Fixed-Pool Data Recipe Search

• Fixed-pool data recipe search is a framework for selecting optimal data subsets from a static pool using composable, order-sensitive operators.
• The method employs filtering, deduplication, and stratified sampling to transform and curate data under strict compute and evaluation constraints.
• Controller designs using reinforcement learning, Bayesian search, and heuristic methods efficiently navigate the complex search space to boost model performance.

Fixed-pool data recipe search refers to the search for an optimal or near-optimal data selection or curation procedure ("data recipe") within a static pool of available samples, sources, or labeled/unlabeled data, typically under compute or evaluation budget constraints. Unlike data augmentation or instance synthesis, fixed-pool approaches restrict selection, filtering, and transformation operations to rearrangements and subsetting of an immutable base pool; sampling formulas, multi-stage filtering, operator sequencing, and mixture optimization are all possible ingredients of the search. This framework has emerged as a unifying setting for LLM adaptation, supervised fine-tuning, retrieval, heuristic learning for search, time-series forecasting, and other scenarios in large-scale machine learning pipelines (Wu et al., 13 May 2026, Chen et al., 11 Feb 2026, Pennino et al., 12 Dec 2025, Gupta et al., 2024).

1. Formalization and Problem Statement

Fixed-pool data recipe search replaces per-instance ranking with the search for an executable data selection pipeline, or recipe, composed of order-sensitive operators or mixture weights, acting on a finite initial pool $\mathcal{D}_0$. The most general formulation consists of:

• A static data pool $\mathcal{D}_0 = \{d_1, \ldots, d_N\}$ (instances or source corpora $\{D_i\}$).
• A library of subset transformers (filtering by precomputed scores, deduplication, mixing, stratified or importance sampling, operator parameters $\theta\in\Theta$).
• Recipes $r = [(o_1,\theta_1), \ldots, (o_L,\theta_L)]$ (sequences of operator+parameter pairs), each deterministically mapping $\mathcal{D}_0$ to a subset $S(r)$.
• Evaluation function $y(r)$: downstream metric (e.g., SFT accuracy, RL reward, held-out loss), requiring model fine-tuning or inference on $S(r)$.

The search problem is

$$r^* = \arg\max_{r\in\mathcal{R}} y(r) \quad \text{s.t. } \mathrm{cost}(r) \leq B$$

where $\mathcal{D}_0 = \{d_1, \ldots, d_N\}$0 is the space of all admissible recipes (subject to operator syntax and budget $\mathcal{D}_0 = \{d_1, \ldots, d_N\}$1 of expensive evaluations) (Wu et al., 13 May 2026, Chen et al., 11 Feb 2026). In mixture-based settings, recipes may be parameterized as $\mathcal{D}_0 = \{d_1, \ldots, d_N\}$2-way mixtures with simplex-constrained weights $\mathcal{D}_0 = \{d_1, \ldots, d_N\}$3, with $\mathcal{D}_0 = \{d_1, \ldots, d_N\}$4 built by stratified sampling (Pennino et al., 12 Dec 2025).

2. Operator Libraries, Recipe Structures, and Search Spaces

Recipe operators in the fixed-pool paradigm are composable, subset-transforming building blocks:

• Filtering: Rank-and-threshold by model/dataset scores (activation similarity, varentropy, n-gram entropy, difficulty metrics, or domain distance), typically parameterized by top-$\mathcal{D}_0 = \{d_1, \ldots, d_N\}$5 cutoffs.
• Deduplication: Remove near-duplicate samples via embedding-space or n-gram similarity above a preset threshold.
• Mixing/Union: Set unions with previously high-value subsets.
• Sampling: Importance or stratified sampling with data-, search-, or task-driven weighting.
• Mixture weights: For clustering-based recipes, clusters act as categories, with sampling weights tuned via black-box optimization (Pennino et al., 12 Dec 2025).

Recipes may include both order-sensitive sequencing and parametric mixture weighting. This leads to a high-dimensional, often discrete-continuous, nonconvex search space (Wu et al., 13 May 2026, Chen et al., 11 Feb 2026).

3. Search Algorithms and Controller Design

With nontrivial search space complexity and expensive black-box evaluation, recent approaches focus on controller designs that maximize recipe evaluation budget efficiency:

• AutoSelection (Wu et al., 13 May 2026): A two-layer framework. Layer A caches task/data/model signals for fast, operator-level recipe materialization. Layer B proposes and ranks candidate recipes via local edits and Gaussian-process surrogate modeling over a joint recipe-state feature space, updating with full SFT evaluations.
• Reinforcement Learning (RL) Controllers (Chen et al., 11 Feb 2026): DataChef models recipe construction as a Markov Decision Process (MDP) with policy $\mathcal{D}_0 = \{d_1, \ldots, d_N\}$6 over operator/action selection, optimized for downstream validation metrics via RL (Group Relative Policy Optimization/GRPO), using learned proxy rewards (e.g., LLM-based Data Verifier regression oracles) to amortize evaluation cost.
• Bayesian/Black-box Search (Pennino et al., 12 Dec 2025): Mixture weights are tuned by black-box optimizers—e.g., Optuna TPE—iteratively sampling, evaluating, and refining $\mathcal{D}_0 = \{d_1, \ldots, d_N\}$7 to minimize target loss.
• Stratified/importance subsampling: Oracle-derived importance weights, e.g., for A* LM heuristics, using task-derived contribution functions in sampling distributions (Gupta et al., 2024).

