A team of researchers from Allen Institute for Artificial Intelligence (Ai2), University of Washington and CMU introduce Fluid Benchmarking, an adaptive LLM evaluation method that replaces static accuracy with 2-parameter IRT ability estimation and Fisher-information–driven item selection. By asking only the most informative questions for a model’s current ability, it yields smoother training curves, delays benchmark saturation, improves external validity at small budgets, and filters mislabeled items.

Fluid Benchmarking replaces static accuracy with an adaptive, psychometrics-grounded procedure. A two-parameter logistic IRT model maps responses to a latent ability score and selects each next item by maximizing Fisher information at the model’s current ability estimate. Across six popular benchmarks and multiple model checkpoints, it improves validity (smaller rank distance), reduces variance (lower normalized total variation), delays saturation (more monotonic training curves), and avoids mislabeled items by ~100× compared to random sampling at equal budget.

What problem does Fluid Benchmarking solve?

Static subsets and plain accuracy conflate item quality and item difficulty, inflate step-to-step variance, and hit benchmark saturation early (training curves flatten while the model still improves). Fluid Benchmarking reframes both aggregation and selection: score in a latent ability space and adapt the item subset to the current ability, rather than treating all items equally or fixing them a priori.

How does it work?

1) Ability, not accuracy

Fit a 2-parameter logistic (2PL) IRT model on historical LM responses: for item j with discrimination aj​ and difficulty bj​, the probability a model with ability θi​ answers correctly is

p(uij​=1)=logistic(aj​(θi​−bj​))

At evaluation, estimate the MAP ability θ^i​ for the candidate LM by maximizing the 2PL likelihood over its observed right/wrong responses on the administered items. Items are weighted by their discrimination and difficulty, unlike accuracy which weights all equally

2) Dynamic item selection via Fisher information

At each step t, select the next item qj​ that maximizes Fisher information at the current ability estimate θ^(t):

I(θi​,aj​,bj​)=aj2​logistic(aj​(θi​−bj​))(1−logistic(aj​(θi​−bj​)))

High-information items minimize the variance of the ability estimate. As training progresses, the most informative items shift from easy to hard, so the administered subset evolves with model capability.

What does “better evaluation” mean here?

Fluid evaluates four dimensions with concrete metrics:

• Validity: external agreement with “true” model ranking; measured by mean rank distance (lower is better).
• Variance: normalized total variation of the training curve across checkpoints (lower is better).
• Saturation: monotonicity (Spearman rank correlation between checkpoint index and predicted performance; higher is better).
• Efficiency: quality at small item budgets.

How strong are the results?

Across six benchmarks (e.g., ARC-C, GSM8K, HellaSwag, MMLU, TruthfulQA, WinoGrande) and six LMs with 61–94 checkpoints each:

• Validity: On the smallest subset (AP-10), mean rank distance drops from 20.0 → 10.1; on AP-50, 15.2 → 8.8.
• Variance: Total variation shrinks markedly; e.g., 28.3 → 10.7 (AP-10) and 19.1 → 6.5 (AP-50).
• Saturation: Monotonicity improves from 0.48 → 0.76 (AP-10) and 0.62 → 0.86 (AP-50).
• Small-budget efficiency: With 10 items, Fluid improves mean rank distance by 9.9 vs. random; at 500 items, the improvement is 0.8—consistent with diminishing returns as budget grows.

In pretraining runs, accuracy space often looks flat late in training, but ability space continues to rise, delaying apparent saturation (e.g., HellaSwag monotonicity 0.91 → 0.99 for random vs. Fluid).

Fluid also avoids mislabeled items: on MMLU-Redux with 100-item budgets, mislabeled items per session drop from 0.75 (random) to 0.01 (Fluid)—about two orders of magnitude fewer.

Ablations isolate where the gains come from: IRT aggregation raises validity, but only dynamic selection lowers variance; “RANDOM-IRT” can even exceed random’s variance at large budgets, underscoring selection as the key lever.

Does it stop early when confident?

Yes. Fluid supports dynamic stopping using the standard error of the ability estimate; terminate when SE falls below the average ability gap between rank-adjacent LMs on the Open LLM Leaderboard. In practice, required items vary widely over training (≈20 early, >80 mid-run), showing why fixed budgets are suboptimal.

Where does it fit in the evaluation stack?

Fluid is benchmark-refinement: it does not invent new tasks; it re-weights and re-orders existing items to maximize information against a latent ability metric. It generalizes beyond pretraining to post-training and to other modalities, assuming enough responses to fit/update an IRT model. As models improve, IRT parameters must be refreshed to resolve difficulty among items that were previously “too hard,” otherwise the top of the scale compresses.

Summary

Fluid Benchmarking makes LLM evaluation budget-efficient and stable by scoring models in ability space and selecting items by Fisher information, yielding lower variance, better rank validity, and delayed saturation with far fewer questions. The trade-offs are operational: maintain fresh response matrices, periodically refit IRT parameters, and ensure reliable right/wrong binarization for open-ended tasks. As these practices standardize, Fluid becomes a practical default for in-loop pretraining and post-training evals across evolving benchmarks.

Check out the Paper, GitHub Page and Technical details. Feel free to check out our GitHub Page for Tutorials, Codes and Notebooks. Also, feel free to follow us on Twitter and don’t forget to join our 100k+ ML SubReddit and Subscribe to our Newsletter.

[Recommended Read] 🧵 NVIDIA AI Open-Sources ViPE (Video Pose Engine): A Powerful and Versatile 3D Video Annotation Tool for Spatial AI