REVIEW 9 cited by
Assessing Judging Bias in Large Reasoning Models: An Empirical Study
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Assessing Judging Bias in Large Reasoning Models: An Empirical Study
read the original abstract
Large Reasoning Models (LRMs) like DeepSeek-R1 and OpenAI-o1 have demonstrated remarkable reasoning capabilities, raising important questions about their biases in LLM-as-a-judge settings. We present a comprehensive benchmark comparing judging biases between LLMs and LRMs across both subjective preference-alignment datasets and objective fact-based datasets. Through investigation of bandwagon, authority, position, and distraction biases, we uncover four key findings: (1) despite their advanced reasoning capabilities, LRMs remain susceptible to the above biases; (2) LRMs demonstrate better robustness than LLMs specifically on fact-related datasets; (3) LRMs exhibit notable position bias, preferring options in later positions; and (4) we identify a novel "superficial reflection bias" where phrases mimicking reasoning (e.g., "wait, let me think...") significantly influence model judgments. To address these biases, we design and evaluate three mitigation strategies: specialized system prompts that reduce judging biases by up to 19\% in preference alignment datasets and 14\% in fact-related datasets, in-context learning that provides up to 27\% improvement on preference tasks but shows inconsistent results on factual tasks, and a self-reflection mechanism that reduces biases by up to 10\% in preference datasets and 16\% in fact-related datasets, with self-reflection proving particularly effective for LRMs. Our work provides crucial insights for developing more reliable LLM-as-a-Judge frameworks, especially as LRMs become increasingly deployed as automated judges.
Forward citations
Cited by 9 Pith papers
-
Cited but Not Verified: Parsing and Evaluating Source Attribution in LLM Deep Research Agents
A new framework parses and evaluates citations in LLM deep research reports across link validity, relevance, and factuality, finding 94%+ link success but only 39-77% factual accuracy.
-
Beyond Semantic Manipulation: Token-Space Attacks on Reward Models
TOMPA performs black-box adversarial optimization in token space to discover non-linguistic patterns that nearly double the reward scores of GPT-5 answers on Skywork-Reward-V2 while producing gibberish text.
-
Fragile Thoughts: How Large Language Models Handle Chain-of-Thought Perturbations
LLMs show heterogeneous robustness to five types of chain-of-thought perturbations, with MathError causing 50-60% accuracy loss in small models but scaling benefits, UnitConversion remaining hard across sizes, and Ext...
-
A Mechanistic View of Authority Hierarchy in LLM Sycophancy
Authority sycophancy in LLMs is a layer-localized erasure of correct answer representations that scales with authority level and resists simple interventions.
-
An Enigma of Artificial Reason: Investigating the Production-Evaluation Gap in Large Reasoning Models
LRMs show a large production-evaluation gap on the VAIR dataset with valid answers but invalid reasoning, driven by answer confirmation bias as evidenced by CoT analysis, linear probes, and causal patching.
-
Label Effects: Shared Heuristic Reliance in Trust Assessment by Humans and LLM-as-a-Judge
Both humans and LLMs trust content more when labeled human-authored than AI-generated, with LLMs showing denser attention to labels and higher uncertainty under AI labels, mirroring human heuristic patterns.
-
Who Endorsed It? Measuring Authority Bias Across Expertise Levels in Language Models
Language models accept misleading endorsements more readily from higher-expertise persona sources, and the effect can be partially steered away by subtracting a residual-stream vector.
-
The Saturation Trap and the Subjectivity of Intervention Timing: Why Affect-Based Triggers and LLM Judges Fail to Time Interventions on Autonomous Agents
Intervention timing for autonomous agents is a low-reliability construct because affective triggers saturate, LLM judges underperform, and human annotators show low inter-rater agreement on intervention points.
-
LLM-as-a-Judge for Human-AI Co-Creation: A Reliability-Aware Evaluation Framework for Coding
LLM judges for human-AI coding co-creation show moderate performance (ROC-AUC 0.59) and low agreement, with co-creation success concentrating early in interactions.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.