betaFeedback

WorksComputer Science

Aligning Reasoning LLMs for Materials Discovery with Physics-aware Rejection Sampling

Abstract

AI-driven materials discovery that couples automated experimentation with algorithmic decision-making requires process aware recipe to property predictors that are accurate, calibrated, and physically admissible. We approach this as a reasoning problem with large reasoning models (LRMs). To instill reasoning capability into language models, we curate reasoning traces from a teacher model to train a student model. However, most training pipelines select reasoning traces using binary correctness or learned preference signals that poorly reflect physical admissibility. We introduce Physics-aware Rejection Sampling (PaRS), a training-time trace selection scheme that favors traces consistent with fundamental physics and numerically close to targets, with lightweight halting to control compute. We instantiate our framework with a large student model fine-tuned on traces synthesized by a larger teacher model, and evaluate under matched token budgets against various rejection sampling baselines. Our method improves accuracy and calibration, reduces physics-violation rates, and lowers sampling cost relative to baselines. These results indicate that modest, domain-aware constraints combined with trace-level selection provide a practical path toward reliable, efficient LRMs for process-aware property prediction and closed-loop materials design.

Introduction

A central goal in materials discovery is to compress the experimental loop by coupling automated experimentation with algorithmic decision making. Within this loop, property prediction is the core module. Accurate models that map composition, structure, and process/recipe variables to target properties (e.g., materials properties or device-level figures of merit) convert combinatorial exploration into tractable optimization and enable closed-loop design via bayesian optimization, provided that the models expose calibrated uncertainty and lightweight physics constraints to keep proposals physically admissible ([39, 10, 11, 36, 18, 41]). Prior work has largely targeted properties from composition or crystal structure ([4] (Xie & Grossman 2018)) and, more recently, has explored text- or instruction-conditioned surrogates with LLMs (Ndayshimiye et al., 2025); yet scaling these predictors to process-aware, recipe to property tasks at the device level remains challenging ([28, 29] (Xie et al. 2024)).

Recently, large reasoning models (LRMs)—language models trained and/or reinforced to produce reliable reasoning traces—have shown dominant performance in diverse areas such as math, coding and scientific QA ([14] (Jaecht et al. 2024; Yang et al. 2025)). Their step-wise reasoning capabilities are a natural fit for recipe to property prediction, where multi-step physical arguments along the classic chemical composition →\to process →\to micro-structure →\to property chain are essential and well established in integrated computational materials engineering ([32]). Despite the promise, the leveraging LRMs on property prediction task remains underexplored relative to LLM and naive knowledge extraction (Zhao et al., 2024; Ndayshimiye et al., 2025). In this paper, we study how to train LRMs that reason effectively about materials recipes and output numerically correct, physically grounded properties.

A prevailing strategy for training LRMs to reason is to use filtered or re-weighted training signals based on the quality of teacher generated traces, complemented by test-time scaling [31]. Concretely, models are fine-tuned on self-generated rationales kept only when they reach correct outcomes (Zelikman et al., 2022b; Yuan et al., 2023), or on samples ranked by a learned reward/verifier [7] (Dong et al. 2023; Zheng et al. 2023b), or they aggregate multiple samples at decoding [44]. These methods work as standard building blocks for LRMs along with SFT-only pipelines and RL-style post-training [5] (Xu et al. 2025).

We argue that training property prediction LRMs from generated reasoning traces requires more sophisticated, physically grounded rejection sampling. Two characteristics of this task drive the need: (1) High combinatorial design space. The composition, process, structure, property chain creates high-dimensional, multi-mechanism spaces; inverse maps are often non-unique, yielding traces that seem plausible yet are scientifically incorrect [40, 35, 26] (Xiang et al. 1995; Yang et al. 2022). Therefore, effective learning requires sufficient exploration that searches both the design and reasoning trace space. (2) Physically grounded outputs. Targets are physical quantities whose magnitudes are constrained by physics and even small numeric deviations matter. Filters must therefore enforce admissible ranges and physical constraints from conservation laws and constitutive relations rather than rely solely on binary correctness.

Motivated by these challenges, we propose Physics-aware Rejection Sampling (PaRS), a domain-tailored approach to optimize reasoning traces. Unlike prior methods that depend on binary correctness or learned reward models, our method couples rejection sampling with task-native, continuous error metrics derived from wet-lab experiments. Concretely, for each device recipe, we sequentially generate candidate traces, accepting the first trace that satisfies physics-aware acceptance gates and halting sampling early when further candidates show negligible variance or improvement.

We adopt Qwen3-32B2 as the backbone model, fine-tuned via supervised fine-tuning (SFT) on internal prompts, with teacher reasoning traces synthesized from Qwen3-235B1 (Yang et al., 2025). We benchmark against various rejection sampling methods under matched token budgets. Empirically, our method achieves the highest overall accuracy and calibration, while also delivering superior compute efficiency compared to existing baselines.

Our contributions are threefold.

  • We formulate recipe to property prediction as a reasoning task with LRMs where physics-aware verification is essential.

  • We propose novel physically grounded rejection sampling for optimizing reasoning traces, introducing the combination of powerful gating and halting techniques.

  • We conduct (1) a teacher-side ablation, comparing our physics-aware sampler against six baselines in terms of trace accuracy and sampling efficiency, and (2) a student-side evaluation, fine-tuning an open-source LRM to demonstrate consistent gains in accuracy, calibration, and compute efficiency over all baselines.

