- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper introduces BINGO, an agentic framework for discovering strong and diverse bias factors in LLM judges. The method follows a dual-loop optimization process. The inner loop uses an optimizer LLM to generate candidate bias factors, filters semantically redundant candidates using embedding similarities, and selects behaviorally novel candidates by comparing their induced judge-decision vectors against previously discovered factors. The outer loop evaluates selected factors on a larger validation set and uses a reflector LLM to summarize successful, unsuccessful, and redundant directions to guide subsequent exploration.
Critical strengths
Existing bias discovery approaches may focus on previously known factors or concentrate on a limited range of successful candidates. Encouraging exploration of distinct judge vulnerabilities is a meaningful direction, particularly as LLM-based evaluation becomes increasingly widespread. The dual-loop architecture provides a clear separation between diversity-oriented candidate selection and strength-oriented validation. The combination of semantic filtering, behavioral comparison, and reflective feedback is intuitive and straightforward to implement.
Critical weaknesses
- The operational definition of bias does not fully distinguish undesirable bias from legitimate differences in response quality. The paper defines a bias factor as something that systematically influences an LLM judge's decision independently of response quality or correctness. However, its principal bias-strength measure is based on the preference of the judge for a modified response over an initially identical reference response. The experimental design assumes that because the injector preserves the original content and conclusion, any systematic preference for or against the injected response reflects bias. This assumption requires further justification. Preserving correctness or the main conclusion does not necessarily imply preserving overall response quality.
- Behavioral diversity may not adequately reflect the diversity of underlying bias mechanisms. A central contribution is the explicit optimization of semantic and behavioral diversity. Semantic novelty is measured using cosine similarity between factor descriptions, while behavioral novelty is measured using the absolute Pearson correlation between vectors of judge decisions. Although these measurements provide useful operational proxies, it is unclear whether they accurately capture diversity in the underlying bias mechanisms. In particular, low behavioral correlation can result from differences in which individual examples are affected, even when two candidate factors correspond to closely related underlying mechanisms. Conversely, distinct mechanisms might influence the same examples in similar ways and therefore exhibit high correlation.
Questions for authors
- Can the authors independently validate the diversity of the discovered bias factors?
- How does BINGO compare with simpler diversity-aware search methods under matched computational budgets?
- What constitutes a genuinely novel bias factor? Could the authors provide more systematic evidence that factors such as Personal Anecdotes represent previously uncharacterized judge vulnerabilities rather than variations of known style-related preferences?
7 comments
Response to Reviewer bPSi [2/4]
[W2] & [Q1] Validating Whether Diversity Proxies Capture Distinct Bias Mechanisms
We thank the reviewer for raising this question. To assess whether BINGO's diversity proxies capture meaningful differences beyond same-mechanism variation, we first compare two types of factor pairs: 1) same-mechanism pairs (A vs. A′), created by paraphrasing factor definitions while preserving their mechanisms and2) different-factor pairs (A vs. B), drawn from the final top-5 factors within each run. Using 15 factors from three runs with GPT-4o as the target judge and the same 500 test queries as Table 2, we find that different-factor pairs exhibit substantially lower semantic similarity and behavioral correlation.
| Factor Pairs | Cosine Sim. | |
|---|---|---|
| Same-mechanism ( vs. ) | 0.931 ± 0.025 | 0.756 ± 0.065 |
| Different-factor ( vs. ) | 0.375 ± 0.001 | 0.206 ± 0.045 |
Furthermore, to examine whether BINGO's diversity measures capture meaningful differences beyond factor descriptions and judge decision patterns, we conduct an additional validation at the response level. We take each of the final top-5 factors in turn as a reference factor (A) and compare its response modifications against those induced by (1) semantically filtered candidates (description cosine similarity > 0.7 with A) and (2) the remaining four top-5 factors. We repeat this comparison for all five reference factors and report the average similarity for each group. To measure response-level similarity, we apply each factor to the same original responses and compute the cosine similarity between their response modification vectors (injected-response embedding − original-response embedding). Importantly, this metric is used only for validation and is not part of BINGO's optimization.
| Run | Semantically Filtered Candidates | Final Top-5 Factors |
|---|---|---|
| 1 | 0.699 | 0.310 |
| 2 | 0.735 | 0.318 |
| 3 | 0.670 | 0.345 |
| Mean | 0.701 | 0.324 |
Across three runs using GPT-4o as the target judge, filtered candidates exhibit substantially higher response-modification similarity than the final top-5 factors. This suggests that semantic filtering removes redundant modifications, while behavioral novelty-based selection retains factors producing more distinct response modifications.
