Source-linked AI summary
When Wording Steers the Evaluation: Framing Bias in LLM judges
Yerin Hwang, Dongryeol Lee, Taegwan Kang, Minwoo Lee, Kyomin Jung
TL;DR
The impact of prompt framing on LLM-based evaluation has been largely unexplored, despite concerns that fixed-formulation evaluations may depend on question construction. The paper tests paired predicate-positive and predicate-negative prompts across four evaluation tasks and finds that all tested judges lack robustness, with bias patterns varying across model families and tasks.
Problem
Existing LLM evaluation frameworks rely on fixed prompt formulations and do not account for how framing may affect assessments.
Source: §1 · p. 1Method
The study compares paired predicate-positive and predicate-negative prompts across four high-stakes evaluation tasks.
Source: §1 · p. 1Results
All tested LLM judges lack robustness to predicate-positive and predicate-negative framings, with agreement patterns varying across model families and tasks.
Source: §5 · p. 4Takeaways & Limitations
Framing effects should be accounted for when designing fair and reliable LLM-based evaluation protocols.
Source: §5 · p. 4The study operationalizes framing as negation-based rephrasing and does not cover broader forms such as euphemism, conditional wording, or other pragmatic cues.
Source: Limitations · p. 5
Abstract
from arXiv · showhide
Large language models (LLMs) are known to produce varying responses depending on prompt phrasing, indicating that subtle guidance in phrasing can steer their answers. However, the impact of this framing bias on LLM-based evaluation, where models are expected to make stable and impartial judgments, remains largely underexplored. Drawing inspiration from the framing effect in psychology, we systematically investigate how deliberate prompt framing skews model judgments across four high-stakes evaluation tasks. We design symmetric prompts using predicate-positive and predicate-negative constructions and demonstrate that such framing induces significant discrepancies in model outputs. Across 14 LLM judges, we observe clear susceptibility to framing, with model families showing distinct tendencies toward agreement or rejection. These findings suggest that framing bias is a structural property of current LLM-based evaluation systems, underscoring the need for framing-aware protocols.
1 Introduction
Prompt phrasing can steer LLM responses, but its effects on LLM-based evaluation have been underexplored. This study tests paired framings across four high-stakes tasks and finds that framing bias affects evaluation outcomes.
1 Introduction
- Prompt-phrasing sensitivity raises concerns about the reliability and fairness of LLM applications as their use expands across domains.Source: §1 · p. 1
- Existing evaluation frameworks use fixed prompt formulations and do not account for framing, potentially making assessments depend on how questions are constructed.Source: §1 · p. 1
- The study pairs predicate-positive and predicate-negative prompts about the same content across truthfulness, jailbreak, toxicity, and grammatical acceptability evaluation.Source: §1 · p. 1
- All 14 evaluated LLM judges are vulnerable to framing bias, with model families and tasks showing distinct tendencies toward agreement or rejection.The findings motivate framing-aware protocols for fair and reliable evaluation.Source: §1 · p. 2
2 Related Works
Prior work describes LLM sensitivity to how tasks are framed, but its impact on LLM-based evaluation remains largely unexplored. This paper investigates framing-induced distortions in evaluation.
2.1 Framing bias in LLMs
- LLMs can produce different outputs from lexical or syntactic framing even when the underlying task remains the same.Source: §2.1 · p. 2
- Prior studies report framing-related inconsistent responses in mathematical reasoning and moral judgment, alongside a tendency to echo users’ viewpoints.Source: §2.1 · p. 2
- The effect of framing on LLM-based evaluation remains largely unexplored, which this work addresses by investigating framing-induced assessment distortions.Source: §2.1 · p. 2
3 Framing Bias in LLM judges
The section describes paired prompt framings, four evaluation tasks, and metrics for measuring inconsistent judgments and agreement or disagreement tendencies.