Related Work

LLMs for materials design Recent advances in large language models (LLMs) have demonstrated strong generalization capabilities in materials design, drawing on interdisciplinary knowledge from chemistry, physics, and engineering [30]. Beyond general-purpose LLMs, domain-specific models such as MatSciBERT [15], MatBERT [42], and MELT [21] have been trained on large-scale materials science corpora, successfully capturing fundamental concepts that link structure, properties, processes, and performance. A key step in the materials design pipeline is accurate property prediction from symbolic representations, which serves as a surrogate for expensive experiments and enables rapid candidate screening. Leveraging their ability to process unstructured scientific data, recent studies have applied LLMs to this task without the need for elaborate feature engineering. For example, LLM-Prop [34] employs LLMs to predict crystalline material properties, while Li et al. [24] integrate LLMs with graph neural networks for improved prediction accuracy. In the context of quantum-dot materials, Choi et al. [6] developed LLM-based synthesis protocol generation and property prediction models, fine-tuned on proprietary synthesis datasets. LLMs have also been explored as surrogate models in optimization frameworks. LLAMBO (Liu et al.) utilizes the exploration capability of LLMs within Bayesian optimization, and BOPRO [2] incorporates LLM-based search strategies that exploit evolving uncertainty estimates to propose promising candidates in each iteration, thereby accelerating the discovery of globally optimal solutions.

Large Reasoning Models (LRMs) The paradigm of next-token prediction has undergone a significant shift with the introduction of "thought" concept—a sequence of intermediate steps representing a model’s internal reasoning process [19, 14, 31] (Yang et al. 2025). This innovative approach enables LLMs to mimic complex human reasoning, such as reflective thinking and tree search. Chain-of-Thought (CoT) prompting [45] initially demonstrated that few-shot rationales could unlock sophisticated reasoning in LLMs, while Self-Consistency (Wang et al.) further enhanced reliability by marginalizing over multiple reasoning paths. Tree-of-Thoughts (ToT) (Yao et al., 2023) later reframed inference as a search over partial thought sequences, incorporating look-ahead and backtracking. Beyond prompting, process supervision has emerged as a powerful technique, training step-level verifiers or reward models to guide the reasoning process. This approach has been shown to outperform models trained with outcome-only labels, particularly in mathematical reasoning tasks [25]. More recently, GPT-o1 [19], Qwen3 (Yang et al., 2025), and DeepSeek-R1 [14] have popularized modern Large Reasoning Models (LRMs) by integrating long-form thinking with process supervision, RL-based post-training, and test-time scaling. Building on these foundational advances, our work adapts these reasoning mechanisms to the domain of materials design, specifically targeting physically grounded recipe to property prediction.

Rejection sampling for LLMs Rejection sampling is widely recognized as an effective data-filtering method that promotes higher-quality supervision. In the context of RLHF and preference-optimization, it narrows multiple generated outputs per prompt to only the high quality responses, as determined by a reward model during post-training adjustments. For example, RAFT (Dong et al.) aligns generative models efficiently by using a reward model and abundant candidate samples; it discards outputs demonstrating undesired behaviors and fine-tunes the model solely on the selected high-quality subset. Building on this, Reinforce-Rej (Xiong et al., 2025) proposes a minimalist policy-gradient extension that filters out both entirely incorrect and entirely correct samples, enhancing stability and efficiency. In reasoning tasks, STAR-like models [16, 22] (Zelikman et al. 2022a) eliminate expensive human annotations by employing a self-taught reasoning loop—generating chain-of-thought traces, self-verifying correctness, and fine-tuning only on reliable examples. Additionally, Rejection Sampling Fine-Tuning (RFT) (Yuan et al., 2024) enhances mathematical reasoning by incorporating model-generated reasoning traces filtered for correctness into training.

Method

Task Definition

Quantum dot light-emitting diodes (QD-LEDs) are electroluminescent devices that use colloidal semiconductor quantum dots as the emissive layer, offering narrowband spectra and composition tunable color [37] (Li et al. 2024). In practice, closed-loop materials design must optimize figures of merit measured at the device level—e.g., peak external quantum efficiency and operational stability—because these metrics ultimately determine application viability for emissive displays and lighting (e.g., televisions, monitors) [3, 17] (Li et al. 2024). We therefore frame inputs as complete device recipes: a multi-layer stack typically consist of anode/ITO, hole-injection layer (HIL), hole-transport layer (HTL), quantum-dot emitting layer (EML), electron-transport layer (ETL), electron-injection layer/cathode with per layer materials and process parameters. Each layer records identifiers (material, formulation), geometric parameters (e.g., thickness), and process variables (e.g., solution concentration, spin profile, bake/anneal temperature and duration, atmosphere), along with post-process steps (e.g., UV–ozone, plasma, solvent rinse). See Figure 1 for a recipe example.

Image not extracted

Figure 1. Structured QD-LED recipe example.

We formulate QD-LED device property prediction as a reasoning LLM task. Let D={(xi,yi)}i=1N\mathcal{D} = \{(x_i, y_i)\}_{i=1}^{N} where xix_i is a QD-LED recipe as above and yi∈Ry_i \in\mathbb{R} is a device-level target; in this work we focus on max external quantum efficiency (i.e., yi=EQEmax⁡y_i = \mathrm{EQE}_{\max}). Given a task-specific prompt (See Figure 2) and recipe xix_i, a large reasoning model fθf_\theta produces a reasoning trace τ\tau along with a numeric prediction y^\hat{y}:

Image not extracted

Figure 2. Prompt for the property prediction task with large reasoning models (LRMs).

fθ(prompt,xi)→(τ,y^).(1)f_\theta(\mathrm{prompt}, x_i) \rightarrow(\tau, \hat{y}). \tag*{(1)}

PaRS: Physics-aware Rejection Sampling

For training LRMs, the supervision signal extends beyond the final answer to include the sampled reasoning traces themselves. Achieving high-quality supervision for reasoning LLMs therefore requires rejection sampling to filter out suboptimal traces. As illustrated in Figure 3, we propose Physics-aware Rejection Sampling (PaRS), which integrates physics-aware gating and halting mechanisms to optimize reasoning traces

Image not extracted

Figure 3. PaRS workflow: the teacher generates a mini-batch of candidates; a candidate is accepted only if it passes gates (range, near-truth tolerance, physics envelope). If none pass, halting checks decide whether to stop or raise temperature and continue to next sampling round. Accepted traces supervise the student model as training data.

