acceptodds
Under review as a conference paper at ICLR 2027

RxnMiner: A Multimodal Agent for Structured Reaction Data Extraction from Chemistry Figures

14 views

Abstract

Reaction data underpin machine learning for reaction prediction, condition recommendation, and retrosynthetic planning, yet extracting these data from literature figures still requires substantial manual effort. Many figures encode multiple experiments through shared templates, substituent tables, and product grids. Recovering concrete reaction records therefore requires molecular recognition, association of evidence with individual experimental entries, and reconstruction of indirectly specified participants.We present RXNMINER, a multimodal agent built on Pi that combines an expert-informed extraction procedure with a chemistry toolbox to coordinate these operations. The agent reads molecular structures, resolves shared and entry-specific evidence, and instantiates participants from templates to assemble source-linked reaction records. We also introduce CHEMRXN-BENCH, a benchmark of 603 reaction figures with 8,547 manually annotated and reviewed reaction records, evaluated under a common protocol for exact molecular identity and reactant–product pairing.On this benchmark, RXNMINER raises the locally deployed Qwen3.8-Flash-Next backbone's stereochemistry-aware reaction macro-F1 from 21.0% under vanilla prediction to 71.0%, compared with 25.1% for the strongest existing extraction baseline. Gains over Bare Pi across all five evaluated backbones support the combined value of the extraction procedure and chemistry toolbox. Applied without human intervention to 10,107 literature figures, RXNMINER produces a silver corpus of 88,202 source-linked reaction records. We release the benchmark, scorer, agent artifacts, and extracted corpus to support reproducible evaluation and scalable curation of reaction knowledge.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
chemical reaction extraction, chemistry figures, multimodal document understanding, optical chemical structure recognition, reaction parsing, multimodal large language models, scientific agents, benchmark

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SLJNOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents RxnMiner, a multimodal agent that integrates a vision-language model, the Pi Agent framework, MolParser, R-group reconstruction tools, and RDKit to extract structured reaction data from chemical figures. The authors also introduce ChemRxn-Bench and report an improvement in reaction extraction F1 from 21% to 71% using Qwen as the backbone. However, the paper has notable limitations in methodological novelty and experimental design.

Critical strengths
  1. Extracting complete reaction records from complex chemical literature is a valuable task, particularly when dealing with shared templates and implicitly specified reactants.

  2. ChemRxn-Bench contains 603 figures and 8,547 manually annotated reaction records, providing a more rigorous evaluation benchmark for complete reaction extraction.

Critical weaknesses

W1. Existing systems like OpenChemIE and ChemEAGLE already support R-group substitution, reactant reconstruction, and chemistry-specific tool integration. The paper does not sufficiently clarify its contributions beyond these prior approaches.

W2. Table 2 reports F1 scores of 71.0, 69.0, and 64.0 for Qwen, Kimi, and DeepSeek, respectively, whereas Appendix D.2 reports 78.4, 76.3, and 71.2 for the corresponding models, without adequately explaining these discrepancies. Additionally, the abstract reports ChemEAGLE's F1 as 25.1%, while Table 1 shows that 25.1% corresponds to Precision and the actual F1 is 25.0%.

Questions for authors

Can the authors report average inference time, token consumption, the number of tool call to assess its practical efficiency?

6 comments
AuthorsOpenReview ↗

[Response 1/3] - Score Corrections and Method Contributions

Dear Reviewer,

We sincerely thank you for the careful reading and for recognizing the value of complete reaction extraction and of a rigorous benchmark for it. Below we respond to each point; all changes are in the revised PDF and marked in red.

W2: Table 2 reports F1 scores of 71.0, 69.0, and 64.0 for Qwen, Kimi, and DeepSeek, whereas Appendix D.2 reports 78.4, 76.3, and 71.2, without adequately explaining these discrepancies. Additionally, the abstract reports ChemEAGLE's F1 as 25.1%, while Table 1 shows that 25.1% corresponds to Precision and the actual F1 is 25.0%.

Thank you for carefully identifying these inconsistencies. We apologize for the errors in reporting these results. After re-examining the original evaluation records, we identified two issues and have corrected them in the revised manuscript.

