Reasoning or Just Prompting? Probing the Mechanism of Latent Chemical Reasoning

Kaiyuan Lu; Jiatong Li; Yatao Bian

cs.LGOct 25, 2026

Reasoning or Just Prompting? Probing the Mechanism of Latent Chemical Reasoning

Authors: Kaiyuan Lu, Jiatong Li, Yatao Bian

Organizations: National University of Singapore

Abstract

Latent reasoning performs intermediate computation in continuous space rather than through explicit tokens, which is particularly appealing in chemistry, where errors in explicit reasoning can propagate to the final answer. Yet if latent states truly constitute an implicit reasoning chain, they should be causally important. We find otherwise: swapping or moderately perturbing intermediate latent states has little effect on performance, while removing them entirely sometimes causes a substantial score drop. Thus, the latent states matter, but their molecule-specific content appears to matter far less. This points to a simpler interpretation. Rather than encoding a step-by-step reasoning chain, latent states may primarily steer subsequent model computation, much like soft prompts. Indeed, a single fixed prefix shared across molecules reproduces LatentChem's performance, and a learned static soft prompt performs comparably to per-input latent generation in controlled models. The tolerance of latent sequences to substitution and moderate perturbation further supports this connection. We therefore suggest that latent chemical reasoning may be better viewed as soft prompting than as a reasoning chain hidden in continuous space.

Figures & tables

Figure 1: Response organization for one ChEBI-20 test molecule. Matched and Swapped Z yield identical responses; the orange molecule supplies

Z_{B}

. X-only produces explicit steps, and a hard prompt lengthens the final answer while retaining those steps. Bold highlights chemical descriptions; dark red marks errors. Ellipses denote omissions; counts cover complete outputs (Appendix E ). Latent squares are schematic.

Figure 2: LatentChem interventions. (a) Scores relative to each condition’s paired Matched run: ChEBI-20 METEOR and optimization property improvement. Gaussian points average three draws; hard prompt is ChEBI-20 only. (b) Full-response chemical-correctness preferences on 200 ChEBI-20 molecules, with three draws per molecule for Gaussian. Supplementary intervention results are collected in Appendix B .

Figure 3: Prefix length and noise tolerance on ChEBI-20. (a) Step-free responses across prefix lengths (160 molecules);

H_{0}

markers show ten-position Matched/Fixed prefixes and X-only with the initial context. (b) METEOR relative to the paired Matched run and step-free responses under full-coordinate noise (200 molecules, three draws). Amplitude

\alpha

is relative to each position’s latent RMS;

\alpha=0

is Matched Z. Points and error bars in (b) show means and sample standard deviations across draws.

Figure 4: Shared trajectories, molecular readout, and attention. (a) Centered variance decomposition; the annotation summarizes six optimization tasks (100 molecules each). LC: LatentChem; Curr.: Curriculum. Controlled representations in (a–b) are from Latent reasoning models and use 200 inputs per task. (b) Readout gains over size alone on the same 200 ChEBI-20 molecules. (c) Attention to

Z

at chemical-content positions minus position-matched controls (153 ChEBI-20 molecules); filled/open markers denote Latent reasoning/Soft prompting. Error bars in (b–c) show 95% molecule-bootstrap intervals.

Training	Model / condition	ChEBI-20		Reaction
		METEOR	Identical (%)	Exact (%)	Identical (%)
Curriculum	Latent reasoning (base)	0.5727	–	77.0	–
	+ Swapped Z	0.5703 ( $-0.42\%$ )	75.0	77.0 ( $+0.00\%$ )	99.0
	+ Mean Z	0.5693 ( $-0.60\%$ )	79.5	77.0 ( $+0.00\%$ )	98.0
	Soft prompting	0.5769 ( $+0.72\%$ )	–	78.0 ( $+1.30\%$ )	–
Direct	Latent reasoning (base)	0.5739	–	77.0	–

Table 1: Controlled latent reasoning and soft prompting on 200 official test inputs per task. Parentheses show relative score changes from the Matched Z baseline within each training recipe. Identical denotes exact answer agreement with that baseline after substitution. Soft prompts are trained separately under the same recipe; paired confidence intervals and removal controls are in Appendix C .