Physics-aware Gating For each recipe xix_i, we generate reasoning traces sequentially from the teacher model up to KK. We accept the first sample that satisfies all of following acceptance gates:

y^(k)∈[0,100],(2)\hat{y}^{(k)} \in[0, 100], \tag*{(2)}
∣y^(k)−yi∣≤εMAE,(3)\left|\hat{y}^{(k)} - y_i\right| \le\varepsilon_{\mathrm{MAE}}, \tag*{(3)}

As EQEmax⁡\mathrm{EQE}_{\max} is reported as a percentage, Eq. (2) enforces range consistency. Unlike categorical correctness filters ([44, 31] (Zelikman et al. 2022b; Yuan et al. 2023; Dong et al. 2023)), Eq. (3) uses a continuous error against wet-lab ground truth, yielding richer learning signals.

With external quantum efficiency (EQE) as the prediction target, we can define an empirical upper bound U(x)U(x) as follows. EQE is commonly factorized as EQE=ηout⋅ηrad⋅γ\mathrm{EQE} = \eta_{\mathrm{out}} \cdot\eta_{\mathrm{rad}} \cdot\gamma with ηout≤1\eta_{\mathrm{out}} \le1 and γ≤1\gamma\le1 by definition. In regimes where solid-state PLQY limits the radiative yield, the conservative relation ηrad≤PLQY\eta_{\mathrm{rad}} \le\mathrm{PLQY} holds, implying EQE≤PLQY\mathrm{EQE} \le\mathrm{PLQY} ([38, 12] (OE2 2014)). PLQY (photoluminescence quantum yield) denotes the ratio of emitted to absorbed photons; we use the solid-state PLQY of the EML under device-relevant conditions. We therefore define an empirical, recipe-specific upper bound as U(x)=UPL(x)U(x)=U_{\mathrm{PL}}(x), where UPL(x)U_{\mathrm{PL}}(x) is instantiated by the highest measured film photoluminescence quantum yield of the emissive layer in the recipe xx. We add this upper bound to the acceptance gates:

y^(k)≤U(xi)(4)\hat{y}^{(k)} \le U(x_i) \tag*{(4)}

It prevents reasoning traces that predict physically implausible overshoots for target property. If no sampled candidate satisfies Eqs. (2)–(4), we discard the example.

Adaptive Halting We sample in mini-batches of size bb and proceed round by round until the total budget Kmax⁡K_{\max} is exhausted. In round rr (r=1,2,…r = 1, 2, \ldots), we draw exactly bb candidates {(τ(r,j),y^(r,j))}j=1b\{(\tau^{(r,j)}, \hat{y}^{(r,j)})\}_{j=1}^{b} at temperature TrT_r. For each candidate, we apply the acceptance gates in Eqs. (2)–(4). If any candidate passes, we accept the earliest passing one and terminate sampling.

If no candidate in the mini-batch passes, we apply two halting checks before proceeding. (i) Variance-based halting (from round 1): stop when the within-batch error variance falls below a threshold, indicating insufficient diversity to justify further exploration. (ii) Improvement-based halting (from round 2): stop when the best error in the current round fails to improve over the previous round by at least a small margin. We also halt once the cumulative number of sampled candidates reaches Kmax⁡K_{\max}. If none of these conditions trigger, we increase TrT_r and continue to the next round to encourage exploration. Refer Appendix § 7.§ 1 for the details of halting methods.

Experiments

Baselines

We curate 11k QD-LED device dataset and split into 10k for training and 1k for testing. We construct prompts from all 11k dataset for a property prediction task and query 10k train prompts with QWEN3-235B to sample teacher reasoning traces. To ensure fair comparison, each reasoning trace has a same sampling budget of K=12K=12, except for no sampling. We compare the following methods for selecting reasoning traces including our method.

  1. No sampling: use the first generated trace.

  2. Random sampling: uniformly sample one of the KK traces.

  3. Self-consistency aggregation: select the trace whose final answer is closest to the median across all KK answers [44].

  4. Longest trace: select the trace with the largest token length.

  5. LLM-as-a-judge: score all traces with a larger judge model (DeepSeek-R1 {sup}‡`3) and select the top-ranked trace {citep}(Zheng et al., 2023a). Refer {xref}Appendix A.2` for details of the LLM-as-a-Judge prompting.

  6. Multi-sampling: retain all KK traces as supervision [13].

  7. PaRS (ours): mini-batch size b=4b=4 with a temperature schedule T∈{0.6,0.8,1.0}T \in\{0.6, 0.8, 1.0\} increasing by 0.2 per round. We set εMAE=1\varepsilon_{\mathrm{MAE}}=1, εvar=1\varepsilon_{\mathrm{var}}=1 and δimp=1\delta_{\mathrm{imp}}=1 by analyzing the data distribution. If no candidate passes the gates within the budget, the example is discarded.

After rejection sampling, we fine-tune QWEN3-32B as the student model for a single epoch on the traces selected by each method, using AdamW (learning rate 2×10−52 \times10^{-5}) on 32 × A100 (80 GB) GPUs; unless stated otherwise, all other training hyperparameters are shared across methods.