Finally, to independently validate the diversity of discovered bias mechanisms, we employ three external LLM raters (Grok 4.3, Qwen 3.7 Flash, and Gemini 3.8 Flash) to assess factor distinctness on a 1–5 scale, using both factor descriptions and examples of their modifications to the same original responses. Across three runs per method, we evaluate 1) pairs among the top-5 discovered factors (30 pairs), 2) discovered factors against five pre-discovered biases (75 pairs; Table 6), and 3) each factor against its own paraphrase as a control (15 pairs). As shown in the table, BINGO consistently outperforms both baselines across all raters, achieving distinctness scores of 4.30–4.72 within discovered factors and 4.38–4.66 against pre-discovered biases. Meanwhile, paraphrase controls uniformly score 1.00, indicating that raters distinguish genuinely different mechanisms from alternative descriptions of the same mechanism. These findings independently support the diversity of BINGO's discovered bias factors.
| Judge | Method | Within-run | vs. Pre-discovered | Paraphrase Control |
|---|---|---|---|---|
| GPT-4o | BiasScope | 3.61 | 3.81 | 1.00 |
| OPRO | 3.24 | 4.12 | 1.00 | |
| BINGO | 4.72 | 4.66 | 1.00 | |
| Sonnet | BiasScope | 3.40 | 4.12 | 1.00 |
| OPRO | 3.89 | 4.20 | 1.00 | |
| BINGO | 4.37 | 4.53 | 1.00 | |
| DeepSeek | BiasScope | 3.53 | 3.90 | 1.00 |
| OPRO | 3.75 | 4.19 | 1.00 | |
| BINGO | 4.37 | 4.56 | 1.00 | |
| GPT-5.6 | BiasScope | 3.92 | 4.00 | 1.00 |
| OPRO | 3.97 | 4.17 | 1.00 | |
| BINGO | 4.30 | 4.38 | 1.00 |
Together, these results provide complementary evidence that BINGO's diversity proxies capture meaningful differences. These analyses are presented in Appendices K and L of the revised manuscript.
Response to Reviewer bPSi [3/4]
[Q2] Comparison with Simpler Diversity-Aware Baselines Under Matched Computational Budgets
We thank the reviewer for this valuable suggestion. To the best of our knowledge, BINGO is the first to explicitly incorporate diversity into automated LLM-judge bias discovery, making direct comparisons difficult. However, to address this concern, we adapt simpler diversity-aware methods from other domains and compare them under matched budgets.
Baselines. All methods share BINGO's injector, judge (GPT-4o), data, and seed factor. While these baselines promote diversity through relatively simpler strategies, such as diversified generation or independent search trajectories, BINGO employs a more refined selection mechanism based on semantic and behavioral novelty.
-
OPRO [3]. Our original baseline, representing BINGO's simplest variant without its key components. For a fair comparison, we explicitly encourage diversity through the optimizer prompt (Appendix B.3).
-
Verbalized Sampling [4]. Originally developed to diversify LLM generation, this method produces multiple outputs with estimated probabilities. We adapt it to generate and evaluate five bias factors per iteration, updating the prompt with the strongest factors found so far. Then, we pick the strongest top-5 from all evaluated factors.
-
Island Model [5]. Originally developed for evolutionary program search, this method promotes diversity by maintaining multiple independent search groups (islands) each exploring different directions. We run four separate OPRO searches, each using only its own factor history. At each step, one island is randomly selected to generate and evaluate a factor. Every 12 evaluated candidates, the two weakest islands restart from the best factor of a randomly selected stronger island. Finally, we select the top-5 factors from all four searches.
We conduct two comparisons: (1) cost-matched, with the same per-run budget as BINGO (~$16.25), and (2) validated candidate-matched, where baselines are run until they evaluate the same number of candidates as BINGO. A validated candidate is considered as a bias candidate that undergoes injection-and-judgment evaluation, excluding initial seeds and duplicate proposals. BINGO evaluates three candidates per iteration during behavioral novelty selection, yielding 42 evaluated candidates (3 × 14 iterations, excluding iteration 1 that uses initial factor only). OPRO and Island Model evaluate one candidate per iteration; Verbalized Sampling evaluates up to five, but evaluates fewer if any duplicate previously proposed factors. We report averages over three runs per method.