The first is that the paper mixed two ways of computing reaction F1. Table 2 computes F1 for each figure and then averages over figures. App. D.2 instead averaged precision and recall over figures first and computed F1 from those averages, which gives higher numbers. App. D.2 now uses the same per-figure definition as Table 2.

The second is rounding: some values were rounded before they went into the tables, which is why many of them ended in ".0". The per-figure results themselves were correct, and we have now recomputed every number from the stored outputs. The ChemEAGLE value in the abstract was its precision by mistake and is now its F1.

ItemBeforeAfter
RxnMiner (Qwen / Kimi / DeepSeek), App. D.278.4 / 76.3 / 71.270.8 / 69.0 / 63.6
Same systems, Table 271.0 / 69.0 / 64.070.8 / 69.0 / 63.6
ChemEAGLE, abstract25.125.0

For these three backbones, the Table 2 scores changed by at most 0.4 points. The larger changes in App. D.2 reflect the change in aggregation. The ranking is unchanged. We apologize for the confusion.

W1: Existing systems like OpenChemIE and ChemEAGLE already support R-group substitution, reactant reconstruction, and chemistry-specific tool integration. The paper does not sufficiently clarify its contributions beyond these prior approaches.

Thank you for this important concern. We agree that OpenChemIE and ChemEAGLE introduced R-group substitution, reactant reconstruction, and chemistry-specific tool integration;RxnMiner builds on these ideas and extends them in several ways: (1) more expressive R-group substitution, supporting complex Markush structures and preserving stereochemistry; (2) more accurate reconstruction of complete reactions, including reactants, products, and their corresponding labels; and (3) a richer chemistry toolchain, integrating tools such as MolParser and special-designed tools for structure recognition, reconstruction, and validation. Together, these improvements lead to substantially more accurate reaction extraction from real figures, and we added App. D.6 and D.7 to show this directly.

Firstly,RxnMiner is ahead on every source, including the figures each baseline was originally developed on:

SystemAll (603)ChemEAGLE set (145)OpenChemIE set (60)Rxnminer unique set (398)
RxnMiner (Qwen)70.877.979.566.9
ChemEAGLE25.038.442.017.5
OpenChemIE5.61.534.62.7

Much of the gap comes from substrate-scope figures, where every product has to be paired with its own label and reactants. OpenChemIE struggles with more complex substrate-scope figures. The three ChemEAGLE examples below illustrate two assembly errors: assigning products to the wrong reaction role and merging distinct experiments into one record (App. D.6).

Figure (App. D.6)ReactionsChemEAGLE RF1RxnMiner RF1
Cyclopropane scope460.00100.00
Cross-coupling scope510.0098.04
Ligand screening380.00100.00
AuthorsOpenReview ↗

[Response 2/3] - Reconstruction Examples and Component Tests

RxnMiner generalizes better across different types of reaction figures. RxnMiner handles the following cases, which both baselines get wrong (App. D.6):

CaseOpenChemIEChemEAGLERxnMiner
Ring size given as a repeat count (12-membered ring)6-membered6-membered12-membered
The same change needed at two sitesOne siteOne siteBoth sites
Color-coded fragment originsFragments swappedFragments swappedCorrect assignment
Aromatic precursors (Ar¹ and Ar²)Rings read as saturatedRings read as saturatedAromatic rings recovered
Nested aryl substituentScaffold misreadTemplate could not be parsedCorrect

To isolate the substitution step itself, we also gave our engine and the OpenChemIE / ChemEAGLE functions exactly the same inputs (App. D.7):

TestRxnMinerOpenChemIE / ChemEAGLE
Stereocentres and double-bond geometry kept after substitution (6 probes)6 / 62 / 6
Same result for 24 equivalent atom orderings of one stereoisomer24 / 2412 / 24
R group "somewhere on a ring" (ChemEAGLE's own Fig. 6d), correct connectivity5 / 6cannot be expressed

Two design choices explain these results. RxnMiner uses MolParser's Extended SMILES representation to encode ring-position-variable groups, whereas the tested baseline interfaces require fixed attachment atoms (App. C.1). And our engine keeps each atom's neighbour order when it substitutes, so wedge bonds survive. When reconstruction does fail, the reason goes back to the model, which can rebuild the structure from the figure or mark it as unresolved instead of returning the generic template. To be fair about our own limits: on Fig. 6d only 2 of the 6 products pass a strict round trip, because of a stereochemistry issue elsewhere in our pipeline.