Evaluation metrics We report two groups of metrics. (1) Teacher-side trace quality: mean absolute error of the selected trace’ prediction, and the average number of sampled traces per prompt to quantify the cost of constructing the selected traces. To further assess trace quality, we also employ a larger LLM (DeepSeek-R1) as an external judge, providing external evaluation of reasoning quality on a 0–10 scale. (2) Student-side performance: MAE, R2R^2, Spearman’s ρ\rho, and a physics violation rate—the fraction of predictions that fall outside [0,100][0, 100] or exceed the empirical upper bound U(x)U(x) on the hold out test set.

For each test prompt, we run five independent inferences with the trained student models and take the median of the five predictions before computing MAE, R2R^2, and Spearman’s ρ\rho. For the physics violation rate, we evaluate the constraint indicator on each of the five predictions and report the average violation frequency across all ensemble. Details of the LLM-as-a-Judge prompting procedure are provided in Appendix § 7.§ 2.

Results

Our experiments show that PaRS effectively optimize the teacher’ reasoning traces to induce reasoning capability for property prediction in student model. We present our findings in two parts: first, an analysis of the trace selection process itself and second, an evaluation of the trained student LRMs on the property prediction task.

Analysis of reasoning trace sampling

We evaluate rejection sampling strategies on the traces generated by the teacher QWEN3-235B. As summarized in Table 1, PaRS achieves lower prediction error while requiring fewer generations on average, placing it on the empirical quality–efficiency Pareto front (Figure 4). A simple MAE gate could make low error appear trivial, yet an external LLM-as-a-judge that is not optimized for our metric still assigns

MethodKavgK_{\mathrm{avg}}Budget (→ selected)LLM-as-a-Judge (score, 0–10)MAE
No sampling11 → 15.972.440
Random sampling1212 → 15.862.327
Longest trace1212 → 16.102.274
Self-consistency1212 → 15.781.829
LLM-as-a-Judge1212 → 16.552.223
Multi-sampling1212 → 125.892.356
PaRS (Ours)6.412 → 0.8†7.510.829

Table 1. Analysis of reasoning trace selection. All traces are generated with QWEN3-235B. Higher LLM-as-a-Judge score and lower MAE is better. KavgK_{\mathrm{avg}} is the average number of generated reasoning traces per prompt. Our Halting logic yields Kavg=6.4K_{\mathrm{avg}}=6.4 on average (fewer required tokens for selected trace). ‘Budget→selected” counts teacher generates per prompt and the number retained after selection.

Image not extracted

Figure 4. Compute–accuracy frontier for rejection sampling methods. Our approach achieves the lowest teacher MAE with substantially fewer required tokens, forming the empirical Pareto front. The x-axis shows average required tokens for generating reasoning trace per prompt and the y-axis shows teacher MAE. See Appendix § 7.§ 3 for details.

Table 4: Distillation results. Each row fine tunes the same student (QWEN3-32B) on selected traces generated by the indicated teacher models with PaRS.

Distiled models# trainMAER2R^2Spearman ρ\rhoViol. (%)
DeepSeek-R1-Distill-Qwen-32B78971.7550.4450.71725.1
Qwen3-235B-Distill-Qwen-32B79801.8080.4240.70527.7
Qwen3-235B-FP8-Distill-Qwen-32B81121.8010.4130.71228.9

Table 4.

PaRS the highest overall score. Self-consistency, which got the lowest MAE except for our method, receives the weakest judge score among the baselines. This gap suggests that generic preference signals do not fully align with physics-grounded proximity to ground truth. By combining physics-aware gates with temperature scheduling and early stopping, PaRS concentrates supervision on high-fidelity, physically admissible traces rather than on merely “good-looking” rationales.

Evaluation of distilled LRMs Training the student QWEN3-32B on traces selected by our method yields consistent gains in accuracy, correlation with ground truth, and physical admissibility, while using substantially fewer supervision traces than competing approaches (See Table 2). The LLM-as-a-Judge baseline is competitive in error and correlation but does not match the reduction in violations or the calibration gains achieved by PaRS. Retaining all sampled traces enlarges supervision volume yet underperforms, consistent with amplified label and reasoning noise when traces remain unfiltered. Recent work [13] finds that multi-sampling can help by preserving reasoning diversity, but in our task the same unfiltered diversity amplifies trace noise, increases physics violation rate and weakens calibration. PaRS reconciles these views by keeping diversity where it matters, namely multiple near correct physical pathways, while trimming supervision to numerically consistent traces that respect simple physics. The result is a better accuracy and admissibility with about 15× fewer traces than the multi sampling baseline.

Training data (prompt + reasoning trace)# trainMAER2R^2Spearman ρ\rhoViol. (%)
No sampling10 0002.0010.3760.60735.8
Random sampling10 0001.9610.3580.62135.2
Longest trace10 0001.9420.3750.61435.3
Self-consistency10 0001.9330.3770.62936.8
LLM-as-a-Judge10 0001.8890.4080.66735.4
Multi-sampling120 0001.9840.3350.63236.6
PaRS (ours)79801.8080.4240.70527.7

Table 2. QD-LED property prediction with QWEN3-32B trained on teacher traces selected by each method from QWEN3-235B. Lower MAE is better; higher R2R^2 and Spearman ρ\rho are better. Viol.% is computed on test predictions without post hoc clipping to U(x)U(x). # train is the number of supervision traces used for SFT: multi-sampling retains all K=12K=12 traces per example (12× number of train prompts), whereas our method keeps only candidates that pass our gates, resulting in fewer traces.

These improvements arise from aligning the acceptance rubric with a continuous, physically grounded target. The range check in Eq. (2) and the empirical envelope in Eq. (4) for target EQE suppress implausible overshoots. The continuous gate in Eq. (3) rewards numerical proximity to wet-lab measurements rather than binary correctness. Since the mapping from recipe to property is many to one and complex, enforcing zero error steers supervision toward outliers and collapses trace diversity. The tolerance εMAE\varepsilon_{\mathrm{MAE}} instead admits near-correct and physically plausible traces, which preserves multiple valid pathways, improves calibration, and lowers violation rates. The halting logic further reduces redundant sampling and concentrates supervision on diverse, high-fidelity trajectories.