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Epochs	Updates	Examples seen	Initial LR	Trainable params
Curriculum / Latent reasoning	4	1800	14,400	5e-05	13,631,488
Curriculum / Soft prompting	4	1800	14,400	5e-05	13,664,256
Direct / Latent reasoning	1	450	3,600	1e-04	13,631,488
Direct / Soft prompting	1	450	3,600	1e-04	13,664,256

Appendix

Table 2: Controlled-model training exposure. Curriculum epochs sum four separate stages; comparisons are matched within each recipe.

Task	Evaluation cohort	Exact overlap	Mean NN	NN $\geq 0.9$	Shared scaffold
ChEBI-20	Curated validation	0/200	0.703	70/200	100/128
ChEBI-20	Official test	0/200	0.528	28/200	68/146
Forward reaction	Curated validation	0/200	0.375	0/200	—
Forward reaction	Official test	0/200	0.353	0/200	—

Appendix

Table 3: Proximity to task-matched adaptation training. Each training pool contains 1,800 distinct canonical inputs. Exact overlap counts canonical input matches; mean NN is the mean maximum training-set Tanimoto similarity. Shared scaffold counts exact Murcko matches to any ring-containing training molecule, with ring-containing evaluation molecules as the denominator.

Task	Matched	X-only	Zero	Swapped	Mean	Swap identity (%)
ChEBI-20	0.1558	0.0610	0.0815	0.1535	0.1564	16.5
QED	0.1862	0.2031	0.2148	0.1923	0.1921	93
LogP	1.3615	1.2575	1.2868	1.3496	1.3680	92
Solubility	1.4772	1.3804	1.0710	1.4723	1.4819	90
DRD2	0.3041	0.2295	0.2144	0.3110	0.3118	90
JNK3	0.0907	0.0521	0.0841	0.0857	0.0861	87

Appendix

Table 4: LatentChem sequence interventions. ChEBI-20 uses METEOR on 200 molecules; optimization uses winsorized property improvement on 100 inputs per task. Identity is exact agreement between Swapped and Matched outputs. Zero Z keeps the latent positions attended but sets their vectors to zero, providing a control for removal.

Condition	METEOR	Tag onset (%)	Answer words	Steps
Matched Z (main run)	0.1558	94.5	123.2	0
X-only (main run)	0.0610	0.0	16.7	5
Matched Z (follow-up)	0.1507	94.0	120.5	0
X-only (follow-up)	0.0689	0.0	17.7	5
X-only + hard prompt	0.1157	0.0	41.6	1

Appendix

Table 5: Response organization and chemical correctness on ChEBI-20. (a) Final-answer word counts are means and step counts are medians. The hard-prompt comparison uses the follow-up run’s paired baselines; explicit steps occur in 98% of hard-prompt responses. (b) Complete-response preferences against Matched Z; Better favors the listed condition. Repeated conditions contain three draws on the same 200 molecules, not 600 independent molecules.

Task	Change to donor	95% CI
ChEBI-20 (METEOR)	-0.0006	[-0.0060, +0.0049]
QED	+0.0011	[-0.0007, +0.0040]
LogP	+0.0010	[-0.0007, +0.0031]
Solubility	+0.0005	[-0.0025, +0.0038]
DRD2	-0.0014	[-0.0047, +0.0013]
JNK3	+0.0008	[-0.0007, +0.0025]

Appendix

Table 6: Change in similarity to the latent donor after swapping. Description uses METEOR against the donor reference; optimization uses Tanimoto similarity to the donor input. Intervals are paired 95% confidence intervals.

Task	Matched	Fixed	Relative (%)	Identity (%)	Change [95% CI]
ChEBI-20	0.1507	0.1522	100.98	14.5	+0.0015 [-0.0039, +0.0069]
QED	0.1870	0.1872	100.10	97.0	+0.0002 [-0.0007, +0.0014]
LogP	1.3761	1.3640	99.12	92.0	-0.0121 [-0.0601, +0.0323]
Solubility	1.4772	1.5015	101.64	88.0	+0.0243 [-0.0238, +0.0904]
DRD2	0.3082	0.3064	99.43	87.0	-0.0017 [-0.0094, +0.0048]
JNK3	0.0907	0.0877	96.66	87.0	-0.0030 [-0.0101, +0.0042]

Appendix

Table 7: Fixed-prefix endpoint results on 200 ChEBI-20 molecules and six optimization tasks with 100 molecules per task. Scores use each condition’s paired Matched run.