Controllers may employ warmup probing, local edits, stagnation-triggered reseeding, and surrogate priors for exploitation/exploration balancing (Wu et al., 13 May 2026).

Method/paper	Controller type	Search granularity
AutoSelection	Summarizer+GP+Reranker	Recipes: operator seq
DataChef	RL policy + verifier	Pipeline, Py/SMPL
Optuna search	TPE sampler (black-box)	Cluster weights
Heuristic coreset	Oracle importance sample	Node-level (A*)

4. Empirical Results and Application Domains

Empirical validations across domains show that fixed-pool data recipe search achieves systematic improvements over baselines:

• LLM supervised fine-tuning: AutoSelection outperforms all single-score selectors and full-data runs (gains up to +3.7 points) on OpenHermes/LESS/Alpaca-52K mixes for Qwen2.5-1.5B/3B and Llama3.2-1B, and remains competitive OOD (e.g., GraphWiz/NLGraph) (Wu et al., 13 May 2026).
• LLM adaptation RL: DataChef-32B matches or exceeds human-curated recipes and commercial LLMs across six held-out benchmarks; on AIME’25, DataChef’s oracle achieves 66.7, beating the industry pipeline for Qwen3-1.7B (Chen et al., 11 Feb 2026).
• A* LM heuristic training: Sampling nodes by contribution-weighted importance yields up to 15× reduction in node expansions, 5× speedup over uniform or full-data coresets (Gupta et al., 2024).
• Time series forecasting: Clustering+mixture search achieves 19.41% MSE reduction while using only ≈43% of data (Pennino et al., 12 Dec 2025).

5. Insights, Limitations, and Design Implications

Fixed-pool data recipe search offers several key advantages:

• Operator composition and structure matter: Searchable multi-operator pipelines consistently outperform single-threshold selectors, and ablations confirm order/parameterization is critical (Wu et al., 13 May 2026).
• Data, task, and model-side signals: Cached statistics (activation similarities, entropy, domain distances, etc.) provide inexpensive proxies for recipe screening, amplifying full-evaluation budgets (Wu et al., 13 May 2026, Pennino et al., 12 Dec 2025).
• Cross-domain generalization: The framework generalizes to vision, NLP, tabular, and sequence tasks wherever pool-based recipe construction is feasible (Pennino et al., 12 Dec 2025).
• Proxy rewards and search efficiency: RL-based controllers using learned proxies (LLM “Data Verifier”) can amortize hundreds of downstream training runs per task (Chen et al., 11 Feb 2026).

Principal limitations include:

• Fixed-pool restriction: No synthesis or expansion beyond the initial pool; rare or missing sample types remain unrepresented (Chen et al., 11 Feb 2026).
• Combinatorial scaling: Growth of the operator library or number of source pools increases search complexity—future strategies may require hierarchical or Bayesian sampling (Chen et al., 11 Feb 2026).
• Surrogate accuracy: Proxy models for reward prediction may underperform on highly specialized or novel domains, leading to degraded policy learning (Chen et al., 11 Feb 2026, Pennino et al., 12 Dec 2025).

6. Research Extensions and Practical Recommendations

Emerging research directions span:

• Automated data curation: Treating data selection workflows as programmatic objects lends interpretability, extensibility, and reproducibility to LLM/adaptation pipelines (Wu et al., 13 May 2026, Chen et al., 11 Feb 2026).
• Operator sets and recipe search augmentation: Incorporating new operator types (semantics-aware deduplication, adaptive importance functions, resume-from-best), expanding modularity and domain flexibility.
• Dynamic, online recipe search: Fixed-pool algorithms may be adapted for curriculum learning or streaming data with evolving pool statistics.
• Cross-task transfer: Transferability of recipe structures (found on small-scale models) to larger models is moderate (Spearman ρ ≈ 0.82 for cross-scale transfer in AutoSelection), but target-model evaluation remains essential (Wu et al., 13 May 2026).
• Practitioner guidance: Practitioners are advised to cache relevant data/model signals, begin with warmup probes covering a range of subset sizes, and exploit local edits and surrogate evaluations for maximal efficiency under tight budgets (Wu et al., 13 May 2026, Pennino et al., 12 Dec 2025).

• "From Instance Selection to Fixed-Pool Data Recipe Search for Supervised Fine-Tuning" (Wu et al., 13 May 2026)
• "DataChef: Cooking Up Optimal Data Recipes for LLM Adaptation via Reinforcement Learning" (Chen et al., 11 Feb 2026)
• "Optimizing the Training Diet: Data Mixture Search for Robust Time Series Forecasting" (Pennino et al., 12 Dec 2025)
• "A Training Data Recipe to Accelerate A* Search with LLMs" (Gupta et al., 2024)