3.1 Framing Protocol
- Predicate-positive prompts ask whether property P holds, while predicate-negative prompts ask whether its negation holds for the same underlying content.A robust judge is expected to give consistent evaluations across these semantically equivalent framings.Source: §3.1 · p. 2
3.2 Evaluation Tasks
- The four tasks are truthfulness, jailbreak detection, toxicity detection, and grammatical acceptability, each evaluated with chain-of-thought prompting.The benchmarks comprise 1,000 TruthfulQA examples, 300 JailbreakBench items, 1,000 Jigsaw examples, and a 1,000-example CoLA subset.Source: §3.2 · pp. 2–3
3.3 Metrics
- Inconsistency measures incompatible judgments across framings, while acquiescence bias measures overall agreement and task-induced bias captures task-level agreement tendency beyond model-wide acquiescence.Source: §3.3 · p. 3
4 Experiments
Across four evaluation tasks, experiments with 14 LLM judges assess susceptibility to predicate-positive and predicate-negative prompt framings. No tested judge is fully robust, and framing direction varies across model families and tasks.
4 Experiments
- The experiments test how predicate-positive and predicate-negative framings affect judges and how tasks may inherently induce acquiescence bias.The study uses 14 general-purpose and reasoning-focused LLMs, including GPT, Qwen, and LLaMA models; temperature is set to zero where supported.Source: §4 · p. 3, §4.1 · p. 3
4.2 Results
- Some models with fewer than 3 billion parameters exceed 50% average inconsistency; GPT-5-mini has 5.69% average inconsistency, the lowest reported.No evaluated model demonstrates full robustness to framing effects.Source: §4.2 · p. 3, Table 1 · p. 4
- LLaMA models tend toward agreement, OpenAI models toward rejection, and Qwen variants show mixed directional patterns.The authors suggest family-level consistency may reflect intrinsic model properties, potentially shaped by instruction tuning, alignment, or decoding preferences.Source: §4.2 · pp. 3–4
- After controlling for model-level bias, grammatical acceptability and toxicity tend toward agreement, while jailbreak detection and truthfulness tend toward rejection.These are task-level directional tendencies, not uniform outcomes for every model.Source: §4.2 · p. 4
- Average inconsistency is highest for jailbreak detection, followed by truthfulness, toxicity, and grammatical acceptability.Figure 2’s right panel presents task-level inconsistency averaged across judges.Source: §4.2 · p. 4, Figure 2 · p. 4
- Qualitative analysis finds that models sometimes give inconsistent judgments across framings and then attempt to justify them with flawed reasoning.Representative examples are in Appendix C.Source: §4.2 · p. 4
5 Conclusion
The study finds that predicate-positive and predicate-negative framings produce inconsistent evaluations across tested LLM judges. Agreement-bias patterns vary across model families and tasks, motivating attention to framing in LLM-based evaluation protocols.
5 Conclusion
- Predicate-positive and predicate-negative framings pose logically equivalent questions in affirmative and negated forms, yet all tested judges lack robustness to them.The study reports agreement-bias patterns that vary across model families and tasks.Source: §5 · p. 4
Limitations
The study tests framing bias with four high-stakes evaluation tasks, while noting that its scope is limited by its task selection, English-language experiments, and negation-based definition of framing. Its controlled prompt pairs isolate wording effects, but differ from end-to-end judge prompts that can vary along multiple dimensions.
Limitations
- The evaluation covers four representative high-stakes tasks, leaving domains such as healthcare, financial risk, and code safety for future study.The authors also call for cross-lingual and cross-domain evaluation to establish whether framing sensitivity persists or changes in multilingual and specialized settings.Source: Limitations · p. 5
- The experiments are conducted primarily in English, so multilingual and specialized settings remain unevaluated.The authors identify cross-lingual and cross-domain robustness as future work.Source: Limitations · p. 5
- Framing is operationalized as predicate-positive versus predicate-negative rephrasing, excluding broader cues such as euphemism and conditional wording.Expanding the framing taxonomy is left for future work.Source: Limitations · p. 5
- Controlled prompt pairs isolate wording effects, but differ from end-to-end judge prompts that may also vary in rubrics, examples, context, and formatting.The authors frame the paired setup as a diagnostic stress test for stable evaluation under targeted rephrasing.Source: Limitations · p. 5
Ethics Statement
The study uses public datasets for evaluation, follows model intended-use and dataset policies, and reports AI assistance in sentence-level drafting and revision. Experiments run on two NVIDIA A100 GPUs.