Budget	Condition	METEOR	Tag onset (%)	Steps (%)
0	Matched	0.0724	0.00	70.62
2	Matched	0.0611	0.00	100.00
2	Fixed	0.0630	0.00	100.00
4	Matched	0.1384	86.25	4.38
4	Fixed	0.1379	87.50	7.50
6	Matched	0.1403	99.38	0.00

Appendix

Table 8: Description length and initial-context controls on the same 160 evaluation molecules. Post-hoc masked Z suppresses the block after ten native steps; it differs from native zero-budget decoding. The

H_{0}

X-only row has no latent positions or boundary markers; its budget label identifies the paired run.

Task	Positions	Score		No explicit steps (%)
		Matched	Fixed	Matched	Fixed
QED	0	0.2104	–	17.0	–
	2	0.2332	0.2215	0.0	0.0
	4	0.2281	0.2324	6.0	3.0
	6	0.1938	0.1827	100.0	100.0
	10	0.1916	0.1856	100.0	100.0

Appendix

Table 9: Optimization length sweeps. Each row uses all 100 task inputs; the same calibration prefix is truncated to the required length.

Task	Matched Z		Fixed Z		X-only
	Score	No steps (%)	Score	No steps (%)	Score	No steps (%)
QED	0.2049	100.0	0.2015	100.0	0.2049	0.0
LogP	1.3210	100.0	1.3229	100.0	1.0593	0.0
Solubility	1.5410	100.0	1.5697	100.0	1.1432	0.0
DRD2	0.2982	100.0	0.2949	100.0	0.2031	0.0
JNK3	0.0971	100.0	0.0947	100.0	0.0380	0.0

Appendix

Table 10: Initial-context controls on the six optimization tasks. All conditions use

H_{0}

; Matched and Fixed use ten latent positions and retain the boundary markers, while X-only removes both the latent positions and the markers.

Task	Paired Matched	Gaussian mean	Relative (%)	SD (pp)
ChEBI-20	0.1558	0.1544	99.0	1.1
QED	0.1849	0.1865	100.8	1.3
LogP	1.3930	1.3702	98.4	1.1
Solubility	1.4801	1.4878	100.5	2.8
DRD2	0.3104	0.3025	97.5	1.4
JNK3	0.0897	0.0882	98.4	2.9

Appendix

Table 11: Gaussian-prefix scores, averaged over three draws on the same molecules. Relative scores use the paired Matched run; SD is the sample standard deviation of the relative score across draws, in percentage points (pp).

(a) Molecular description: relative score $\pm$ SD (%)
Task	$\alpha=0.1$	$\alpha=0.25$	$\alpha=0.5$	$\alpha=1$
ChEBI-20	100.0 $\pm$ 0.5	99.1 $\pm$ 0.9	92.7 $\pm$ 1.3	61.4 $\pm$ 1.8

Appendix

Table 12: Full-coordinate RMS perturbations across all seven tasks. (a) ChEBI-20 METEOR on 200 molecules. (b) Optimization scores, response organization, and answer changes on 100 molecules per task. Relative scores use the paired Matched run (100%); SD is the sample standard deviation across three draws, in percentage points. Score-change intervals in (b) are paired molecule-bootstrap 95% intervals, not between-draw SDs.

Recipe	Task	Condition	Matched	Score	Change [95% CI]	Identity (%)
Curriculum	Description	Swapped Z	0.5727	0.5703	-0.0024 [-0.0095, +0.0044]	75.0
Curriculum	Description	Mean Z	0.5727	0.5693	-0.0034 [-0.0106, +0.0036]	79.5
Curriculum	Description	Soft prompting	0.5727	0.5769	+0.0041 [-0.0079, +0.0166]	44.0
Curriculum	Description	X-only	0.5727	0.5777	+0.0049 [-0.0048, +0.0152]	51.5
Direct	Description	Swapped Z	0.5739	0.5751	+0.0011 [-0.0040, +0.0065]	79.5
Direct	Description	Mean Z	0.5739	0.5720	-0.0019 [-0.0086, +0.0040]	81.5

Appendix

Table 13: Paired controlled-model contrasts on 200 official test inputs per task. Description uses METEOR; reaction scores are percentages and changes are percentage points. Soft prompting is separately trained under the same recipe. All rows use the batched intervention run and its paired baseline.

