PharmAgent: Constraint-Aware Search with Frozen Language Models for Molecular Optimization

Nihui Shao; Guanxing Chen; Jilong Shi; Zhengyang Bai; Haohuai He; Zhenchao Tang; Qiujie Lv; Yu-An Huang

cs.AIOct 4, 2026

PharmAgent: Constraint-Aware Search with Frozen Language Models for Molecular Optimization

Authors: Nihui Shao, Guanxing Chen, Jilong Shi, Zhengyang Bai, Haohuai He, Zhenchao Tang, Qiujie Lv, Yu-An Huang, +1 more

Organizations: City University of Hong Kong (Dongguan) · The Hong Kong Polytechnic University · Tencent AI Lab · Zhengzhou University · Northwestern Polytechnical University

Abstract

Molecular optimization must improve target activity and satisfy developability constraints within limited evaluation budgets. Classical methods require tailored rules or training to incorporate chemical instructions and property feedback. Frozen language models can condition edits on this information, but need explicit constraint control and relevant experience. We therefore present PharmAgent, a constraint-aware molecular search method driven by adaptive external state. Its Lagrangian controller translates violations in accepted states into accumulated constraint pressure, keeping this history separate from current property measurements. Structure-indexed replay complements this feedback with relevant evaluated transitions that guide subsequent proposals. As a curriculum progressively activates constraints, candidates and the incumbent are compared under the same current objective, and the accepted state determines the next multiplier update. We derive an exact identity that characterizes how accepted-state violations accumulate in the controller's multipliers. Across five tasks with five independent runs, PharmAgent achieves a summed area under the target-score curves (AUC) of 3.9208 in target-only search, improving over MOLLEO by 37.3%. With online constraints, it achieves a property-adjusted AUC of 0.7076, improving over the strongest online baseline, ExLLM, by 53.8%. These results rank first among all evaluated methods in both target-only and constraint-aware search. The online comparison covers all five baseline frameworks. The full system leads every ablation variant in target quality, property-adjusted performance, and Pareto hypervolume. All five molecular cases reach feasible final states, documenting target gains and trade-offs.

Figures & tables

Method	Online constraints	Adaptive weights	Structure-based retrieval	Cross-task reuse
Graph-GA / REINVENT4	✗	✗	✗	✗
MOLLEO / ExLLM / SEISMO	✗	✗	✗	✗
Online adaptations	✓	✗	✗	✗
PharmAgent	✓	✓	✓	✓

Table 1: Mechanisms in evaluated configurations (✓ enabled, ✗ not enabled). Baseline rows show original configurations. Online adaptations use fixed-penalty (FP) rewards or feasibility-first (FF) rewards, selection, or conversational guidance. Retrieval and cross-task reuse concern optimization attempts. Appendix D specifies protocols.

Figure 1: Motivation and the PharmAgent feedback loop. (a) Two feedback gaps in molecular search. High target scores need not imply feasibility, and past evaluations require relevant retrieval. (b) Frozen Generator–Reviewer workflow for proposal, evaluation, acceptance, and feedback. Candidates and the incumbent share the current objective. (c) External search state comprising Chemical-CER, curriculum, and Lagrangian control. Evaluated attempts enter replay, while accepted states update multipliers. Current measurements remain separate from historical pressure. Property labels denote molecular weight (MW), the logarithm of the octanol–water partition coefficient (logP), quantitative estimate of drug-likeness (QED), synthetic accessibility (SA), cytochrome P450 2D6 (CYP2D6), the human ether-à-go-go-related gene potassium channel (hERG), and Caco-2 cell permeability. Icons are schematic. Solid routes show flow or updates, whereas dashed routes show conditioning.

Figure 2: Task-level comparisons against (a–c) original baselines in E1 and (d–f) all six online strategies in E2. Bars show means, dots all five runs, and percentages gains over the strongest baseline in each panel. The selected tasks highlight leading means. Full task results appear in Appendix B . GA, R4, MOL, EX, and SEI abbreviate Graph-GA, REINVENT4, MOLLEO, ExLLM, and SEISMO. FF denotes feasibility-first and FP fixed-penalty search. SEISMO uses terminal-score carry-forward.

\dagger

marks independent reimplementations.

