The rapid release of both language models and benchmarks makes it increasingly costly to evaluate every model on every dataset. In practice, models are often evaluated on different samples, making scores difficult to compare across studies. To address this, we propose a framework based on multidimensional Item Response Theory (IRT) that uses anchor items to calibrate new benchmarks to the evaluation suite while holding previously calibrated item parameters fixed. Our approach supports a realistic evaluation setting in which datasets are introduced over time and models are evaluated only on the datasets available at the time of evaluation, while a fixed anchor set for each dataset is used so that results from different evaluation periods can be compared directly. In large-scale experiments on more than 400 models, our framework predicts full-evaluation performance within 2-3 percentage points using only 100 anchor questions per dataset, with Spearman ρ≥0.9 for ranking preservation preservation. This shows that benchmark suites can grow over time while preserving score comparability, since adding a new dataset requires running existing models only on that dataset's anchors. Code and data are available at: https://eliyahabba.github.io/growing-pains/
Figures & tables
Figure 1: Fixed parameter calibration enables extensible evaluation as new benchmarks are added over time. (1) Base datasets (e.g., MMLU, GSM8K) are calibrated jointly on reference models to define initial anchor item parameters. (2) At each subsequent step t, a new dataset is integrated by estimating its item parameters while holding all previously calibrated anchor parameters fixed (locked icons). (3) Once calibrated, the accumulated anchors serve as a compact proxy for the full suite, enabling performance prediction from anchor responses alone.
Table 1: Summary of the benchmark suites used in the experiments.
Dataset
Total Items
N=50
N=100
ARC Challenge
1,212
4.1%
8.3%
GSM8K
1,359
3.7%
7.4%
HellaSwag
10,082
0.5%
1.0%
MMLU
14,042
0.4%
0.7%
TruthfulQA
857
5.8%
11.7%
Winogrande
1,307
3.8%
7.7%
Table 2: Anchor coverage as a percentage of total items for the Open LLM Leaderboard suite (left) and for representative subjects in the MMLU suite (right).
Figure 2: Fixed parameter calibration evaluates previously tested models only on the anchors of each newly added dataset, keeping the evaluation size constant as the suite grows. Concurrent calibration uses an increasingly large anchor set as the suite grows. Random sampling has the same fixed evaluation size but higher prediction error.
Figure 3: A small anchor budget suffices for accurate prediction across chain steps, with diminishing returns from larger sets. Across both benchmark suites, even compact sets (e.g., N=10 or 25 ) allow fixed parameter calibration and concurrent calibration to maintain low and stable MAE. As the number of anchors increases, random sampling approaches the performance of IRT-based methods, narrowing the gap between all three approaches. Shaded regions denote 95% confidence intervals.
Figure 4: Prediction error remains low under time-ordered and family-held-out model splits on the Open LLM Leaderboard. Fixed parameter calibration tracks concurrent calibration and outperforms random sampling under both splits.
Figure 5: Prediction error stays low when an out-of-domain dataset is linked last, at increasing chain distance (the number of datasets added before it): GSM8K on the Open LLM Leaderboard (left), and the MMLU STEM (center) or mathematics (right) subjects. Fixed parameter calibration stays no worse than concurrent calibration and ahead of random sampling. Shaded regions denote 95% confidence intervals.
Open LLM Leaderboard
MMLU
Method \ Anchors ( N )
25
100
200
10
25
50
100
Random
0.82
0.91
0.96
0.68
0.85
0.91
0.98
Concurrent Calibration
0.89
0.94
0.97
0.73
0.83
0.90
0.98
Fixed Parameter Calibration
0.88
0.94
0.97
0.72
0.84
0.90
0.98
Table 3: Ranking preservation (Spearman ρ ) between predicted and full-evaluation model orderings. IRT-based calibration methods are competitive with random anchor sampling, with the clearest gains at small anchor counts.
Figure 6: The effect of reference model count varies across suites. On the Open LLM Leaderboard (top), a reliable fixed parameter calibration is only achieved once the reference pool reaches 100 or more models. On MMLU (bottom), prediction quality is relatively robust even with only 25 reference models. Shaded regions denote 95% confidence intervals.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Item maps showing the difficulty ( b ) and discrimination ( a ) of all items, with selected anchor items highlighted. Clustering-based selection (circles) distributes anchors across the full parameter space, while top- K selection (diamonds) concentrates them in a narrow high-discrimination region. Representative MMLU subjects.
