[논문 리뷰] Neuro-Symbolic Verification on Instruction Following of LLMs

Yiming Su, Kunzhao Xu외 5명|arXiv (Cornell University)|2026. 01. 25.

Topic Modeling인용 수 0

한 줄 요약

Nsvif는 프롬프트 준수를 제약-충족 문제로 형식화하여 LLM 출력이 프롬프트를 따르는지 확인하는 보편적 신경-기호 검증기(verifier)를 제시합니다. 이는 논리/의미 제약을 하나의 해석기(unified solver)와 결합하고 평가를 위해 VifBench를 도입합니다.

ABSTRACT

A fundamental problem of applying Large Language Models (LLMs) to important applications is that LLMs do not always follow instructions, and violations are often hard to observe or check. In LLM-based agentic workflows, such violations can propagate and amplify along reasoning chains, causing task failures and system incidents. This paper presents NSVIF, a neuro-symbolic framework for verifying whether an LLM's output follows the instructions used to prompt the LLM. NSVIF is a universal, general-purpose verifier; it makes no assumption about the instruction or the LLM. NSVIF formulates instruction-following verification as a constraint-satisfaction problem by modeling user instructions as constraints. NSVIF models both logical and semantic constraints; constraint solving is done by a unified solver that orchestrates logical reasoning and semantic analysis. To evaluate NSVIF, we develop VIFBENCH, a new benchmark for instruction-following verifiers with fine-grained data labels. Experiments show that NSVIF significantly outperforms LLM-based approaches and provides interpretable feedback. We also show that feedback from NSVIF helps improve LLMs' instruction-following capability without post-training.

연구 동기 및 목표

• LLM의_instruction_를_적절하게_따르도록 하는 신뢰할 수 있는_verification의 필요성을 제시하여 에이전트 워크플로우에서의 위반 확산을 방지합니다.
• 프롬프트를 제약으로 모델링하고 논리 및 의미 분석을 통해 출력을 확인하는 보편적 검증기를 제안합니다.
• 어느_제약이_위반되었고 출력이 어떻게 그 제약을 실패시키는지 식별하여 해석 가능한 피드백을 제공합니다.
• 사후 학습 없이 LLM 기반 판사보다 효과적임을 입증하고 LLM의 프롬트 준수 개선 가능성을 보여줍니다.]
• method([
• 제약과 그것의 관계를 일차 도퀘 논리 CSP 인코딩으로 제시합니다.
• 제약을 논리적(기호적) 제약과 의미적(신경 기반) 제약으로 분해하고 통합 SMT 해석기(Z3)를 사용하여 타당성을 검사합니다.
• 다중 에이전트 시스템을 사용합니다: 형식화 에이전트(제약 추출 및 Z3 프로그램 구성), 검사 에이전트(제약 검사기 생성), 해석기 에이전트(S CSP를 실행합니다).
• 제약 추출, 제약 유형 분류, 의미 평가에 LLM을 활용하고, 검사기 코드를 보완하기 위한 자기성찰 루프를 도입합니다.
• 제약 수준 주석이 포함된 큐레이션된 데이터 세트인 VifBench를 도입하여 검증기를 평가합니다.
• 사후 학습 없이 LLM의 프롬프트 준수 향상을 위한 피드백 메커니즘을 제공합니다.

제안 방법

• Formulate instruction following as a first-order logic CSP encoding constraints and their relations.
• Decompose constraints into logic (symbolic) and semantic (neural) constraints and use a unified SMT solver (Z3) to check feasibility.
• Use a multi-agent system: formulation agent (extracts constraints and builds a Z3 program), checking agent (generates constraint checkers), and solver agent (executes the CSP).
• Leverage LLMs for constraint extraction, constraint type classification, and semantic evaluation, with self-reflection loops to repair checker code.
• Introduce VifBench as a curated dataset with fine-grained constraint-level annotations to evaluate verifiers.
• Provide feedback mechanisms to improve LLM instruction following without post-training.

실험 결과

연구 질문

• RQ1 Can a universal neuro-symbolic verifier accurately determine whether an LLM output follows its prompting instruction across diverse instructions and LLMs?
• RQ2 How does decomposing instructions into logic and semantic constraints impact verification accuracy and interpretability?
• RQ3 Does constraint-level feedback from Nsvif translate into improved LLM instruction following without post-training?
• RQ4 How does Nsvif compare to LLM-based judges in terms of accuracy, precision/recall, and robustness across models?
• RQ5 What is the effect of instruction complexity on verification performance?

주요 결과

• Nsvif significantly outperforms LLM-as-a-judge baselines across multiple models in F1/precision/recall metrics.
• The CSP-based, neuro-symbolic approach yields interpretable results by pinpointing violated constraints and the reasoning path.
• A multi-turn feedback loop with Nsvif can guide LLMs to improve instruction following without post-training.
• A dedicated benchmark (VifBench) with fine-grained constraint labels exposes the limitations of single-label evaluations and enables deeper analysis.
• Ablation studies show that symbolic reasoning (SMT solver) substantially boosts verification performance over purely semantic checks, while still benefiting from semantic analysis.