{"id":"5399279e-52e7-46bb-a3f0-67952da7533f","arxiv_id":"2501.13824","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Adding hallucinated molecule descriptions to prompts improves ROC-AUC for several LLMs on molecular property prediction, with GPT-4o-generated text giving the largest consistent gains.","lead":"Adding made-up molecule descriptions generated by AI chatbots to a drug-property question can improve the chatbot's accuracy on some benchmark tasks, even compared with factual descriptions. The result challenges the usual assumption that hallucinations are always harmful and could change how AI is used in early-stage drug discovery.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported AUC gains are single-run point estimates on very small test sets; without error bars or significance tests, the headline improvements may be noise.","rationale":"I read the paper in good faith. The experimental setup is clear, the tables are transparent, and the idea is interesting. The core issue is not the novelty or the framing but the evidential weight of the headline numbers. The reader's weakest assumption (hallucination measured against MolT5) is a valid interpretive concern: the paper labels descriptions as hallucinated based on HHEM-2.1-Open agreement with MolT5, not chemical ground truth. However, even if that proxy were perfect, the central empirical claim would still need the performance differences to be statistically credible. They are not, as reported. The paper gives no measures of uncertainty, and three test sets are small enough that AUC estimates have wide intervals. The abstract's 'outperforms all baselines' wording is internally contradicted by the paper's own full table, which suggests the point estimates are being read too strongly. I therefore recommend keeping the reader's CONDITIONAL verdict: the paper should be accepted only if the authors supply bootstrap confidence intervals or multi-seed runs demonstrating that the headline gains are not noise. The hallucination-typing claim about structural misdescriptions is also circular as the reader notes, but it is secondary; the performance numbers are the load-bearing element. My agreement with the reader is partial because I differ on which assumption is the weakest.","tokens_in":21870,"tokens_out":9266,"duration_ms":84565,"concrete_test":"Using the reported test sets and the same models/prompts, reconstruct the per-instance predictions for Table 7 (author code/data would make this direct), then compute a stratified bootstrap (1,000 resamples) over test instances to obtain 95% CIs for average ROC-AUC and for the paired differences vs SMILES, MolT5, and PubChem. Run DeLong's paired test on the two largest datasets (HIV, Tox21). The central claim survives only if the Falcon3+GPT-4o vs PubChem and Llama-3.1+Llama-3 vs SMILES CIs exclude zero and the corresponding p-values are below 0.05 after multiple-comparison correction. If the CIs overlap zero, the headline improvements are not statistically distinguishable from chance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim is that hallucinated descriptions improve ROC-AUC, with headline gains such as Falcon3-Mamba-7B +8.22% over PubChem and Llama-3.1-8B +15.80% over SMILES. These numbers are single-run point estimates. Table 5 shows that three of the five test sets are very small: BBBP 205 instances, Clintox 148 instances (11 positives), SIDER 143 instances. For Clintox, the difference between 68.08 and 35.30 ROC-AUC corresponds to a handful of re-ranked examples; the standard error of AUC on 11 positives is large. The reported 'average' gains are therefore vulnerable to one or two noisy datasets. No confidence intervals, bootstrap estimates, DeLong tests, or multiple-seed runs are reported anywhere in the paper. The abstract even overstates the result: 'Falcon3-Mamba-7B outperforms all baselines when hallucinated text is included' is contradicted by Table 7, where Falcon3 with Ministral-generated text averages 44.50, below the PubChem baseline of 45.47. This inconsistency reinforces the fragility of the point estimates. If the headline differences are within noise, the paper's central claim—that hallucinations help—is not established, regardless of the hallucination-measurement proxy. The hallucination-proxy issue raised by the reader is important for interpretation, but the statistical reliability of the performance numbers is the more load-bearing link: it is the foundation on which the rest of the argument rests.