Source-linked AI summary
Wider and Deeper LLM Networks are Fairer LLM Evaluators
Xinghua Zhang, Bowen Yu, Haiyang Yu, Yangyu Lv, Tingwen Liu, Fei Huang, Hongbo Xu, Yongbin Li
TL;DR
Evaluating whether generated responses align with human preferences is challenging, while common metrics correlate weakly with human judgments and human evaluation is costly. The paper proposes a role-specialized, multilayer LLM evaluator and evaluates it on a diverse benchmark. A wider two-layer network performs best, and the approach also supports Chinese LLM assessment with reduced evaluation time and annotation cost.
Problem
Open-ended response evaluation must assess alignment with human intent, but traditional metrics correlate weakly with human judgments and human evaluation is costly.
Source: §1 · p. 2Method
The method assigns distinct evaluation perspectives to first-layer LLM neurons and has later layers integrate representations from all neurons in the preceding layer.
Source: §1 · pp. 2–3Results
A wider two-layer LLM network performs best, and WideDeep reports a 0.06 enhancement in kappa correlation coefficient.
Source: §1 · p. 3Takeaways & Limitations
For Chinese LLM assessment, WideDeep accelerated evaluation 4.6 times and decreased average annotation cost per sample by 60%.
Source: §1 · p. 3
Abstract
from arXiv · showhide
Measuring the quality of responses generated by LLMs is a challenging task, particularly when it comes to evaluating whether the response is aligned with human preference. A novel approach involves using the LLM itself to make evaluation and stabilizing the results through multiple independent evaluations, similar to a single-layer narrow LLM network. This network consists of a fixed number of neurons, with each neuron being the same LLM. In this paper, we draw upon the extensive research on deep neural networks to explore whether deeper and wider networks can lead to fairer evaluations. Specifically, inspired by the observation that different neurons in a neural network are responsible for detecting different concepts, we first adaptively generate as many neuron roles as possible for each evaluation sample. Each perspective corresponds to the role of a specific LLM neuron in the first layer. In subsequent layers, we follow the idea that higher layers in deep networks are responsible for more comprehensive features, each layer receives representations from all neurons in the previous layer, integrating the locally learned evaluation information to obtain a more comprehensive evaluation result. Interestingly, this network design resembles the process of academic paper reviewing. To validate the effectiveness of our method, we construct the largest and most diverse English evaluation benchmark LLMEval$^2$ for LLM evaluators, comprising 15 tasks, 8 abilities, and 2,553 samples. Experimental results demonstrate that a wider network (involving many reviewers) with 2 layers (one round of discussion) performs the best, improving kappa correlation coefficient from 0.28 to 0.34. We also leverage WideDeep to aid in the assessment of Chinese LLMs, which has accelerated the evaluation time by 4.6 times, resulting in a 60% cost saving. WideDeep achieves a remarkable 93% agreement level among humans.
1 Introduction
The introduction frames LLM evaluation as a human-alignment problem that existing evaluators and narrow benchmarks do not adequately address. It proposes role-specialized, multilayer LLM evaluators and introduces a broad benchmark, reporting the best results with a wider two-layer design.
1 Introduction
- Open-ended responses make alignment with human intent difficult; traditional metrics correlate weakly with human judgments, while human evaluation is costly.Prior LLM evaluators also face position bias, and existing evaluation benchmarks lack diverse capabilities.Source: §1 · p. 2
- Each sample’s candidate evaluation perspectives become first-layer neuron roles, while later layers combine prior neurons’ scores and reasons into broader evaluations.The authors adaptively generate perspectives and explicitly assign them to LLM neurons.Source: §1 · pp. 2–3
- A wider LLM network with two layers performs best, resembling a paper-review process with many reviewers and one discussion round.The paper compares its network design to reviewers independently scoring, discussing, and then having a chair consolidate opinions.Source: §1 · p. 3
- LLMEval2 covers 15 tasks and 8 abilities with 2,553 human-annotated preference samples, 31 times FairEval’s dataset size.The introduction presents it as the largest and most diverse benchmark for LLM evaluators.Source: §1 · p. 3
- WideDeep experiments report a 3.5-point increase in accuracy and a 0.06 enhancement in kappa correlation coefficient.The introduction also reports validation on two existing benchmarks and LLMEval2.Source: §1 · p. 3
- 6 points in accuracy, 5.5 points in F1 score, and 0.09 in kappa correlation coefficient accompany Chinese evaluation, with 74% labeling accuracy and 93% human agreement.WideDeep accelerated evaluation 4.6 times and decreased average annotation cost per sample by 60%.Source: §1 · p. 3
2 Related Work
Related work describes limitations in both conventional text metrics and LLM-based evaluation. Prior LLM judges address some biases through response-order swaps and repeated evaluations, while the paper frames its approach as a wider, deeper alternative.
2 Related Work
- N-gram and embedding metrics measure lexical or semantic similarity but correlate relatively weakly with human judgments of response quality.Traditional multiple-choice benchmarks may also fail to distinguish aligned models from base models.Source: §2 · p. 3
- LLM evaluators can assess generation tasks, but investigations identify position and verbosity biases in their judgments.Prior methods swap candidate order and run multiple evaluations to mitigate bias and improve reliability.Source: §2 · p. 4
- The paper characterizes prior LLM evaluators as narrow, single-layer networks and explores assigning distinct roles across wider, deeper layers.Each layer takes evaluation outputs from all neurons in the preceding layer.Source: §2 · p. 4
3 Methodology
The methodology widens and deepens LLM evaluation by assigning sample-specific roles to first-layer neurons, passing prior-layer assessments upward, and aggregating judgments. It explains this design through a three-stage academic paper-review analogy.