Method	GSK3 $\beta$	JNK3	DRD2	Sitagliptin	Albuterol	Sum
Graph-GA	0.4600	0.1814	0.6939	0.2213	0.4423	1.9988
REINVENT4	0.3792	0.1742	0.6917	0.1513	0.3947	1.7911
MOLLEO †	0.5028	0.3658	0.8041	0.2999	0.8840	2.8565
ExLLM †	0.5150	0.3302	0.7981	0.2657	0.8860	2.7950
SEISMO	0.4850	0.2326	0.9756	0.1344	0.9791	2.8067
PharmAgent	0.7564	0.9186	0.9494	0.3472	0.9492	3.9208

Table 2: Standard PMO target AUC Top-10 (higher is better), averaged over five runs. Sum adds the five task means. All LLM methods use MiniMax-M3.

\dagger

and SEISMO’s original short-run protocol with terminal-score carry-forward are defined in Section 4.1 . Best/second-best means are bold/underlined.

Method	Target AUC Sum	AA-AUC	PHV	Top-1	Top-10
Graph-GA (FF)	1.0675	0.2536	0.4181	0.4375	0.3219
Graph-GA (FP)	1.1809	0.2711	0.4630	0.4837	0.3575
REINVENT4 (FP)	1.8337	0.2859	0.5622	0.6221	0.4663
MOLLEO † (FF)	2.0465	0.4006	0.6430	0.6788	0.5864
ExLLM † (FF)	2.2732	0.4602	0.6514	0.6647	0.6352
SEISMO (FF)	2.9319	0.3838	0.5795	0.6405	0.5978

Table 3: Online-constraint results across five tasks and five runs. Target AUC Sum adds task means. The other metrics average tasks and runs. All methods receive online constraints. FF denotes feasibility-first and FP fixed-penalty search. Best/second-best means are bold/underlined. Figure 3 reports uncertainty.

Figure 3: Online comparison across five tasks and five runs. (a–c) Target-AUC sum, AA-AUC, and PHV for six baseline strategies and PharmAgent E2. Feasibility-first (FF) search prioritizes constraint satisfaction. Fixed-penalty (FP) search penalizes total normalized violation with a weight fixed within each run. Bars show means with unadjusted 95% intervals after within-run task aggregation. All methods have the same budget allowance. SEISMO retains its short-run protocol and terminal-score carry-forward.

Figure 4: Five molecular cases (run 0, before full constraint activation at round 59 versus final accepted state). (a–e) Sigmoid desirability profiles under all seven requirements. Outward is favorable, and the 0.5 ring marks thresholds. CYP2D6 denotes inhibition. Target scores and violation counts accompany each profile. (f) Recorded GSK3

\beta

structures with an arrow spanning subsequent rounds. ADMET endpoints are predictions.

Variant	Target AUC Sum	AA-AUC	PHV
PharmAgent (full system)	3.6936	0.7076	0.8248
w/o Lagrangian	3.6535	0.4781	0.7815
w/o Chemical-CER	3.4008	0.5459	0.7904
w/o Curriculum	3.1981	0.6260	0.7295
w/o MedChem-RAG	2.9977	0.4146	0.6593
w/o Reviewer agent	3.3024	0.5566	0.7100

Table 4: ADMET-PMO ablations over five tasks and five runs. The full-system reference matches Table 3 . Target AUC Sum adds task means. AA-AUC and PHV average tasks and runs. Best/second-best means are bold/underlined.

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A1: Online control improves feasible-solution search. (a) Best feasible target score versus unique oracle calls, including initialization. Lines average five task-averaged runs, with pointwise 95% intervals. (b) E2–E1 AA-AUC differences for every task and the fixed five-task mean, with run-index-matched 95% intervals (

n=5

, unadjusted). Shared indices do not imply shared LLM draws. An independent-group aggregate check remains positive (Appendix B ).

Figure A2: Target AUC across all six online baseline strategies and PharmAgent E2. Diamonds denote means, grey dots the five runs, and bars unadjusted 95% Student-

t

intervals. Panels (a–e) show the five tasks. Panel (f) averages tasks within each run before estimating uncertainty. Intervals are not clipped to

[0,1]

\dagger

denotes independent MiniMax-M3 reimplementations. *SEISMO follows its short-run protocol with terminal-score carry-forward. FF denotes feasibility-first and FP fixed-penalty search.