Cost-matched
| Method | Cost/run ($) | S(b)\ ↑ | CosSim Avg ↓ | CosSim Max ↓ | Avg ↓ | Max ↓ |
|---|---|---|---|---|---|---|
| OPRO | 16.66 ± 0.24 | 0.377 ± 0.006 | 0.545 ± 0.032 | 0.618 ± 0.025 | 0.458 ± 0.099 | 0.523 ± 0.082 |
| Verbal Sampling | 16.54 ± 0.15 | 0.346 ± 0.039 | 0.442 ± 0.054 | 0.486 ± 0.051 | 0.403 ± 0.021 | 0.482 ± 0.016 |
| Island Model | 16.38 ± 0.21 | 0.362 ± 0.019 | 0.537 ± 0.055 | 0.661 ± 0.059 | 0.465 ± 0.040 | 0.588 ± 0.061 |
| BINGO | 16.25 ± 0.20 | 0.416 ± 0.022 | 0.375 ± 0.001 | 0.448 ± 0.004 | 0.206 ± 0.045 | 0.297 ± 0.052 |
Validated candidates matched
| Method | Cost/run ($) | ↑ | CosSim Avg ↓ | CosSim Max ↓ | Avg ↓ | Max ↓ |
|---|---|---|---|---|---|---|
| OPRO | 33.27 ± 1.06 | 0.358 ± 0.031 | 0.550 ± 0.026 | 0.601 ± 0.009 | 0.447 ± 0.046 | 0.518 ± 0.046 |
| Verbal Sampling | 30.11 ± 0.23 | 0.371 ± 0.016 | 0.443 ± 0.036 | 0.518 ± 0.042 | 0.390 ± 0.050 | 0.444 ± 0.047 |
| Island Model | 31.09 ± 0.31 | 0.379 ± 0.010 | 0.463 ± 0.032 | 0.583 ± 0.060 | 0.389 ± 0.047 | 0.487 ± 0.055 |
| BINGO | 16.25 ± 0.20 | 0.416 ± 0.022 | 0.375 ± 0.001 | 0.448 ± 0.004 | 0.206 ± 0.045 | 0.297 ± 0.052 |
Under a matched cost, BINGO achieves the highest bias strength and lowest redundancy across all four diversity metrics. Moreover, even when we increase the baselines' budgets by approximately 2× to match BINGO's 42 evaluated candidates, all three methods continue to underperform BINGO in both bias strength and diversity. These results demonstrate that BINGO's improvements cannot be replicated simply by allocating larger search budgets to simpler diversity-aware methods. We have added these analyses to Appendix J.
Response to Reviewer bPSi [1/4]
Dear Reviewer bPSi,
We greatly appreciate your thoughtful and detailed analysis of our work. Below, we provide a thorough response to each of the points you raised.
[W1] Differentiating Judge Bias from Genuine Quality Differences
We thank the reviewer for raising this important concern. While it is true that correctness may not guarantee to fully capture all aspects of reasoning quality, our correctness-based evaluation provides a meaningful and objective anchor for assessing reasoning quality. We consider this approach for the following reasons. 1) Reasoning quality is inherently difficult to define and assess comprehensively, and while LLM-based evaluation is a viable alternative, its assessments may vary across models and prompts.In contrast, 2) answer reconstruction provides a well-defined and objectively verifiable criterion for assessing whether the original answer remains derivable from the modified reasoning. For these reasons, constrained injection with answer reconstruction approach builds upon established practices in prior work [1, 2], and we carefully adopt this approach and further validate its reliability through additional experiments in Appendix I.1.
Nevertheless, to more comprehensively assess overall response quality beyond correctness, we conducted additional analyses.
(1) Neutral Rewriting Control. To separate the effects of bias factors from those of rewriting itself, we generate neutral rewrites rn under identical injection settings but without a bias factor. Using GPT-5.1, we assess whether the overall response quality is similar or different for six factors (four from Appendix I and two additional factors), using 50 questions per factor. Evaluations are blinded and repeated with swapped response positions. Both orders yield similar quality judgments in 100% (r vs. rn), 97.3% (rn vs. r~), and 96.3% (r vs. r~) of cases, suggesting that rewriting alone does not introduce meaningful quality differences.
(2) Overall Quality Preservation. Beyond the correctness-focused checks in Appendix I, we directly assess overall response quality without mentioning correctness in the evaluation instructions. GPT-5.1 compares original and injected responses for the same six factors across 50 examples each, using blind, position-swapped evaluations. Additionally, seven human annotators each evaluate five pairs per factor, yielding 210 judgments. As the table below shows, GPT-5.1 detects a consistent quality difference in only 0.3% of comparisons, while human judgments confirm quality preservation in 94.8% of cases.
Together, these complementary evaluations provide further support that the observed preference shifts primarily arise from sensitivity to injected presentation styles rather than substantive differences in response quality. We have these results in Appendix I.2 of the revised manuscript.