Figure 8: Item maps showing the difficulty ( b ) and discrimination ( a ) of all items, with selected anchor items highlighted. Clustering-based selection (circles) distributes anchors across the full parameter space, while top- K selection (diamonds) concentrates them in a narrow high-discrimination region. Open LLM Leaderboard datasets.
Figure 9: Replacing the clustering-based anchor selection with top- K selection by discrimination parameter leads to substantially higher MAE on the Open LLM Leaderboard, while the rest of the fixed parameter calibration pipeline remains identical. Shaded regions denote 95% confidence intervals.
Figure 10: Replacing the clustering-based anchor selection with top- K selection by discrimination parameter leads to substantially higher MAE on MMLU, while the rest of the fixed parameter calibration pipeline remains identical. Shaded regions denote 95% confidence intervals.
Figure 11: Prediction error remains low under time-ordered (top) and family-held-out (bottom) model splits, on the Open LLM Leaderboard (left) and MMLU (right). Fixed parameter calibration tracks concurrent calibration and outperforms random sampling, matching the random-split result in Figure 3 . Shaded regions denote 95% confidence intervals.
Current LLM evaluation relies on two complementary but often disconnected signals: static benchmarks with objective correctness labels and arena-style preference data that better reflect open-ended user interactions. We introduce DualEval, a latent model-item calibration framework that represents models and evaluation items in a shared space, jointly estimating model ability together with item difficulty and sharpness. We apply DualEval across four domains: coding, math, miscellaneous domain-knowledge tasks, and generic everyday user queries. Our evaluation uses 18 frontier LLMs, static benchmark labels, and reward-model scores validated against held-out human preferences for open-ended model responses. Empirically, our framework produces reliable and balanced model rankings, and its learned item-level profiles support downstream applications such as benchmark compression for sample-efficient evaluation and anomaly detection for contamination or outlier analysis. Overall, DualEval unifies static and arena-style evaluation through joint model-item calibration, producing model rankings and item-level diagnostics that support more sample-efficient, interpretable, and auditable evaluation pipelines.
Aaron J. Li, Hao Huang, Youngmin Park +6
University of California, Berkeley · Arena · University of California, Los Angeles
Evaluating large language models (LLMs) typically requires thousands of benchmark items, making the process expensive, slow, and increasingly impractical at scale. Existing evaluation protocols rely on average accuracy over fixed item sets, treating all items as equally informative despite substantial variation in difficulty and discrimination. We introduce ATLAS, an adaptive testing framework based on Item Response Theory (IRT) that estimates model ability using Fisher information-guided item selection. ATLAS reduces the number of required items by up to 90% while maintaining measurement precision. For instance, it matches whole-bank ability estimates using only 41 items (0.157 MAE) on HellaSwag (5,600 items). We further reconstruct accuracy from ATLAS's ability estimates and find that reconstructed accuracies closely match raw accuracies across all five benchmarks, indicating that ability θ preserves the global performance structure. At the same time, θ provides finer discrimination within accuracy-equivalent models: among more than 3,000 evaluated models, 23-31% shift by more than 10 rank positions, and models with identical accuracies receive meaningfully different ability estimates. Code and calibrated item banks are available at https://github.com/Peiyu-Georgia-Li/ATLAS.git.
Safety benchmarks for language models are typically evaluated using static paradigms that treat all items as equally informative for all models, an assumption that is particularly problematic for adversarial, highly heterogeneous safety items. Applied in full to modern benchmark suites, the current evaluation procedures would require on the order of 105 responses, most of which provide little ranking signal. We analyze a suite of widely used safety benchmarks and make three contributions toward more efficient safety evaluation. First, we show that Item Response Theory (IRT) recovers interpretable structure on safety benchmarks, with ability estimates resolving differences among models that cluster at the ceiling of raw safety metrics. Second, we show that adaptive item selection, which dynamically chooses informative items for each model based on its responses, approximates full-benchmark rankings while reducing evaluation cost by at least 80% on benchmarks where Spearman's ρ>90% with full-benchmark is attainable, and by up to 99.9% on AIR-Bench 2024. Third, we introduce a practical procedure for extracting a fixed, informative subset of items reusable across models, providing an alternative to adaptive selection with savings of up to 99.8% on AIR-Bench 2024. Together, these results establish that psychometric methods enable benchmark-aware reductions in evaluation costs across the safety evaluation pipeline.
Fabio Spagliardi, Mírian Silva, Ayan Datta +3
Supervised Program for Alignment Research (SPAR) · Yale University · Mohamed bin Zayed University of Artificial Intelligence (MBZUAI)