Ethics Statement
- Public datasets are used for their original purposes; toxicity data are used only for evaluation under the dataset’s usage policies and content-handling guidelines.The authors state that no personally identifying or offensive content was included.Source: Ethics Statement · p. 5
- The authors used AI assistants for sentence-level drafting and revision.Models were sourced from official websites or Hugging Face and operated in line with intended use.Source: Ethics Statement · p. 5
A.1 Computing Resources
- Experiments used two NVIDIA A100 GPUs, each with 80 GB of VRAM, and Python 3.10.14 with PyTorch 2.6.0 and CUDA 12.4.Source: §A.1 · p. 8
A.2 Experimental Configuration of LLMs
The study evaluates closed- and open-source GPT, QWEN2.5-INSTRUCT, and LLAMA-3-INSTRUCT models. Configurable models use zero-temperature decoding, with medium effort for reasoning-enabled models with a configurable effort parameter.
A.2 Experimental Configuration of LLMs
- The GPT evaluation includes o4-mini, GPT-4o, GPT-5-mini, and GPT-5, accessed through the official OpenAI platform.The listed API identifiers are o4-mini-2025-04-16, gpt-4o-2024-08-06, gpt-5-mini-2025-08-07, and gpt-5-2025-08-07.Source: §A.2 · p. 8
- The open-source roster includes six QWEN2.5-INSTRUCT models from 1.5B to 72B parameters and four LLAMA-3-INSTRUCT models, accessed through Hugging Face.The listed LLAMA models are LLAMA-3.1-8B-INSTRUCT, LLAMA-3.1-70B-INSTRUCT, LLAMA-3.2-1B-INSTRUCT, and LLAMA-3.3-70B-INSTRUCT.Source: §A.2 · p. 8
- Configurable models use zero-temperature decoding, and configurable reasoning-enabled models use medium effort; other settings remain at their defaults.Results are reported from a representative run.Source: §A.2 · p. 8
B Metric Definitions
The metrics compare binary judgments under predicate-positive and predicate-negative framings, using paired parseable items and their outcome proportions. They quantify inconsistency, agreement, model-level acquiescence bias, and task-induced bias, with optional aggregation across tasks.
B Metric Definitions
- The P–¬P cell proportions distinguish the four paired outcomes: (P=1, ¬P=0), (P=0, ¬P=1), (P=1, ¬P=1), and (P=0, ¬P=0).The empirical yes rates under each framing are computed from these cell proportions.Source: B Metric Definitions · p. 8
- Pairwise Inconsistency Rate (PIR) measures disagreement between paired binary judgments under P and ¬P framings.The metrics use paired, parseable items and the proportions of each P–¬P judgment combination.Source: B Metric Definitions · p. 8
- The metrics include model-level acquiescence bias and task-induced bias controlled for the models scored on each task.The task-level agreement rate is also defined, and values may optionally be aggregated across tasks.Source: B Metric Definitions · p. 8
C Qualitative analysis
The qualitative examples show framing-related instability in judges’ reasoning, while a repeated-trial control distinguishes rephrasing inconsistency from run-to-run randomness. The evaluation prompts use chain-of-thought cues but score only the final canonicalized label.
C Qualitative analysis
- In two jailbreak examples, GPT-5 gives identical verdicts under opposite framings while producing contradictory rationales.The rationales are also erroneous and mutually inconsistent.Source: C Qualitative analysis · pp. 8–9, Table 2 · p. 9
- Qwen 2.5 3B frequently assigns the same terminal label, No, to the same output under both framings.This is the characteristic default-to-No pattern found in the qualitative inspection.Source: C Qualitative analysis · p. 9
D Prompts for LLM-based Evaluation
- Evaluation templates use a brief chain-of-thought cue and request a final canonicalized label; only that label is scored, while free-form rationale is ignored.The prompts elicit intermediate reasoning, but rationale does not enter metric computation.Source: D Prompts for LLM-based Evaluation · p. 9
E Stochastic Baseline for Inconsistency
- 5.69% framing inconsistency versus 0.8% stochastic inconsistency for GPT-5-mini suggests rephrasing instability is not primarily sampling noise.The control repeated each example three times under identical prompts.Source: E Stochastic Baseline for Inconsistency · p. 9