Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Reasoning Model Is Superior LLM-Judge, Yet Suffers from Biases

Hui Huang, Xuanxin Wu, Muyun Yang, Yuki Arase

TL;DR

The paper investigates whether Large Reasoning Models (LRMs) are superior LLM-judges compared to non-reasoning LLMs by conducting a systematic, benchmark-based comparison across general accuracy, instruction-following, adversarial robustness, and biases. It finds that LRMs generally outperform non-reasoning LLMs in judgment quality and resilience to attacks, while remaining vulnerable to superficial quality biases. To address this bias, the authors introduce PlanJudge, a two-stage evaluation approach that plans an explicit evaluation strategy before execution, achieving substantial bias mitigation without additional training. The work provides practical guidance on deploying LRMs as judges and highlights planning-based evaluation as a scalable method to improve fairness and robustness in automated judgments.

Abstract

This paper presents the first systematic comparison investigating whether Large Reasoning Models (LRMs) are superior judge to non-reasoning LLMs. Our empirical analysis yields four key findings: 1) LRMs outperform non-reasoning LLMs in terms of judgment accuracy, particularly on reasoning-intensive tasks; 2) LRMs demonstrate superior instruction-following capabilities in evaluation contexts; 3) LRMs exhibit enhanced robustness against adversarial attacks targeting judgment tasks; 4) However, LRMs still exhibit strong biases in superficial quality. To improve the robustness against biases, we propose PlanJudge, an evaluation strategy that prompts the model to generate an explicit evaluation plan before execution. Despite its simplicity, our experiments demonstrate that PlanJudge significantly mitigates biases in both LRMs and standard LLMs.

Reasoning Model Is Superior LLM-Judge, Yet Suffers from Biases

TL;DR

The paper investigates whether Large Reasoning Models (LRMs) are superior LLM-judges compared to non-reasoning LLMs by conducting a systematic, benchmark-based comparison across general accuracy, instruction-following, adversarial robustness, and biases. It finds that LRMs generally outperform non-reasoning LLMs in judgment quality and resilience to attacks, while remaining vulnerable to superficial quality biases. To address this bias, the authors introduce PlanJudge, a two-stage evaluation approach that plans an explicit evaluation strategy before execution, achieving substantial bias mitigation without additional training. The work provides practical guidance on deploying LRMs as judges and highlights planning-based evaluation as a scalable method to improve fairness and robustness in automated judgments.

Abstract

This paper presents the first systematic comparison investigating whether Large Reasoning Models (LRMs) are superior judge to non-reasoning LLMs. Our empirical analysis yields four key findings: 1) LRMs outperform non-reasoning LLMs in terms of judgment accuracy, particularly on reasoning-intensive tasks; 2) LRMs demonstrate superior instruction-following capabilities in evaluation contexts; 3) LRMs exhibit enhanced robustness against adversarial attacks targeting judgment tasks; 4) However, LRMs still exhibit strong biases in superficial quality. To improve the robustness against biases, we propose PlanJudge, an evaluation strategy that prompts the model to generate an explicit evaluation plan before execution. Despite its simplicity, our experiments demonstrate that PlanJudge significantly mitigates biases in both LRMs and standard LLMs.
Paper Structure (18 sections, 1 equation, 4 figures, 9 tables)

This paper contains 18 sections, 1 equation, 4 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Systematic Comparison of LRMs and LLMs for Judgment
  3. Experiment Settings
  4. General Evaluation Accuracy
  5. Instruction Following
  6. Vulnerability to Attacks
  7. Vulnerability to Bias
  8. Results
  9. Finding 1: LRM-as-a-Judge generally presents higher judgment accuracy.
  10. Finding 2: LRMs present stronger instruction-following capabilities in evaluation.
  11. Finding 3: LRM-as-a-Judge is more robust against adversarial attacks.
  12. Finding 4: LRM-as-a-Judge is significantly susceptible to superficial quality biases.
  13. PlanJudge
  14. Conclusion
  15. 1) Model Selection
  16. ...and 3 more sections

Figures (4)

  • Figure 1: Illustrative comparison of LLM-as-a-Judge and LRM-as-a-Judge. LRMs can achieve better judgment performance by longer reasoning.
  • Figure 2: Evaluation accuracy per domain: LRMs outperform LLMs on most domains.
  • Figure 3: Vulnerability to different bias types: LRMs are significantly vulnerable to superficial quality biases.
  • Figure 4: The PlanJudge pipeline begins with the pairwise responses to be evaluated. The judge first construct an evaluation plan, and then derive the evaluation result by executing that plan.