	Description			Reaction
Model	Matched	X-only	Zero Z	Matched	X-only	Zero Z
Curr. / Latent reasoning	0.5727	0.5777	0.5694	77.0	77.5	78.0
Curr. / Soft prompting	0.5769	0.5739	0.5728	78.0	78.0	78.5
Direct / Latent reasoning	0.5739	0.5670	0.5751	77.0	78.5	78.0
Direct / Soft prompting	0.5733	0.5723	0.5745	77.5	78.0	78.0

Appendix

Table 14: Latent removal and zeroing in all four controlled models. Description scores are METEOR; reaction scores are exact-match percentages. Every listed condition has 100% strict answer-tag onset and no detected explicit steps.

Representation / property	Size-only $R^{2}$	Combined $R^{2}$	$\Delta R^{2}$ [95% CI]
Z mean / MW	0.9838	0.9834	-0.0004 [-0.0017, +0.0004]
Z mean / logP	0.6393	0.7833	+0.1440 [+0.0473, +0.2575]
Z mean / TPSA	0.8189	0.9083	+0.0894 [+0.0519, +0.1580]
Z mean / Rings	0.4715	0.8628	+0.3913 [+0.3012, +0.5163]
H0 mean / MW	0.9838	0.9839	+0.0001 [-0.0020, +0.0019]
H0 mean / logP	0.6393	0.8514	+0.2121 [+0.1018, +0.3530]

Appendix

Table 15: Size-controlled property readout on ChEBI-20. Intervals resample held-out predictions at the molecule level.

Model	Control	MW	logP	TPSA	Rings
LC	linear	-0.0004	+0.1440 [+0.0473, +0.2575]	+0.0894 [+0.0519, +0.1580]	+0.3913 [+0.3012, +0.5163]
LC	quadratic	-0.0004	+0.1376 [+0.0600, +0.2400]	+0.1055 [+0.0488, +0.2191]	+0.3660 [+0.2102, +0.6414]
Curr.	linear	+0.0022	+0.2291 [+0.1201, +0.3680]	+0.1304 [+0.0846, +0.2072]	+0.4577 [+0.3661, +0.5855]
Curr.	quadratic	+0.0028	+0.1950 [+0.1245, +0.2870]	+0.1370 [+0.0817, +0.2486]	+0.4614 [+0.2811, +0.8030]
Direct	linear	+0.0012	+0.2315 [+0.1144, +0.3702]	+0.1266 [+0.0804, +0.2014]	+0.4564 [+0.3633, +0.5868]
Direct	quadratic	+0.0014	+0.2212 [+0.1234, +0.3480]	+0.1446 [+0.0794, +0.2816]	+0.4667 [+0.2736, +0.8312]

Appendix

Table 16: Linear and quadratic size controls on the same 200 description molecules and scaffold folds. Cells show

\Delta R^{2}

with 95% intervals where available; LC denotes LatentChem, and Curr. and Direct denote controlled latent-reasoning models.

Representation	Macro AP [95% CI]	Shuffled AP
Size only	0.3417 [0.3240, 0.3735]	0.2516
H0 mean	0.7119 [0.6821, 0.7520]	0.2743
Z mean	0.5710 [0.5476, 0.6111]	0.2729

Appendix

Table 17: Readout of 30 functional-group and local-structure labels on ChEBI-20. Macro average precision (AP) and shuffled-label baselines use the same label set.

Model	Task	Molecule (%)	Shared step (%)	Interaction (%)
LatentChem	Description	0.799	96.234	2.966
LatentChem	QED	0.242	98.592	1.166
LatentChem	LogP	0.305	98.691	1.005
LatentChem	Solubility	0.198	98.849	0.953
LatentChem	DRD2	0.230	98.609	1.161
LatentChem	JNK3	0.202	98.839	0.959

Appendix

Table 18: Centered latent variance by source on the official test cohorts (200 description molecules, 100 inputs per optimization task, and 200 controlled-model reaction inputs). The molecule and interaction terms are input-dependent; their sum ranges from 1.151% to 1.417% across the six optimization tasks. The step term is shared across inputs. Static prompts have zero molecule and interaction terms by construction.

