verievol-verifiable-data-construction

name	verievol-verifiable-data-construction
description	Verifiable Evol-Instruct framework for scaling multimodal mathematical reasoning. Type-aware evolution + HTV-Agent verifier with offline hypothesis-test falsification ensures reliable reward labels at scale.
trigger_words	["verifiable data","evol-instruct","hypothesis testing","multimodal reasoning","data verification"]
version	1
arxiv	2606.23543v1
authors	Haoling Li, Kai Zheng, Jie Wu, et al.
date	2026-06-22T00:00:00.000Z
categories	cs.AI, cs.CL, cs.CV, cs.LG

VeriEvol: Verifiable Evol-Instruct for Mathematical Reasoning

Core Insight: Scaling RL requires verifiable data construction. Decouple prompt difficulty (evolution) from answer reliability (falsification) before policy updates.

Problem Addressed

• Reward label reliability decay as data volume grows
• Supervision scaling trusts labeller without verification
• Policy-side methods assume answers already correct

Key Methodology

Two-Axis Decoupling

1. Prompt Difficulty Axis:

   • Type-aware evolution operators
   • Rewrite low-difficulty seeds into harder, image-grounded prompts
   • Route-specific evolution strategies

2. Answer Reliability Axis:

   • HTV-Agent verifier: Offline hypothesis-test falsification
   • Accept answer only after multi-source counter-evidence fails to refute
   • Never trust labeller blindly

HTV-Agent Verification Process

def htv_agent_verify(question, proposed_answer, image):
    # Generate counter-evidence from multiple sources
    counter_evidence = []
    
    # Source 1: Mathematical consistency check
    consistency = check_mathematical_consistency(proposed_answer)
    counter_evidence.append(consistency)
    
    # Source 2: Visual grounding verification
    visual_grounding = verify_image_grounding(proposed_answer, image)
    counter_evidence.append(visual_grounding)
    
    # Source 3: Alternative solution derivation
    alternative = derive_alternative_solution(question)
    counter_evidence.append(alternative)
    
    # Hypothesis-test falsification
    if all_falsification_attempts_failed(counter_evidence):
        return ACCEPT_VERIFIED
    else:
        return REJECT_UNVERIFIED

Performance Gains

• Scaling 10K → 250K samples: 35.42 → 54.73 mean accuracy
• Evolved prompts: +1.82 gain
• HTV-Agent verifier: +2.06 gain