Analysis

Robustness of PaRS across teacher models To validate robustness of our method across the diverse teacher models, we sample reasoning traces with DeepSeek-R1-671B and Qwen3-235B-FP8 in addition to Qwen3-235B. MAE for all traces in Table 3 shows that smaller or low-precision teacher models produce noisier trace distributions than large models. However, after applying PaRS, the selected traces from small quantized teachers converge to the quality of those from large teachers: error and stability tighten, and the retained supervision is physically admissible and consistently high quality. In effect, PaRS equalizes teacher induced variability by filtering out unstable generations and preserving only high fidelity reasoning trajectories.

Table 3: Comparison of reasoning traces generate by three teacher models. We calculate MAE for all traces by using all generated traces (seven traces per prompt in average). MAE from selected traces is computed only over selected traces by PaRS. Selected ratio is the fraction of generated traces accepted by PaRS.

Teacher modelMAE for all tracesMAE for selected tracesSelected ratio
DeepSeek-R1-671B2.1440.7850.789
Qwen3-235B2.3270.8290.798
Qwen3-235B-FP89.0780.8120.811

Table 3.

This equalization propagates to the student. Fine-tuning the same student on selected traces from different teachers yields similar MAE, correlation, and violation rates. The correctness ratio of selected traces also varies little across teachers, indicating that PaRS is broadly compatible with heterogeneous teacher models and precisions. Consequently, even smaller or low-precision teachers can provide supervision competitive with that from larger models.

How trace correctness ratio shapes distillation performance. We study how the fraction of accepted traces (the “correctness ratio” under the tolerance gate εMAE\varepsilon_{\mathrm{MAE}}) relates to student performance (Figure 5). Holding the dataset size and token budget fixed, increasing the acceptance rate rr monotonically lowers MAE and reduces physics violations. Interestingly, recent work [31] shows that curated rationales and test-time scaling can yield strong results even when only 53% of traces are correct. We hypothesize that tasks such as math, coding, and QA exhibit binary correctness with explicit derivations, so trace correctness closely mirrors supervision quality. However, for property prediction, categorical correctness is a weak proxy for supervision quality because two traces may follow different discrete steps yet end within 1% of the true EQE, which is the signal the model must learn.

Image not extracted

Figure 5. Effect of training correctness ratio on student performance. Training dataset size and inference token budget are fixed. Points show means over a 5-model ensemble on test prompts.

Ablation on adaptive halting. We ablate the adaptive-halting mechanism in PaRS to assess the benefit of early stopping. As summarized in Table 5, enabling adaptive halting reduces the average number of sampled candidates from Kavg=8.9K_{\mathrm{avg}}=8.9 to 6.4 and shortens compute time from 20.5,h to 15.1,h, with only a marginal change in accuracy. These results indicate that the variance- and improvement-based early stopping criteria are effective, low-overhead heuristics for pruning unproductive samples without materially degrading the quality of the retained traces. As the teacher model’s parameters are fixed, persistently low variance and minimal improvement suggest limited capacity to produce faithful reasoning traces that recover ground-truth properties with physically plausible values.

MethodKavgK_{\mathrm{avg}}MAECompute time (h)
PaRS w/o Adaptive Halting8.90.81720.5
PaRS w/ Adaptive Halting6.40.82915.1

Table 5. Ablation on adaptive halting. We use 10 nodes of 8×8 \times H100 (80G) to sample reasoning traces from the 10k prompts with Qwen3-235B and measure the compute time.

Conclusion

We cast recipe to property prediction for materials discovery as a reasoning problem and introduce Physics-aware Rejection Sampling (PaRS) to curate supervision signals that are numerically accurate, calibrated, and physically admissible. PaRS replaces binary correctness and generic reward models with domain-grounded gates—range checks, a recipe-specific physical envelope, and a continuous error tolerance—and adds variance and improvement-based halting. Instantiated with a Qwen3-235B teacher and a Qwen3-32B student on QD-LED device recipes, PaRS consistently optimizes teacher-trace quality and yields student LRMs with lower MAE, higher correlation, better calibration, and markedly fewer physics violations. These gains are achieved with substantially less sampling, indicating a favorable quality–efficiency trade-off. While our experiments focus on EQE in QD-LEDs, the framework naturally extends to other device-level properties (e.g., lifetime, luminance) and to materials systems beyond optoelectronics. An important next step is to evaluate PaRS at larger scale and across diverse domains. We also envision coupling PaRS with RL-based adaptive exploration to enable closed-loop recipe design, where models not only predict reliably but also guide autonomous exploration of materials space under explicit physical guarantees.

Implementation details

Sampling schedule

For an example (xi,yi)(x_i,y_i) and rounds r=1,2,…r = 1,2,\ldots with mini-batch indices j=1,…,bj = 1,\ldots,b, draw candidates (τ(r,j),y^(r,j))(\tau^{(r,j)},\hat{y}^{(r,j)}) and define the per-candidate error er,j:=∣y^(r,j)−yi∣e_{r,j} := |\hat{y}^{(r,j)} - y_i|. For each round, let eˉr:=b−1∑j=1ber,j\bar{e}_r := b^{-1}\sum_{j=1}^{b}e_{r,j}, sr2:=(b−1)−1∑j=1b(er,j−eˉr)2s_r^2 := (b-1)^{-1}\sum_{j=1}^{b}(e_{r,j}-\bar{e}_r)^2, and er∗:=min⁡1≤j≤ber,je_r^* := \min_{1\le j\le b}e_{r,j}; for r≥2r \ge2, define the improvement Δr:=er−1∗−er∗\Delta_r := e_{r-1}^* - e_r^*. The total candidate budget is Kmax⁡K_{\max}, and TrT_r denotes the sampling temperature in round rr.