Model	Task	Median $\alpha$ [95% CI]	Range	Identity
LatentChem	Description	0.0767	0.0336–0.2248	16.5%
LatentChem	QED	0.0435	0.0248–0.1582	93.0%
LatentChem	LogP	0.0448	0.0266–0.1178	92.0%
LatentChem	Solubility	0.0384	0.0218–0.2058	90.0%
LatentChem	DRD2	0.0464	0.0198–0.1467	90.0%
LatentChem	JNK3	0.0401	0.0198–0.1296	87.0%

Appendix

Table 19: RMS-normalized amplitudes of actual substitutions. Intervals are reported where available; LatentChem description identity is 16.5%.

Model	Task	Position 1	Position 4	Final
LatentChem	Description	0.1751	0.0229	0.0418
LatentChem	DRD2	0.1078	0.0142	0.0263
LatentChem	LogP	0.1038	0.0168	0.0304
LatentChem	QED	0.1078	0.0141	0.0283
LatentChem	Solubility	0.0957	0.0132	0.0293
LatentChem	GSK3- $\beta$	0.1073	0.0148	0.0240

Appendix

Table 20: Between-molecule dispersion at the first, fourth, and final latent positions. The final position is 10 for LatentChem and 8 for controlled models. Showing the final value also captures the rebound after LatentChem’s early contraction.

Chemical Reasoning in LLMs

See all 16 papers

Latent JEPA: Abstract Future Prediction for Latent Reasoning in Chemistry
Oct 1, 2026Xinjian Zhao, Yaoyao Xu, Xuemin Chen +2Molecular Optimization Chemical Reasoning in LLMs
School of Data Science, The Chinese University of Hong Kong, Shenzhen · Shanghai Artificial Intelligence Laboratory
Chemical Chain-of-Thought Functions as a Hallucination-Prone Molecular Scratchpad
Jul 23, 2026Jiatong Li, Yuxuan Ren, Weida Wang +2CoT Faithfulness Chemical Reasoning in LLMs
Blue Whale Lab, National University of Singapore · Fudan University · Hong Kong Polytechnic University
Does Latent Chain-of-Thought Actually Reason? A Single, Unifying Taxonomy for Latent Reasoning
Oct 15, 2026Jingjie Qiu, Sheldon Yu, Xunyi Jiang +6Latent CoT Reasoning CoT Evaluation
University of California, San Diego · New York University · University of Virginia +1

Oct 1, 2026J/K move · Enter open · S save

Latent JEPA: Abstract Future Prediction for Latent Reasoning in Chemistry

Xinjian Zhao, Yaoyao Xu, Xuemin Chen, Xiaozhuang Song+1

School of Data Science, The Chinese University of Hong Kong, Shenzhen · Shanghai Artificial Intelligence Laboratory

Molecular Optimization Chemical Reasoning in LLMs Joint-Embedding Predictive Architecture Predictive Representation Learning Molecular Representation Learning Latent Reasoning

Open paper AskPDF arXiv

Large language models offer a promising foundation for chemical reasoning, bringing together chemical knowledge and multistep problem solving. Chemical intuition can provide an initial sense of plausible outcomes before the details of a solution are fully worked out. Inspired by how such expectations complement explicit analysis, we study how continuous latent thoughts can be trained to anticipate informative aspects of future solutions without verbalizing every intermediate step. We introduce Latent JEPA, a framework that combines autoregressive learning with joint-embedding prediction of one or more future views. For chemical reasoning, we develop textual and molecular prediction objectives that connect latent thoughts to both subsequent reasoning and molecular outcomes. Experiments on ChemCoTBench show gains in molecular optimization and on several editing and reaction metrics. Representation analyses show that future prediction makes latent thoughts more informative about molecular outcomes and strengthens their correspondence with chemical structure. These findings support abstract future prediction as a learning principle for connecting continuous latent reasoning with scientific outcomes.

Abstract

Chemical Reasoning in LLMs

Explore similar work

Latent JEPA: Abstract Future Prediction for Latent Reasoning in Chemistry

Chemical Chain-of-Thought Functions as a Hallucination-Prone Molecular Scratchpad

Does Latent Chain-of-Thought Actually Reason? A Single, Unifying Taxonomy for Latent Reasoning