Figure A3: AA-AUC across all six online baseline strategies and PharmAgent E2. Diamonds denote means, grey dots the five runs, and bars unadjusted 95% Student-

t

intervals. Panels (a–e) show the five tasks. Panel (f) averages tasks within each run before estimating uncertainty. Intervals are not clipped to

[0,1]

\dagger

denotes independent MiniMax-M3 reimplementations. *SEISMO follows its short-run protocol with terminal-score carry-forward. FF denotes feasibility-first and FP fixed-penalty search.

Figure A4: Pareto hypervolume across all six online baseline strategies and PharmAgent E2. Diamonds denote means, grey dots the five runs, and bars unadjusted 95% Student-

t

intervals. Panels (a–e) show the five tasks. Panel (f) averages tasks within each run before estimating uncertainty. Intervals are not clipped to

[0,1]

\dagger

denotes independent MiniMax-M3 reimplementations. *SEISMO follows its short-run protocol with terminal-score carry-forward. FF denotes feasibility-first and FP fixed-penalty search.

Figure A5: Target AUC for the five original baselines and PharmAgent E1/E2. Baselines and E1 receive post-hoc property annotation, while E2 uses online constraints. Diamonds denote means, grey dots the five runs, and bars unadjusted 95% Student-

t

intervals. Panels (a–e) show the five tasks. Panel (f) averages tasks within each run before estimating uncertainty. Intervals are not clipped to

[0,1]

\dagger

denotes independent MiniMax-M3 reimplementations. *SEISMO follows its short-run protocol with terminal-score carry-forward.

Figure A6: AA-AUC for the five original baselines and PharmAgent E1/E2. Baselines and E1 receive post-hoc property annotation, while E2 uses online constraints. Diamonds denote means, grey dots the five runs, and bars unadjusted 95% Student-

t

intervals. Panels (a–e) show the five tasks. Panel (f) averages tasks within each run before estimating uncertainty. Intervals are not clipped to

[0,1]

\dagger

denotes independent MiniMax-M3 reimplementations. *SEISMO follows its short-run protocol with terminal-score carry-forward.

Figure A7: Pareto hypervolume for the five original baselines and PharmAgent E1/E2. Baselines and E1 receive post-hoc property annotation, while E2 uses online constraints. Diamonds denote means, grey dots the five runs, and bars unadjusted 95% Student-

t

intervals. Panels (a–e) show the five tasks. Panel (f) averages tasks within each run before estimating uncertainty. Intervals are not clipped to

[0,1]

\dagger

denotes independent MiniMax-M3 reimplementations. *SEISMO follows its short-run protocol with terminal-score carry-forward.

Figure A8: All five recorded accepted-state structure pairs from E2 underlying Figure 4 . Panels (a–e) show GSK3

\beta

, JNK3, DRD2, Sitagliptin MPO, and Albuterol similarity. Every task uses run 0, comparing the state before full constraint activation (round 59) with the final accepted state. All structures use a common bond-length scale, atom-label size, and line width. Structures are drawn from original canonical SMILES. Target scores and violation counts appear in Figure 4 . Raw properties are reported in Table A1 .

Task	State	MW	logP	QED	SA	CYP	hERG	Caco-2
Bound		$\leq 500$	$\leq 5$	$\geq 0.3$	$\leq 6$	$\leq 0.5$	$\leq 0.5$	$\geq-5.15$
GSK3 $\beta$	R59	401.418	4.4596	0.3982	2.4764	0.0553	0.5754	-5.5022
GSK3 $\beta$	Final	375.215	4.3756	0.6825	2.3569	0.4371	0.3496	-4.4525
JNK3	R59	378.407	4.0013	0.7018	2.7675	0.0638	0.6968	-4.9562
JNK3	Final	461.497	3.8926	0.4802	3.4456	0.0391	0.4993	-5.1188
DRD2	R59	386.511	4.0715	0.6044	2.9680	0.9720	0.9889	-5.2292

Appendix

Table A1: Raw computed or predicted properties for the ten states in Figure 4 . All cases use run 0. Bold values violate the displayed threshold. CYP denotes CYP2D6 inhibition. MW is in Da and Caco-2 follows the predictor scale. R59 means the accepted state at round 59. Final means the final accepted state.