In short, RxnMiner goes beyond existing systems by supporting more complex R-group substitutions, preserving stereochemistry, and reconstructing complete reactions more accurately through an integrated chemistry toolchain. We have made this clearer in the revised paper.

AuthorsOpenReview ↗

[Response 3/3] - Runtime, Tokens, and Tool Calls

Q1: Can the authors report average inference time, token consumption, the number of tool call to assess its practical efficiency?

Thank you for this practical question. Full statistics are in App. D.8 (Tables 18 and 20).

SystemRF1Time / fig, sTool calls / figTokens in / out per figCost / fig ($)Cost / correct reaction ($)
RxnMiner (Qwen)70.81,34464.84.74 M / 64.7 k0.590.06
RxnMiner (Kimi K3)69.01,62368.33.00 M / 45.9 k1.930.21
RxnMiner (DeepSeek)63.61,91578.76.30 M / 120.9 k0.670.08
Bare Pi (Qwen)59.61,13445.02.73 M / 60.7 k0.330.04
Vanilla Qwen20.8247–6.5 k / 16.0 k0.0070.002
ChemEAGLE (GPT)25.0547–74.4 k / 16.3 k0.55≈0.19
OpenChemIE5.65–local models––

RxnMiner is slower than ChemEAGLE (about 22 vs. 9 minutes per figure), but costs about the same per figure and about one third as much per correctly extracted reaction. Note that ChemEAGLE runs on GPT, so the backbones differ, and our Qwen runs were served locally and priced at the hosted Qwen rate.

We thank the reviewer again for the valuable comments. We believe these clarifications and new experiments strengthen the paper, and they are included in the revised version.

Reviewer_mCyfOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose two things: a manually annotated chemistry figure to structured reaction information dataset, and a agent loop developed based on Pi-agent for complex chemical figure information extraction. The authors evaluates the results across multiple version of VLMs and also make ablation studies to isolate the contribution of the proposed agent loop. The performance is good.

Critical strengths
  • The work is mostly solid, the proposed dataset can greatly improve the development of this field.
  • The writing is clear and easy to follow.
Critical weaknesses

Some evaluation details still required further discussion, see the questions for details. And the visualization is too dense, making the reading difficult.

Questions for authors
  • According to the paper, the dataset are already extracted into a structured data, but why the author still use LLM as judge for reaction condition and the reaction yield. Is that intended to address cases where SMILES, abbreviations, and various common/trivial names are used interchangeably in the reaction condition? What about the yield?
  • Can the authors provide the cost comparison of different pipeline and apis (how many tokens/dollars should cost of the benchmarked api ×\times pipeline combinations). The will be helpful for better clarification on the practical value of the proposed pipeline.
  • Is there any possible to propose a versions with a structed condition formatting, like converting the reagents into SMILES and other fields unified? (This is a minor question, just for discussion.)
4 comments
AuthorsOpenReview ↗

[Response 1/2] - Why We Use an LLM Judge

Dear Reviewer,

We sincerely thank you for the positive assessment and for recognizing the value of the dataset for this field. Below we answer each question; the new material is in the revised PDF and marked in red. We also take the point about dense visualizations and will make the figures easier to read in the final version.

Q1: According to the paper, the dataset are already extracted into a structured data, but why the author still use LLM as judge for reaction condition and the reaction yield? Is that intended to address cases where SMILES, abbreviations, and various common/trivial names are used interchangeably in the reaction condition? What about the yield?

Thank you for this insightful question. Your understanding is correct.

All molecule matching is rule-based, and the main tables use only these rule-based scores. Reactants and products are compared as canonical structures, including stereochemistry. The LLM judge evaluates structures, conditions and outcomes under a complementary rubric; it does not alter the primary scores.