A candidate is accepted if it satisfies all gates in Eqs. (2)–(4), namely y^(r,j)∈[0,100]\hat{y}^{(r,j)} \in[0,100], ∣y^(r,j)−yi∣≤ϵMAE|\hat{y}^{(r,j)}-y_i| \le\epsilon_{\mathrm{MAE}}, and y^(r,j)≤U(xi)\hat{y}^{(r,j)} \le U(x_i). If multiple candidates pass within the same round, accept the one with the smallest jj and terminate for that example.

If no candidate is accepted in round rr, the procedure halts early under any of the following conditions: variance small enough, sr2≤ϵvars_r^2 \le\epsilon_{\mathrm{var}} (available from r=1r = 1); lack of improvement, Δr≤δimp\Delta_r \le\delta_{\mathrm{imp}} (available from r=2r = 2); or budget exhausted, rb≥Kmax⁡rb \ge K_{\max}.

The temperature follows a capped, monotone increase to encourage exploration, for example Tr+1=min⁡{Tmax⁡,γTr}T_{r+1} = \min\{T_{\max},\gamma T_r\} with γ>1\gamma> 1, or Tr+1=min⁡{Tmax⁡,Tr+ΔT}T_{r+1} = \min\{T_{\max},T_r+\Delta T\} with ΔT>0\Delta T > 0, starting from T1=Tmin⁡T_1 = T_{\min}.

The overall procedure is: at round r=1r = 1, sample bb candidates at temperature TrT_r and apply the acceptance gates; if none pass, compute sr2s_r^2, er∗e_r^*, and (for r≥2r \ge2) Δr\Delta_r; if no halting condition triggers, update the temperature according to the schedule and continue to r+1r+1. Discard the example if no candidate is accepted before the budget is spent.

Prompt for LLM-as-a-Judge

We use DEEPSEEK-R1-0528-671B as the LLM-as-a-Judge. The prompt in Figure 6 instructs the judge to evaluate synthesized reasoning traces against five rubrics and to return a numeric score for each rubric on a 0–10 scale with a brief justification. For the LLM-as-a-Judge selection baseline, we score all candidate traces generated for a prompt, compute a composite judge score by averaging the five rubric scores, and select the best one per prompt. For the summary metric reported in Table 1, we evaluate the set of traces selected by each method with the same judge prompt. We compute the composite score for each trace as the mean over the five rubrics and then report the mean across all evaluated prompts. This yields a single 0–10 score per method that is comparable across rejection sampling methods.

Image not extracted

Figure 6. Prompt for evaluating reasoning traces with DeepSeek-R1. We report the sum of average score of the five metrics to Table 1.

Token accounting for trace selection

For each prompt, let Tteach,inT_{\mathrm{teach,in}} and Tteach,outT_{\mathrm{teach,out}} denote the teacher’s input tokens and average output tokens per generated trace. Let KK be the generation budget, GG the random number of traces actually generated before acceptance or budget exhaustion, and Kavg:=E[G]K_{\mathrm{avg}} := \mathbb{E}[G]. Let TselectT_{\mathrm{select}} capture any extra token cost due to a selection pass (if present). Finally, let racc∈[0,1]r_{\mathrm{acc}} \in[0,1] be the probability that a prompt yields at least one accepted trace.

Expected tokens per prompt The expected token cost per prompt, regardless of whether a trace is accepted, is

E[tokens per prompt]=Kavg(Tteach,in+Tteach,out)+Tselect.(5)\mathbb{E}[\text{tokens per prompt}] = K_{\mathrm{avg}}\left(T_{\mathrm{teach,in}} + T_{\mathrm{teach,out}}\right) + T_{\mathrm{select}.} \tag*{(5)}

For offline selection methods (random, self-consistency, and ours), the selection pass is negligible, so Tselect≈0T_{\mathrm{select}} \approx0. By contrast, LLM-as-a-Judge performs an additional inference pass over the concatenated set of generated traces. Approximating the judge pass as comparable in length to the teacher pass yields

E[tokens per prompt]judge≈2Kavg(Tteach,in+Tteach,out).(6)\mathbb{E}[\text{tokens per prompt}]_{\mathrm{judge}} \approx2K_{\mathrm{avg}}\left(T_{\mathrm{teach,in}} + T_{\mathrm{teach,out}}\right). \tag*{(6)}

Expected tokens per accepted trace When some prompts produce no accepted trace, it is useful to normalize by the acceptance rate raccr_{\mathrm{acc}}. The expected tokens per accepted trace are

E[tokens per accepted]=E[tokens per prompt]E[accepted traces per prompt]≈Kavg(Tteach,in+Tteach,out)+Tselectracc,(7)\mathbb{E}[\text{tokens per accepted}] = \frac{\mathbb{E}[\text{tokens per prompt}]}{\mathbb{E}[\text{accepted traces per prompt}]} \approx\frac{K_{\mathrm{avg}}\left(T_{\mathrm{teach,in}} + T_{\mathrm{teach,out}}\right) + T_{\mathrm{select}}}{r_{\mathrm{acc}}}, \tag*{(7)}

with the judge variant obtained by substituting Eq. (6) into Eq. (7).

Under online acceptance (our method), generation halts immediately upon acceptance or when the budget KK is reached. Thus GG follows a truncated geometric-like process, and KavgK_{\mathrm{avg}} reflects both early acceptance on easy prompts and full-budget usage on hard prompts. Methods that commit to a fixed KK without early stopping have Kavg≈KK_{\mathrm{avg}} \approx K.