Figure A9: Online performance and comparative gains across five tasks and five runs. Panels (a–c) show five-run means with unadjusted 95% Student-

t

intervals. Panels (d–f) show PharmAgent E2 minus each baseline with unadjusted 95% independent-group intervals. Target AUC uses five-task sums, while AA-AUC and PHV use task means. Each interval uses five run-level aggregates per method. The zero line in (d–f) denotes equal aggregate performance. All six baseline strategies are retained. FF denotes feasibility-first and FP fixed-penalty search. SEISMO retains its short-run protocol and terminal-score carry-forward.

Optimizer	Weight	AA-AUC	PHV	Selected
Graph-GA	0.1	0.206613	0.349602	No
Graph-GA	1	0.209172	0.294164	No
Graph-GA	10	0.209456	0.290444	Yes
REINVENT4	0.1	0.257339	0.489099	No
REINVENT4	1	0.286928	0.542494	Yes
REINVENT4	10	0.262841	0.506246	No

Appendix

Table A2: Independent development runs for selecting fixed-penalty weights. Each row averages five tasks and two runs. Formal outcomes do not enter selection.

Method	Unique evaluations	Evaluation events	Total unique
Graph-GA (FF)	1,000	1,000	25,000
Graph-GA (FP)	1,000	1,000	25,000
REINVENT4 (FP)	1,000	1,000	25,000
MOLLEO (FF)	1,000	1,000	25,000
ExLLM (FF)	1,000	1,000	25,000
SEISMO (FF)	6–47	37–50	955

Appendix

Table A3: Target-evaluation counts across 25 task/run cells per strategy. Ranges are per cell. FF denotes feasibility-first and FP fixed-penalty search. Repeated SEISMO evaluations add events without adding unique molecules. All AUCs use the same 1,000-call horizon.

Figure A10: PharmAgent E2 minus every online baseline on each task. Panels (a–c) show target AUC, AA-AUC, and PHV. Each cell is a difference between five-run means, calculated before rounding. Positive values favor PharmAgent. A shared signed color scale uses purple for positive differences and orange for negative differences. Cell fills and the color bar follow the same signed fourth-root scale, which makes small signed differences visible. Printed values are rounded to four decimals. All 90 comparisons are retained. Target AUC is reported per task. Colors encode mean differences, with uncertainty reported in Figures A2 – A4 . FF denotes feasibility-first and FP fixed-penalty search.

Figure A11: Online minus original aggregate performance within each framework. Panels (a–c) show changes in target-AUC sum, AA-AUC, and PHV. Diamonds denote mean differences and bars unadjusted 95% independent-group intervals over five run-level task aggregates per configuration. Positive changes favor online search. Graph-GA’s original reference is shared by its two adaptations. PharmAgent compares E2 with E1, using the independent-group sensitivity analysis. Each contrast compares complete original and online configurations. FF denotes feasibility-first and FP fixed-penalty search. SEISMO retains its short-run protocol in both conditions.

Constraint	Property	Type	Threshold	Tool
Lipinski MW	Molecular weight	$\leq$	500 Da	RDKit
Lipinski logP	Lipophilicity	$\leq$	5.0	RDKit
QED	Drug-likeness	$\geq$	0.3	RDKit
SA Score	Synth. accessibility	$\leq$	6.0	RDKit
CYP2D6 inhibition	Metabolic safety	$\leq$	0.5	ADMET-AI
hERG blocking	Cardiac safety	$\leq$	0.5 probability	ADMET-AI

Appendix

Table A4: ADMET constraints in the ADMET-PMO benchmark. The same seven constraints are used for all five tasks evaluated in this paper.

Property	Scale $r_{i}$	Steepness $s_{i}$	Justification
MW	150 Da	5.0	Std. dev. in drug-like space
logP	2.0	5.0	Std. dev. in drug-like space
QED	0.2	5.0	Range $[0,1]$ , std $\approx 0.2$
SA Score	2.0	5.0	Range $[1,10]$ , std $\approx 2$
CYP2D6 inh.	0.3	5.0	Probability scale $[0,1]$
hERG blocking	0.3	5.0	Probability scale $[0,1]$

Appendix

Table A5: Property-specific normalization scales used in the ADMET score and Lagrangian penalty terms.

Quantity	E1	E2
Unique target evaluations	25,000	25,000
Recorded optimization rounds	38,013	49,854
Successful HTTP calls	83,653	108,055
HTTP attempts	83,751	108,168
Input tokens	119,050,374	301,475,520
Output tokens	55,388,732	118,552,476

Appendix

Table A6: Recorded E1/E2 compute over 25 runs per arm. Oracle calls count unique target evaluations. HTTP attempts include retries. Round counts include rounds without new molecules.

State / operation	Parameter	Default
Constraint state	Initial penalty $\rho_{0}$	1.0
	Growth factor $\gamma$	1.5
	Maximum penalty $\rho_{\max}$	100
Experience state	Retrieval count $k$	5
	Archive capacity	1000
	Fingerprint radius / bits	2 / 2048

Appendix

Table A7: Hyperparameters for Lagrangian, Chemical-CER, and Curriculum modules.

System	Multi-agent	Online ADMET	Online ADMET	Memory	Cross-task	Five-task
		evaluation	constraints	retrieval	transfer	PMO
Graph-GA	✗	✗	✗	✗	✗	✓
REINVENT4	✗	✗	✗	✗	✗	✓
MOLLEO-MiniMax-M3 †	✗	✗	✗	✗	✗	✓
ExLLM-MiniMax-M3 †	✗	✗	✗	❍	✗	✓
SEISMO-MiniMax-M3	✗	✗	✗	❍	✗	✓

Appendix

Table A8: Features enabled in the evaluated five-baseline protocol. ✓ indicates enabled features and ✗ indicates features not enabled. The symbol ❍ denotes within-task memo or trajectory reuse without fingerprint-based experience retrieval. Online ADMET columns concern the optimization loop. All five baselines separately receive the same post-hoc ADMET annotation.

\dagger

denotes our independent MiniMax-M3 reimplementations.

Method	Within-task state	Structure-aware	Cross-task
Graph-GA	Evolutionary population	✗	✗
REINVENT4	Learned policy parameters	✗	✗
MOLLEO	Population / parent scores	✗	✗
ExLLM	Experience memo	✗	✗
SEISMO	Interaction trajectory	✗	✗
Chemical-CER (ours)	Retrieved transitions	✓	✓

Appendix

Table A9: State and memory mechanisms in the five baseline frameworks and Chemical-CER. Framework names omit the LLM backend. Implementation details are specified in the experimental setup. Within-task state includes populations, learned policy parameters, and explicit history. Structure-aware denotes fingerprint-based retrieval of optimization experiences. Cross-task denotes reuse of an optimization experience archive in the evaluated configurations.

	Target Score			Property-Aware Quality
Method	Top-1	AUC Top-10	5-Task Sum	AA-AUC	PHV
Graph-GA	0.6029	0.3998	1.9988	0.3367	0.5686
REINVENT4	0.5864	0.3582	1.7911	0.2723	0.5238
MOLLEO †	0.7797	0.5713	2.8565	0.4145	0.7016
ExLLM †	0.7364	0.5590	2.7950	0.3121	0.6065
SEISMO	0.6053	0.5613	2.8067	0.2994	0.5204

Appendix

Table A10: Original baselines with post-hoc property annotation, compared with online PharmAgent E2. Results average five tasks and five runs. The 5-Task Sum adds task-level AUC means. All metrics are higher-is-better. Baselines receive post-hoc ADMET annotation. PharmAgent uses online constraints. Backend and evaluation protocols follow Section 4.1 . Best/second-best means are bold/underlined.

Method	Top-1	Top-10
PharmAgent E1	$\mathbf{0.8641}\pm 0.0645$	$\mathbf{0.8606}\pm 0.0634$
PharmAgent E2	$\underline{0.8514}\pm 0.0561$	$\underline{0.8305}\pm 0.0751$
Graph-GA	$0.6029\pm 0.0194$	$0.5498\pm 0.0157$
REINVENT4	$0.5864\pm 0.0469$	$0.4613\pm 0.0264$
MOLLEO †	$0.7797\pm 0.0401$	$0.7484\pm 0.0406$
ExLLM †	$0.7364\pm 0.0565$	$0.7184\pm 0.0561$

Appendix

Table A11: Final target-score summaries for all configurations across five fixed tasks and five runs. Means

\pm

unadjusted 95% interval half-widths. The lower block contains online adaptations. FF denotes feasibility-first and FP fixed-penalty search. Best/second-best means are bold/underlined across all rows. *SEISMO follows its original short-run protocol.

Explore similar work

Sep 28, 2026cs.LG

M3OS: A Monte Carlo Graph Search-Orchestrated Multi-Agent LLM System for Evidence-Traced Molecular Optimization

Small-molecule optimization integrates medicinal-chemistry reasoning and computational evidence through iterative, multi-objective decisions. When large language models (LLMs) reason over optimization histories stored primarily in conversational context, they must recover candidate identities, prior evaluations, and task constraints to guide subsequent decisions. We present M3OS, a multi-agent LLM system that decouples molecular-design reasoning from optimization-state management through Monte Carlo graph search. A persistent graph links evaluated candidates, parent-child transformations and evaluation evidence, while rewards and visit statistics guide LLM-assisted parent selection. Two branches combine tool-driven candidate generation with knowledge- and case-guided medicinal-chemistry editing. An execution harness controls graph updates through structured output extraction, molecular validation and task-bound evaluation. Agents receive role-specific contexts, while the graph preserves optimization trajectories beyond their active contexts. Across three molecular optimization benchmarks, M3OS achieves higher success rates than baselines, supporting the integration of persistent search state, specialized agents and controlled execution for multi-constraint optimization.

Junjie Wang, Yaowei Jin, Ruohui Tang +7

Lingang Laboratory, Shanghai, 200031, China. · School of Information Science and Technology, ShanghaiTech University, Shanghai, 201210, China. · Global Institute of Future Technology, Shanghai Jiao Tong University, Shanghai, 200240, China.

Molecular Optimization Multi-Agent Large Language Model Systems Monte Carlo Tree Search Monte Carlo

Ask PDF

Apr 24, 2026cs.LG

C-MORAL: Controllable Multi-Objective Molecular Optimization with Reinforcement Alignment for LLMs

Large language models (LLMs) show promise for molecular optimization, but aligning them with selective and competing drug-design constraints remains challenging. We propose C-Moral, a reinforcement learning post-training framework for controllable multi-objective molecular optimization. C-Moral combines group-based relative optimization, property score alignment for heterogeneous objectives, and bottleneck-sensitive non-linear reward aggregation to improve stability across competing molecular properties. Experiments on C-MuMOInstruct and S

^2

-Bench MolOpt show that C-Moral achieves the best performance among compared methods on both benchmarks. On C-MuMOInstruct, C-Moral achieves the best Success Optimized Rate (SOR) of 48.9% on in-domain tasks and 39.5% on out-of-domain tasks while preserving scaffold similarity. On S

^2

-Bench MolOpt, it also achieves the strongest results across LogP, MR, and QED optimization tasks. These results suggest that C-Moral is an effective way to align molecular LLMs with continuous and constrained molecular design objectives. Our code and models are publicly available at https://github.com/Rwigie/C-MORAL.

Rui Gao, Youngseung Jeon, Swastik Roy +2

University of California, Los Angeles · Amazon

Molecular Optimization Optimization Modeling Moral Reasoning Multi-Objective Optimization Language Modeling

Ask PDF

Jul 26, 2024cs.LG

Small Molecule Optimization with Large Language Models

Molecular optimization, the process of designing molecules with desirable properties, represents a critical challenge in drug discovery. The recent advancements in large language models (LLMs) have opened new opportunities for their integration with traditional molecular optimization algorithms to improve performance. In this work, we propose Molecular Language Model powered Evolutionary Algorithm (Mol-E), an evolutionary algorithm that relies on the generative capabilities of LLMs trained on molecules and molecular properties. Scientific Contribution. Mol-E obtains the highest aggregate Top-10 AUC among the comparable full-23-task results considered here, scoring 17.500 in the task-agnostic regime, in which the oracle is treated strictly as a black box, and 20.551 in the task-informed regime, in which the optimizer receives a fixed semantic description of the objective. Mol-E also improves over the evaluated baselines on multi-property optimization with docking against DRD2, MK2, and AChE.

Philipp Guevorguian, Menua Bedrosian, Tigran Fahradyan +3

YerevaNN research lab, Yerevan, Armenia.

Molecular Optimization Drug Discovery Optimization Modeling Evolutionary Search Comparison

Ask PDF