We use a judge there because the same fact can be written in many ways, and several of these cannot be settled by exact matching. Some real cases from the benchmark (App. B.3):

ProblemReal exampleWhy exact matching fails
Different names for one speciesDEA vs. diethylamine; dppf vs. its full nameA dictionary covers common aliases, but models also add names or SMILES that are not in the figure
Distinguishing ordinary and large-scale experimentsAn ordinary experiment reports 89% yield, while a separate scale-up run reports 86% yield and 1.331 gThe same major product appears in both experiments, but they must be treated as separate records; product mass must not be confused with yield
Yield, selectivity, and structural stereochemistryOne entry reports 78% yield, 97% ee, and 9:1 dr, with a specified C=N E configurationSelectivity text, molecular stereochemistry, and reagent amount require separate credits and evidence
Reactant or conditionAn alkyl halide printed above the arrowIts atoms end up in the product, so it is a reactant

Yields have a similar problem: the outcome is often more than one number. A figure may report 93% ee, a 10:1 dr, or "86%, 1.331 g" for a scale-up run. The ee is not a yield and the scale-up is a separate record, so the judge checks that each value is tied to the right quantity and the right entry.

We have added these cases to App. B.3.

AuthorsOpenReview ↗

[Response 2/2] - Cost and Structured Conditions

Q2: Can the authors provide the cost comparison of different pipeline and apis (how many tokens/dollars should cost of the benchmarked api pipeline combinations)?

Thank you for this practical suggestion. Full cost statistics are now in App. D.8 (Tables 18 and 20). our Qwen runs were served locally, so their cost is what the same tokens would cost on the hosted Qwen API.

PipelineRF1Time / fig, sTool calls / figTokens in / out per figCost / fig ($)Max cost / fig ($)Cost / correct reaction ($)
RxnMiner, Qwen70.81,34464.84.74 M / 64.7 k0.592.500.06
RxnMiner, Kimi K369.01,62368.33.00 M / 45.9 k1.937.710.21
RxnMiner, DeepSeek63.61,91578.76.30 M / 120.9 k0.674.280.08
RxnMiner, Doubao46.91,24741.22.84 M / 42.9 k1.014.740.16
RxnMiner, GLM41.12,78647.11.37 M / 187.0 k0.150.380.04
Bare Pi, Qwen59.61,13445.02.73 M / 60.7 k0.334.840.04
Vanilla Qwen20.8247–6.5 k / 16.0 k0.0070.030.002
ChemEAGLE (GPT)25.0547–74.4 k / 16.3 k0.552.30≈0.19

With the open Qwen backbone, a correctly extracted reaction costs about $0.06, roughly one third of ChemEAGLE's cost at 2.8 times its accuracy. "Per correct reaction" divides total cost by the number of recovered reactions, weighting each figure by its F1.

Q3: Is there any possible to propose a versions with a structed condition formatting, like converting the reagents into SMILES and other fields unified?

Thank you for the suggestion. We agree this would be valuable and see it as natural future work.

Our records already keep each condition's role, its printed text, an optional SMILES and its amount. App. B.3 now sketches a fuller format: identity (a normalised name or SMILES next to the printed text), role and scope (catalyst, ligand, additive or solvent, and which steps and entries it covers), amount (equiv, mol%, mass, volume, concentration) and outcome (yield, ee, er, dr, E:Z or scale, each linked to its product).

It is harder than it looks, though. Figures rarely say what role a species plays, some species (metal complexes, ferrocenes, molecular sieves) have no clean SMILES, and the same reagent goes by different names across papers. So we expect a hybrid scorer to work best: rules for well-defined fields such as amounts and yields, and a judge for the rest.

We thank the reviewer again for the encouraging comments and helpful questions, and we will include these additions in the revised paper.

No more questions, i will keep my positive score

We’re glad our response addressed your questions. Thank you for taking the time to follow up and for your encouraging feedback on our work.

Reviewer_4Sa2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The main idea of the paper is solving a problem of extracting specific reaction data from complex figures in the chemistry literature for machine learning tasks such as reaction prediction and condition recommendation. A single figure is often used to describe many experiments using shared templates, footnotes, tables. The authors present three contributions:

  1. RxnMiner – a multimodal agent, that combines a VLM backbone with eight tools and expert informed instructions;
  2. ChemRxn-Bench, a benchmark of 603 figures with 8,547 annotated records, evaluated primarily by macro-average per-figure reaction F1;
  3. A corpus of 88,202 records extracted without human validation. With a locally deployed Qwen3.8-Flash-Next backbone, reaction macro-F1 increases from 21.0 (vanilla) to 59.6 (Bare Pi), and then to 71.0 (RxnMiner). Meanwhile, the strongest of the eight existing systems, ChemEAGLE, reaches only 25.0. Across five evaluated VLM backbones, RxnMiner consistently outperforms Bare Pi by 9.4 to 25.5 percentage points.
Critical strengths
  1. Precise and reusable evaluation protocol. ChemRxn-Bench evaluates the extraction of concrete entry-level records by the exact match of reactants and products using stereochemistry, with one-to-one matching. Failed runs are counted as zero, and tasks are packaged in containers with a published scorer. Previous systems were evaluated on incompatible datasets and matching rules, so this allows for a consistent comparison and shows that existing systems recover few complete records (the best F1 is 25.0).
  2. Design that separates reading, association and deterministic reconstruction. Drawn structures are read by OCSR, the VLM associates conditions and substituents with individual entries, and templated or undrawn reactants are assembled deterministically. With a locally deployed Qwen3.8-Flash-Next backbone, reaction macro-F1 rises from 21.0 (vanilla) and 59.6 (Bare Pi) to 71.0 (78.0 without stereochemistry), and the gain over Bare Pi holds on all five backbones. On 10,107 RxnCaption-11k figures the agent produced 88,202 records, which shows it can scale, although this corpus has not been validated.
Critical weaknesses
  1. The same team created the agent and the benchmark, and the division into development and testing is not described. LESSONS.md consists of error prevention rules, the source data of which is not specified.
  2. Component ablation: In the main text and Table 2, the comparison is only between Vanilla LLM, Bare Pi and the full RxnMiner package. In Appendices B.3 and D.2, the authors explicitly acknowledge that the Bare Pi comparison tests the expert instructions and the chemistry toolbox together and does not isolate the contribution of individual tools. It is impossible to understand from the paper how much each component contributes (the OCSR tool MolParser, the algorithmic R-group substitution, the stereochemistry visual verification step).
Questions for authors
  1. How were SKILL.md and LESSONS.md developed? What figures, if any, were used to create rules in LESSONS.md and have any of them been used in ChemRxn-Bench? If so, could you please report the results for a subset that was not used in the development?
  2. Can you provide a component ablation: for example RxnMiner without the R-group engine, without LESSONS.md, and with an alternative OCSR?
  3. Figures and annotations from ChemEAGLE and OpenChemIE are public and may be in the backbones’ pretraining data. Could you report results separately for these sources and for the remaining figures?
  4. What are the average tokens, tool calls and wall-clock time per figure? What are the timeout and failure rates per backbone?
  5. How many figures are excluded as unsolvable without the context of the article? What is the agent doing in such figures: abstaining or hallucinating?
  6. What tools did Bare Pi have in the container? Did it have access to MolParser and RDKit?
6 comments
AuthorsOpenReview ↗

[Response 1/5] - Development and Held-Out Evaluation

Dear Reviewer,

We sincerely thank you for the careful and constructive review, and for recognizing the value of a precise, reusable evaluation protocol for this task. Below we respond to each point. All new analyses are in the revised PDF and marked in red.

W1/Q1: How were SKILL.md and LESSONS.md developed? What figures, if any, were used to create rules in LESSONS.md and have any of them been used in ChemRxn-Bench? If so, could you please report the results for a subset that was not used in the development?

Thank you for this important question.

We use a development set and a held-out test set. Of the 603 benchmark figures, 308 were used during development and the remaining 295 were not used in development at all. On the development figures we ran the agent, looked at its errors, and improved the harness and the procedure files.

No reference answers ever enter the agent. SKILL.md and LESSONS.md contain only general instructions, with no figure IDs, structures or answers, and a lint check rejects all three. So a development figure can only influence the agent through a rule that would apply to any figure.

In fact, most of the development effort went into the harness rather than the instruction files. Each change addressed a failure that kept recurring across figures: MolParser readings are cached with their image regions and re-read from a crop when a structure is missed; the R-group engine builds substituted structures deterministically, and the assembler checks each record (e.g. stereocentres) and tells the model why a step failed; context compaction lets the agent continue on long tables, and a wrap-up step near the time limit saves partial work; and every molecule records whether it was read by a tool, derived from a template, or written by the model.