3.1 Deeper and Wider LLM Network
- Sample-specific perspectives give first-layer LLM neurons distinct evaluation roles, while later layers combine their predecessors’ representations to form more comprehensive assessments.The roles are generated for each evaluation sample; the paper gives coherence, relevance, harmlessness, and accuracy as example perspectives.Source: §3.1 · pp. 4–5
- Because the LLM neurons have no trainable numerical weights, prompts connect each neuron to all neurons in the preceding layer.The paper describes this prompt-based connection as inspired by Stacked LLMs [30].Source: §3.1 · p. 5
- The network can aggregate judgments by averaging neuron scores before comparing responses, or by having neurons choose and vote for a preferred response.Voting can use neurons across all layers or neurons within each layer.Source: §3.1 · p. 6
3.2 Explain LLM Network as a Academic Paper Review Process
- Blind-review reports correspond to input-layer assessments, discussion rounds to later layers, and the chair’s decision to final aggregation.The analogy is depicted as blind review, reviewer discussion, and chair summary.Source: §3.2 · pp. 6–7, Figure 2 · p. 7
4 LLMEval2 Benchmark
LLMEval2 is presented as a larger, more diverse benchmark for evaluating LLM evaluators. It comprises 2,553 human-labeled samples from 15 datasets and covers varied tasks, abilities, and response preferences.
4 LLMEval2 Benchmark
- LLMEval2 is presented as addressing prior benchmark limits: small test sets in FairEval and MT-bench, and limited ability statistics and source diversity in PandaLM.The paper describes LLMEval2 as the largest and most diverse LLM-evaluator benchmark to date.Source: §4 · p. 8
- 2,553 human-preference-annotated samples were collected across 15 datasets using sampling to balance data diversity and evaluation costs.The benchmark construction draws on datasets used to train reward models.Source: §4 · p. 8
- The benchmark covers eight tasks and eight abilities, with 1,050 samples preferring response 1, 1,021 preferring response 2, and 482 judged similar in quality.The listed tasks include dialogue, retrieval QA, and programming; abilities include logical reasoning, harmlessness, and multilingual capability.Source: §4 · p. 8
5 Experiments
The experiments test whether wider and deeper LLM evaluators improve performance, which neuron roles they prioritize, and whether they accelerate manual annotation.
5.1 Experimental Settings
- The experiments use FairEval, PandaLM, and LLMEval2, report accuracy, Macro-F1, and kappa, and use GPT-3.5-turbo as the neuron on the full dataset.LLMEval2 mini contains 300 samples drawn as 20 from each of its 15 datasets for analytical experiments.Source: §5.1 · p. 8
5.2 Experimental Results
- Accuracy increased 3.2, 4.4, and 3 points, while kappa correlation increased 3.7, 8.4, and 6.3 points across three benchmarks versus FairEval; voting across all layers performed better.The main results compare WideDeep with FairEval and find that voting across all layers performs better than voting in individual layers.Source: Table 1 · p. 9, §5.2 · p. 9
5.3 Experimental Analyses
- With one or two layers, increasing neuron count consistently improves performance; moving from one to two layers also helps, while deeper networks slightly decline, possibly from more homogeneous information.The analysis reports the deeper-network explanation as a possibility.Source: §5.3 · pp. 9–10, Table 2 · p. 10
- Generated roles adapt to task—security for harmlessness, coherence and character for stories, and correctness and efficiency for programming—and removing roles lowers accuracy and Macro-F1.Figure 5 shows roles for dialogue, harmlessness QA, story generation, and programming; the authors say they did not provide task names or definitions when generating roles.Source: Figure 5 · p. 10, §5.3 · pp. 10–11, Table 3 · p. 10
- Across almost all neuron-count constraints, WideDeep outperforms FairEval and continues improving as neurons increase, whereas FairEval saturates at five neurons and then declines.The WideDeep experiments use two layers, reducing the second layer by one neuron when the constraint is odd.Source: Figure 6 · p. 11, §5.3 · p. 11
5.4 Application in Chinese LLM Evaluation
- In Chinese LLM evaluation, WideDeep gains 6, 5.5, and 8.9 points in accuracy, F1, and kappa, respectively, and reaches 74% labeling accuracy and 93% of human agreement.Human agreement is 80%.Source: Table 4 · p. 11, §5.4 · p. 12
- WideDeep speeds evaluation 4.6 times and cuts average annotation cost per sample by 60%; annotators inspect 23% of its predictions versus 37.5% for FairEval and 39.3% for GPT-4.The inspection rates are given for reaching 80% accuracy, aligning with manual annotation.Source: §5.4 · p. 12
6 Conclusion
The paper concludes that a two-layer wider LLM network yields the best results for evaluating generated text. It also reports applying the evaluator to Chinese LLM assessment, where evaluation is faster and annotation costs are lower.
6 Conclusion
- A two-layer wider LLM network yields the best results and significantly enhances LLMs’ ability to evaluate generated text.The network uses distinct evaluation roles for neurons that interact and collaborate across layers.Source: §6 · p. 12
- The evaluator speeds up Chinese LLM assessment by 4.6 times and decreases average annotation cost per sample by 60%.Source: §6 · p. 12
- The paper presents LLMEval2 as the largest and most diverse evaluation benchmark developed to date for LLM evaluators.Source: §6 · p. 12
- The evaluation process is feedforward, with each layer receiving inputs from the previous layer for a comprehensive assessment.The paper likens the network to academic paper reviewing.Source: §6 · p. 12