In compute versus accuracy frontiers (e.g., Figure 4), the x-axis reports the per-prompt token cost in Eq. (5). When comparing methods with materially different raccr_{\mathrm{acc}}, we additionally report the per-accepted-trace cost using Eq. (7).

As a concrete example, suppose Tteach,in=900T_{\mathrm{teach,in}}=900, Tteach,out=2000T_{\mathrm{teach,out}}=2000, Kavg=6.4K_{\mathrm{avg}}=6.4, and racc=0.8r_{\mathrm{acc}}=0.8. Then the per-prompt cost for offline selection is 6.4×(900+2000)=18,5606.4 \times(900 + 2000) = 18{,}560 tokens. The judge variant is about 2×18,560=37,1202 \times18{,}560 = 37{,}120 tokens per prompt. Normalizing by acceptance rate, the per-accepted-trace costs are approximately 23,200 (offline) and 46,400 (judge).

References

  1. [2]Dhruv Agarwal, Manoj Ghuhan Arivazhagan, Rajarshi Das, Sandesh Swamy, Sopan Khosla, and Rashmi Gangadharaiah. Searching for optimal solutions with LLMs via bayesian optimization. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=aVfdr7lXdV. 3
  2. [3]Sang Yun Bang, Yo-Han Suh, Xiang-Bing Fan, Dong-Wook Shin, Sanghyo Lee, Hyung Woo Choi, Tae Hoon Lee, Jiajie Yang, Shijie Zhan, William Harden-Chaters, Chatura Samarakoon, Luigi G. Occhipinti, Soo Deok Han, Sung-Min Jung, and Jong Min Kim. Technology progress on quantum dot light-emitting diodes for next-generation displays. Nanoscale Horizons, 6:68–77, 2021. doi: 10.1039/D0NH00556H. 3DOI
  3. [4]Chi Chen, Weike Ye, Yifei Zuo, Chen Zheng, and Shyue Ping Ong. Graph networks as a universal machine learning framework for molecules and crystals. Chemistry of Materials, 31(9):3564–3572, 2019. doi: 10.1021/acs.chemmater.9b01294. 1DOI
  4. [5]First name initial. Chen et al. Towards training large reasoning models. arXiv preprint, 2025. Please replace with the exact arXiv ID when finalized. 2
  5. [6]So Eun Choi, Miyoung Jang, SoHee Yoon, SangHyun Yoo, Jooyeon Ahn, Minho Kim, Ho-Gyeong Kim, Yebin Jung, Seongeon Park, Young-Seok Kim, et al. Llm-driven synthesis planning for quantum dot materials development. Journal of Chemical Information and Modeling, 65(6):2748–2758, 2025. 3DOI
  6. [7]Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. URL https://arxiv.org/abs/2110.14168. Introduces the GSM8K dataset. 2
  7. [10]Manuel M. Flores-Leonar, Giovanni F. Caramori, Nancy A. Romero, et al. Self-driving laboratories for chemistry and materials science. Chemical Reviews, 124(14):10137–10190, 2024. doi: 10.1021/acs.chemrev.3c00430. URL https://doi.org/10.1021/acs.chemrev.3c00430. 1
  8. [11]Peter I. Frazier. A tutorial on bayesian optimization. arXiv preprint arXiv:1807.02811, 2018. URL https://arxiv.org/abs/1807.02811. 1
  9. [12]Malte C. Gather and Sebastian Reineke. Recent advances in light outcoupling from white organic light-emitting diodes. Journal of Photonics for Energy, 5:057607, 2015. doi: 10.1117/1.JPE.5.057607. 5DOI
  10. [13]Tanmay Guha et al. Openthoughts: A scaling recipe for test-time compute in reasoning. arXiv preprint arXiv:2506.04178, 2025. URL https://arxiv.org/abs/2506.04178. 6, 8
  11. [14]Quanlong Guo et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. URL https://arxiv.org/abs/2501.12948. 1, 3
  12. [15]Tanisha Gupta, Mohd Zaki, NM tAnoop Krishnan, and Mausam. Matscibert: A materials domain language model for text mining and information extraction. npj Computational Materials, 8(1): 102, 2022. 2arxiv.org/abs/2109.15290
  13. [16]Arian Hosseini, Xingdi Yuan, Nikolay Malkin, Aaron C Courville, Alessandro Sordoni, and Rishabh Agarwal. V-star: Training verifiers for self-taught reasoners. CoRR, 2024. 3arxiv.org/abs/2402.06457
  14. [17]Yu-Ming Huang, Konthoujam James Singh, An-Chen Liu, Chien-Chung Lin, Zhong Chen, Kai Wang, Yue Lin, Zhaojun Liu, Tingzhu Wu, and Hao-Chung Kuo. Advances in quantum-dot-based displays. Nanomaterials, 10(7):1327, 2020. doi: 10.3390/nano10071327. 3DOI
  15. [18]Ryan Jacobs, Logan Ward, Anubhav Jain, et al. Accurate predictions, uncertainty estimates, and accessibility: A roadmap for machine-learning-enabled materials discovery. Machine Learning: Science and Technology, 5(4):042501, 2024. doi: 10.1088/2632-2153/ad77aa. URL https://iopscience.iop.org/article/10.1088/2632-2153/ad77aa. 1
  16. [19]Aaron Jaech, William Cohen, et al. The ol system: Towards reasoning in large language models. arXiv preprint arXiv:2412.16720, 2024. URL https://arxiv.org/abs/2412.16720. 1, 3
  17. [21]Junho Kim, Yeachan Kim, Jun-Hyung Park, Yerim Oh, Suho Kim, and SangKeun Lee. Melt: Materials-aware continued pre-training for language model adaptation to materials science. In Findings of the Association for Computational Linguistics: EMNLP 2024, pp. 10690–10703, 2024. 2