Results on the two splits (strict reaction F1, %):

SystemDevelopment (308)Test (295)All (603)
RxnMiner (Qwen)73.268.370.8
Bare Pi (Qwen)60.758.559.6
Vanilla Qwen21.919.720.8
ChemEAGLE24.825.225.0
OpenChemIE7.04.15.6

On the held-out test figures RxnMiner is still clearly ahead, and its margin over Bare Pi with the same backbone is similar on both splits (+12.5 and +9.8).

As for LESSONS.md, it collects general mistakes we saw the agent make repeatedly, such as inventing compound IDs or misreading alkyl groups. To be honest, the ablation below shows that it contributes little: repeating such warnings to the model does not change its behaviour much. The general workflow in SKILL.md, the tools and the harness matter more.

We have added the split, these results and the history of the procedure files to App. D.4 and C.2.

AuthorsOpenReview ↗

[Response 2/5] - Component Ablations

W2/Q2: Can you provide a component ablation: for example RxnMiner without the R-group engine, without LESSONS.md, and with an alternative OCSR?

Thank you for this valuable suggestion. We ran the requested ablation (App. D.5). Each configuration removes or replaces one component and was run once on all 603 figures with a one-hour budget per figure; a figure without an answer scores 0.

IDConfigurationRF1Change vs. full
A0Full RxnMiner63.44–
A1Without R-group tools61.02−2.43
A2Without LESSONS.md64.14+0.70
A3Without stereo check65.31+1.87
A4Without MolParser46.89−16.55
A5MolScribe in place of MolParser54.38−9.07
A6Without LESSONS.md and SKILL.md62.19−1.25

Structure recognition matters most: removing MolParser costs 16.6 points, and swapping it for another OCSR model (MolScribe) costs 9.1.

A2 and A3 come out slightly above the full system. Both skip checking work, so under a one-hour cap more figures finish in time. To check this, we reran A0 to A3 with three hours on the 200 figures where at least one of them had run out of time (App. D.5, Table 15):

Same 200 figures1 h RF1vs. full, 1 h3 h RF1vs. full, 3 h
A0 Full RxnMiner45.8–72.1–
A1 Without R-group tools45.2−0.668.4−3.7
A2 Without LESSONS.md52.5+6.768.3−3.8
A3 Without stereo check47.5+1.771.8−0.3

With enough time the full system is ahead of every variant, so the one-hour gains of A2 and A3 come from the time limit, not from the components being harmful. The stereo check has no measurable effect on the score either way.

A1 understates the R-group engine, because without the dedicated tools the agent can still edit structures with RDKit. So we also tested the engine directly: App. D.6 compares the full systems on 10 figures, and App. D.7 gives our engine and the OpenChemIE / ChemEAGLE functions identical inputs (e.g. stereochemistry kept in 6 of 6 probes vs. 2 of 6).

Each configuration was run once, so we do not read much into differences of one or two points.

AuthorsOpenReview ↗

[Response 3/5] - Results by Figure Source

Q3: Figures and annotations from ChemEAGLE and OpenChemIE are public and may be in the backbones' pretraining data. Could you report results separately for these sources and for the remaining figures?

Thank you for raising this. The results by source are below (App. D.4, Table 12):

SystemChemEAGLE set (145)OpenChemIE set (60)Rxnminer unique set (398)
RxnMiner (Qwen)77.979.566.9
RxnMiner (Kimi)76.979.064.6
RxnMiner (DeepSeek)72.277.258.5
Bare Pi (Qwen)67.760.156.6
Vanilla Qwen27.822.118.0
ChemEAGLE38.442.017.5
OpenChemIE1.534.62.7

We agree exposure may play a role, but the public figures are also simpler than the Rxnminer unique figures, which explains part of the gap.

Importantly, RxnMiner's improvement does not depend on the public sets. On all three benchmarks Rxnminer still clearly beats every baseline:

RxnMiner (Qwen) vs.Gain on Rxnminer unique set
Bare Pi (Qwen)+10.3
Vanilla Qwen+48.9
ChemEAGLE+49.4

Its gain over Bare Pi is about the same on public and remaining set (difference 2.6 points).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.