefines the objective from maximizing standard NDCG to maximizing CertNDCG and decision-evidence coupling, demonstrating its effectiveness against multiple baselines including zero-shot, SFT, and GRPO policies across closed-roster, predicted-roster, and hybrid-roster settings.

Key takeaway

For AI Engineers building decision-support systems requiring transparency, ECPO offers a robust solution. If your current ranking models lack verifiable evidence, you should consider integrating ECPO's joint optimization of ranking and evidence generation. This approach ensures decisions are not only accurate but also auditable via doc_id:span certificates, enhancing trust and compliance. Evaluate ECPO against your existing methods to improve CertNDCG and decision-evidence coupling.

Key insights

The ECPO framework jointly optimizes candidate ranking and verifiable evidence generation using coupled rewards and a deterministic verifier.

Principles

• Decision-support ranking needs verifiable evidence.
• Joint optimization improves evidence coupling.
• Verifier-based rewards enhance certificate validity.

Method

ECPO uses a listwise policy optimization with three coupled rewards: ranking utility, span validity, and an evidence-cycle reward from a label-free verifier, shifting from NDCG to CertNDCG.

In practice

• Apply ECPO for auditable decision-support systems.
• Use doc_id:span certificates for decision recovery.
• Integrate label-free verifiers for evidence quality.

Topics

• Evidence-Certified Ranking
• Policy Optimization
• Decision Support Systems
• MAVEN-ERE
• RAMS
• CertNDCG

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• ECPO: Evidence-Coupled Policy Optimization for Evidence-Certified Candidate Ranking — Ranking systems in decision-support must provide verifiable evidence alongside candidate lists.
• CATPO: Critique-Augmented Tree Policy Optimization — CATPO optimizes LLM reasoning by focusing RL training on informative trees and healing failures with critiques.
• Learning When Not to Act: Mitigating Tool Abuse in Agentic Reinforcement Learning — EAPO enables agentic reinforcement learning models to selectively use tools, avoiding overuse without sacrificing performance.
• Structured Role-Aware Policy Optimization for Multimodal Reasoning — SRPO improves LVLM multimodal reasoning by assigning role-aware, token-level credit for visual evidence and reasoning.
• Towards Pareto-Optimal Tool-Integrated Agents with Pareto Ranking Policy Optimization — ParetoPO optimizes tool-integrated LLMs for superior accuracy-efficiency trade-offs using a two-stage multi-objective framework.
• Uniform-Correct Policy Optimization: Breaking RLVR's Indifference to Diversity — RLVR's diversity collapse stems from objective indifference and on-policy sampling, which UCPO addresses by promoting uniform probability across correct solutions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.