  18. [22]Woosung Koh, Wonbeen Oh, Jaein Jang, MinHyung Lee, Hyeongjin Kim, Ah Yeon Kim, Joonkee Kim, Junghyun Lee, Taehyeon Kim, and Se-Young Yun. Adastar: Adaptive data sampling for training self-taught reasoners. arXiv preprint arXiv:2505.16322, 2025. 3arxiv.org/abs/2505.16322
  19. [24]Youjia Li, Vishu Gupta, Muhammed Nur Talha Kilic, Kamal Choudhary, Daniel Wines, Wei-keng Liao, Alok Choudhary, and Ankit Agrawal. Hybrid-llm-gnn: integrating large language models and graph neural networks for enhanced materials property prediction. Digital Discovery, 4(2): 376–383, 2025. 3
  20. [25]Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. In The Twelfth International Conference on Learning Representations, 2023. 3arxiv.org/abs/2305.20050
  21. [26]Dianjing Liu, Yixuan Tan, Erfan Khoram, and Zongfu Yu. Training deep neural networks for the inverse design of nanophotonic structures. ACS Photonics, 5(4):1365–1369, 2018. doi: 10.1021/acsphotonics.7b01377. 2
  22. [28]Zhe Liu, Nicholas Rolston, Austin C. Flick, Thomas W. Colburn, Zekun Ren, Reinhold H. Dauskardt, and Tonio Buonassisi. Machine learning with knowledge constraints for process optimization of open-air perovskite solar cell manufacturing. Joule, 6(4):834–849, 2022. doi: 10.1016/j.joule.2022.03.003. 1
  23. [29]Yao Lu, Dong Wei, Wu Liu, Juan Meng, Xiaomin Huo, Yu Zhang, Zhiqin Liang, Bo Qiao, Suling Zhao, Dandan Song, and Zheng Xu. Predicting the device performance of perovskite solar cells from the experimental parameters through machine learning of existing experimental results. Journal of Energy Chemistry, 77:200–208, 2023. doi: 10.1016/j.jechem.2022.10.024. 1
  24. [30]Santiago Miret and NM Anoop Krishnan. Are Illms ready for real-world materials discovery? In AI for Accelerated Materials Design-Vienna 2024. 2arxiv.org/abs/2402.05200
  25. [31]Niklas Muennighoff et al. S1: Simple test-time scaling. arXiv preprint arXiv:2501.19393, 2025. URL https://arxiv.org/abs/2501.19393. 2, 3, 5, 9
  26. [32]National Research Council. Integrated Computational Materials Engineering: A Transformational Discipline for Improved Competitiveness and National Security. National Academies Press, Washington, DC, 2008. doi: 10.17226/12199. URL https://nap.nationalacademies.org/catalog/12199. 1DOI
  27. [34]Andre Niyongabo Rubungo, Craig Arnold, Barry P Rand, and Adji Bousso Dieng. Llm-prop: predicting the properties of crystalline materials using large language models. npj Computational Materials, 11(1):186, 2025. 3DOI
  28. [35]Ben Ren et al. Inverse deep learning methods and benchmarks for artificial electromagnetic material design. arXiv preprint arXiv:2112.10254, 2021. URL https://arxiv.org/abs/2112.10254. 2
  29. [36]Elif Sabanzagil et al. Multi-fidelity bayesian data-driven design of energy absorbing spinoid cellular structures. arXiv preprint arXiv:2507.22079, 2025. URL https://arxiv.org/abs/2507.22079.
  30. [37]Yasuhiro Shirasaki, Geoffrey J. Supran, Moungi G. Bawendi, and Vladimir Bulović. Emergence of colloidal quantum-dot light-emitting technologies. Nature Photonics, 7:13–23, 2013a. doi: 10.1038/nphoton.2012.328. 3DOI
  31. [38]Yasuhiro Shirasaki, Geoffrey J. Supran, Moungi G. Bawendi, and Vladimir Bulović. Emergence of colloidal quantum-dot light-emitting technologies. Nature Photonics, 7(1):13–23, 2013b. doi: 10.1038/nphoton.2012.328. 5DOI
  32. [39]Eric A. Stach, Brian L. DeCost, A. Gilad Kusne, Jason Hattrick-Simpers, Keith A. Brown, Katia G. Reyes, Joshua Schrier, Simon J. L. Billinge, Tonio Buonassisi, and Ian Foster. The rise of self-driving laboratories in chemical and materials sciences. Nature Synthesis, 1:774–786, 2022. URL https://www.nature.com/articles/s44160-022-00149-2. 1
  33. [40]Ichiro Takeuchi, Robert Bruce van Dover, and Hideomi Koinuma. Combinatorial synthesis and evaluation of functional inorganic materials using thin-film techniques. MRS Bulletin, 27(4):301–308, 2002. doi: 10.1557/mrs2002.97. 2DOI
  34. [41]Vladimir Varivoda, Kunpeng Dong, Shahrzad Omeе, Zhenpeng Hu, et al. Materials property prediction with uncertainty quantification: A benchmark study. arXiv preprint arXiv:2211.02235, 2022. URL https://arxiv.org/abs/2211.02235. 1
  35. [42]Yuwei Wan, Tong Xie, Nan Wu, Wenjie Zhang, Ch unyu Kit, and Bram Hoex. From tokens to materials: Leveraging language models for scientific discovery. arXiv preprint arXiv:2410.16165, 2024. 2arxiv.org/abs/2410.16165
  36. [44]Xuezhi Wang, Jason Wei, Dale Schuurmans, et al. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022. URL https://arxiv.org/abs/2203.11171. 2, 5, 6
  37. [45]Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022. 3arxiv.org/abs/2201.11903

Paper details

vlnc:W_z2rc6ha4
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 6

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents