TMCS: Tool-Grounded Multi-Agent Reasoning for Compositional Chemical Problem Solving

Shengqin Wang; Jie Jin; Yu Cheng; Yihang Chen; Weilin Luo; Yuan Xie; Zhizhong Zhang

cs.AISep 28, 2026

TMCS: Tool-Grounded Multi-Agent Reasoning for Compositional Chemical Problem Solving

Authors: Shengqin Wang, Jie Jin, Yu Cheng, Yihang Chen, Weilin Luo, Yuan Xie, Zhizhong Zhang

Organizations: East China Normal University, Shanghai Innovation Institute, University College London, Huawei Noah’s Ark Lab

Abstract

Despite the promise of Large Language Models (LLMs) in computational chemistry, rigorous combinatorial chemistry problems remain difficult because they require quantitatively constrained molecular modification, candidate validation, and systematic revision after failed attempts. Existing tool-augmented chemical agents demonstrate useful planning and tool use, but they rarely provide a unified loop for property-driven molecular optimization and workflow-level composition. To bridge this gap, we propose Tool-Grounded Multi-Agent Reasoning for Compositional Chemical Problem Solving (TMCS), a step-by-step multi-agent framework that formalizes chemical problem solving as an interpretable, tool-augmented workflow. At the task level, specialized agents leverage external tools, few-shot trajectory memory, and structured reflection to iteratively refine solutions. At the workflow level, TMCS chains generation, understanding, editing, description, and optimization into a closed-loop pipeline. Evaluations across multiple chemical tasks demonstrate that TMCS consistently enhances chemical reasoning across both open- and closed-source base models, achieving state-of-the-art performance.

Figures & tables

Figure 1: Mean property improvement ( $\Delta$ ) across six molecular optimization objectives. The grouped bar charts compare the proposed TMCS framework against its corresponding generative baselines: Llama-3.3-70B-Instruct (left) and GPT-5.4 (right). TMCS consistently achieves superior $\Delta$ values across both physicochemical constraints and bioactivity targets. Higher positive values indicate successful property enhancements, whereas values near or below zero signify property degradations.

Figure 2: The overall macroscopic architecture of the TMCS framework. It illustrates the compositional chemical pipeline supported by the Task Router, Memory Bank, and Tool Service.

Figure 3: The task-level tool-grounded Agent Loop and memory construction process. It details how raw chemical data is transformed into few-shot memory to guide the self-evolving refinement cycle.

Method	Bounded reflection	Task routing	Closed-loop optimization	Multi task
ChemCrow	✗	$\triangle$	✗	✗
DrugAgent	✗	$\triangle$	$\triangle$	✗
MADD	$\triangle$	✓	$\triangle$	✗
ChemCRAFT	$\triangle$	✗	$\triangle$	✗
TMCS	✓	✓	✓	✓

Table 1: Capability comparison with representative tool-augmented chemical agents. ✓: full support;

\triangle

: partial support; ✗: absent or not a primary focus.

Figure 4: Example of QED attribute optimization.

Methods	FG $\downarrow$	Ring $\downarrow$	Murcko $\uparrow$	Ring-sys $\uparrow$	Eq $\uparrow$	Add	Delete	Sub	Design	Capt
General Models
GPT-4o	0.17	1.35	0.21	80.0	72	80	80	65.0	0.07	0.01
DeepSeek-R1	0.27	1.55	0.34	45.0	65	70	70	68.3	0.22	0.04
GPT 5.4	0.29	1.05	0.42	45.0	84	80	80	88.3	0.19	0.13
Gemini-2.5-pro	0.11	0.60	0.51	87.5	82	100	85	81.7	0.29	0.04
Llama-3.3-70B-Instruct	0.52	1.80	0.12	68.3	67	60	80	50.0	0.03	0.02

Table 2: Results on individual chemical tasks. 70B represents Llama-3.3-70B-Instruct,

\downarrow

indicates lower is better, while

\uparrow

indicates higher is better.

Methods	LogP		Solubility		QED		DRD2		JNK3		GSK3- $\beta$
Methods	$\Delta$	SR	$\Delta$	SR	$\Delta$	SR	$\Delta$	SR	$\Delta$	SR	$\Delta$	SR
General Models
GPT-4o	-0.09	37	0.92	80	0.13	70	0.07	48	-0.02	30	-0.00	39
DeepSeek-R1	0.47	69	0.80	80	0.17	72	0.12	62	-0.02	29	0.01	41
Gemini-2.5-pro	-0.22	76	1.06	70	0.28	84	0.36	74	-0.02	35	0.06	68
Llama-3.3-70B-Instruct	0.02	35	0.72	81	0.15	61	0.00	31	-0.01	30	0.01	40

Table 3: Optimization results on multiple molecular objectives.

\Delta

is the mean property improvement, where a negative

\Delta

indicates that most optimizations are property degradations. SR denotes success rate (%).

Model	Strategy	LogP		Solubility		QED		DRD2		JNK3		GSK3- $\beta$
Model	Strategy	$\Delta$	SR	$\Delta$	SR	$\Delta$	SR	$\Delta$	SR	$\Delta$	SR	$\Delta$	SR
70B	Baseline	0.152	41.0	0.004	15.0	-0.003	16.0	0.000	17.0	0.001	15.0	0.002	14.0
70B	TMCS (Ours)	1.056	89.0	0.416	65.0	0.123	81.0	0.094	69.0	0.028	68.0	0.085	74.0
GPT-5.4	Baseline	-0.417	82.0	1.305	80.0	0.095	78.0	0.002	42.0	0.004	30.0	0.003	35.0
GPT-5.4	TMCS (Ours)	1.101	98.0	1.062	92.0	0.178	94.0	0.148	80.0	0.054	96.0	0.124	90.0

Table 4: Workflow-level optimization-stage results: comparison between the task-specific Pure LLM Baseline and TMCS across six molecular objectives. We report

\Delta

(mean net property change over all 100 cases, including positive and negative changes) and SR (success rate, %). 70B represents Llama-3.3-70B-Instruct.

Method	$\Delta$	SR	$\Delta$	SR
	LogP		DRD2
Ours	1.10	96	0.30	90
w/o few-shot	0.96	88	0.27	84
w/o tool	0.96	86	0.26	80

Table 5: Ablation results on molecular optimization tasks.

\Delta

is the mean property improvement, where a negative

\Delta

indicates that most optimizations are property degradations, and SR denotes success rate.

Setting	Solubility $\Delta$
Baseline	0.72
Ablation on reflection rounds
w/o reflection	0.87
1 round	0.90
5 rounds	0.93
3 rounds (Ours)	0.92

Table 6: Ablation results on reflection rounds for molecular optimization.

\Delta

denotes mean property improvement; TMCS uses three rounds by default.

Model	Start Source	Strategy	DRD2	GSK3- $\beta$	JNK3
GPT-5.4	Gen-Start	Baseline	40.9	31.9	32.9
	Gen-Start	TMCS	78.8	94.2	95.7
	GT-Start	Baseline	44.1	41.9	23.3
	GT-Start	TMCS	82.4	80.6	96.7

Table 7: Ablation results on initialization source for bioactivity tasks (DRD2, GSK3-

\beta

, JNK3).

Explore similar work

One Policy Is Enough: Single-Agent Reinforcement Learning Outperforms Tree Search for Chemistry Tool Learning
Aug 31, 2026Armin Dariani, Sifan Wu, Bang Liu +1Chemistry Monte Carlo Tree Search
1DIRO, Université de Montréal · 2Mila - Quebec Artificial Intelligence Institute · 3Innovation Campus Delaware, Air Liquide
A Multitask Large Reasoning Model for Molecular Science
Mar 13, 2026Pengfei Liu, Shuang Ge, Xiaobo Wang +6Moleculenet Large Reasoning Models
Guangzhou Key Laboratory of Forensic Multi-Omics for Precision Identification, Southern Medical University, Guangzhou, Guangdong, 510515, China · Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, Guangdong 518000, China · Guangdong Provincial Key Laboratory of Psychoactive Substances Monitoring and Safety, Anti-Drug Technology Center of Guangdong Province (National Anti-Drug Laboratory Guangdong Regional Center), Guangzhou, 510230, China +4
Can Agents Price a Reaction? Evaluating LLMs on Chemical Cost Reasoning
May 8, 2026Yuyang Wu, Yue Huang, Shuaike Shen +8Reaction
Carnegie Mellon University · University of Notre Dame · University of North Carolina, Chapel Hill +1

Aug 31, 2026, cs.LGJ/K move · Enter open · S save

One Policy Is Enough: Single-Agent Reinforcement Learning Outperforms Tree Search for Chemistry Tool Learning

Armin Dariani, Sifan Wu, Bang Liu, Entao Yang

1DIRO, Université de Montréal · 2Mila - Quebec Artificial Intelligence Institute · 3Innovation Campus Delaware, Air Liquide

Chemistry Monte Carlo Tree Search Offline Reinforcement Learning Trees

Open paper AskPDF arXiv

Chemistry questions often demand exact computation and database lookups that a language model cannot supply from its parameters, so it must reach for external tools. Tool use here is a three-part problem: select the right tool from a large pool, fill it with correctly typed arguments, and chain calls so that each consumes the outputs of the last. CheMatAgent, a previously published system, addresses this with hierarchical evolutionary MCTS: separate policy and execution models searching tool-call trees under two learned critics, one regressed partly onto GPT-assigned scores. We show that a single policy suffices. Our model interleaves reasoning, tool calls, and returns in one left-to-right generation, trained by a supervised warm-up and then outcome-level reinforcement learning against a programmatic reward read directly off the gold call chain, which leaves no learned critic and no judge in the training loop. On ChemToolBench multiple-tool comprehensive chemistry, on both backbones CheMatAgent use, we improve Tool F1 by 5.5% and Return F1 by 9.6% on Qwen-2.5-7B, and by 3.7% and 3.9% on Llama-3.1-8B, compared with their strongest search configuration, at one model invocation per question, against a search whose cost grows with the tree; we also lead answer Pass Rate on Qwen-2.5-7B.

Abstract

Explore similar work

One Policy Is Enough: Single-Agent Reinforcement Learning Outperforms Tree Search for Chemistry Tool Learning

A Multitask Large Reasoning Model for Molecular Science

Can Agents Price a Reaction? Evaluating LLMs on Chemical Cost Reasoning