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper asks whether hallucinated natural-language descriptions of molecules can improve LLM-based molecular property prediction. For each SMILES string, the authors prompt one of seven LLMs to generate a free-text description, append it to a binary classification prompt, and compare the resulting ROC-AUC against three non-LLM description settings: SMILES-only, MolT5 reference descriptions, and rule-based PubChem metadata. Across five MoleculeNet datasets, they report that some models benefit substantially, e.g., Falcon3-Mamba-7B improving by 8.22% over PubChem with GPT-4o descriptions and Llama-3.1-8B improving by 15.80% over SMILES with Llama-3 descriptions. The paper also proposes a four-type hallucination taxonomy, uses an external LLM to label beneficial hallucinations, and reports ablations over model size and generation temperature.","tokens_in":22096,"tokens_out":8402,"duration_ms":69411,"significance":"The question is timely and the experimental setup is broad: seven instruction-tuned models, five datasets, three baselines, and a separate evaluation of which hallucination source helps most. The core comparison against SMILES, MolT5, and PubChem involves no fitted parameters and is therefore not circular, and the hallucination taxonomy with a human-agreement check is a useful contribution. If the performance differences survive uncertainty quantification and the hallucination measurement is validated against actual molecular structure, the result would be a valuable empirical finding for prompt engineering in scientific LLM use. As it stands, the evidence is suggestive rather than conclusive, because the headline numbers are single-run point estimates, the main table is selected after seeing performance, and the term 'hallucination' is attached to divergence from a neural reference rather than to verified chemical inaccuracy.","major_comments":[{"comment":"All reported ROC-AUC values are single-run point estimates with no confidence intervals, bootstrap estimates, DeLong tests, or repeated seeds. This matters because Table 5 reports very small test sets: BBBP has 205 instances, Clintox has 148 instances with only 11 positives, and SIDER has 143 instances. For example, Llama-3.1-8B's Clintox score moves from 35.30 with SMILES to 68.08 with Llama-3 descriptions; with 11 positives, a few re-ranked examples can produce this difference, and the standard error of AUC is large. Since the headline gains (e.g., +15.80% over SMILES for Llama-3.1-8B and +8.22% over PubChem for Falcon3-Mamba-7B) are averages over five datasets, one noisy small dataset can drive the average. The paper should provide per-dataset uncertainty estimates, significance tests, or repeated runs before claiming that hallucinations 'significantly improve' predictive accuracy. In addition, SIDER and Tox21 use a single label selected for balance, and no sensitivity analysis for that choice is reported.","section":"Table 5 / Table 7"},{"comment":"The main results table is curated after seeing performance. Table 2 is captioned 'curated set of descriptions,' and the text states that the hallucination source shown for each model category is the one whose generated hallucinations 'resulted in the greatest average performance gains in ROC-AUC.' Thus the headline comparisons are selected from the full results in Table 7 rather than following a fixed, pre-specified protocol, which can inflate the apparent benefit of hallucinated text. The abstract's claim that 'Falcon3-Mamba-7B outperforms all baselines when hallucinated text is included' is also contradicted by Table 7: with Ministral-generated descriptions, Falcon3-Mamba-7B averages 44.50, below the PubChem baseline of 45.47. The paper should present the full results as primary or clearly label the selected results as best-case, and it should temper the abstract accordingly.","section":"Table 2 / 'Do Hallucinations Improve LLMs?'"},{"comment":"The definition of hallucination used in the paper is not the quantity that is measured. The Hallucination Generation section defines hallucination as 'any information in the generated text that is not supported by the molecule's actual structure or known properties,' but Figure 1 and Table 6 measure factual consistency with HHEM-2.1-Open relative to MolT5 reference descriptions. MolT5 is a generative model, not ground-truth chemistry, so divergence from MolT5 is not evidence of hallucination against the molecule; a description can diverge from MolT5 and still be chemically accurate. Therefore statements such as 'the majority of generated descriptions contain some degree of hallucination' and the low factual-consistency scores in Figure 1 are not established by the experiment. The authors should either verify descriptions with a structure-aware chemical tool or explicitly reframe the investigation as the effect of descriptions that diverge from a neural reference.","section":"Hallucination Generation / Figure 1"},{"comment":"The beneficial-hallucination analysis is selection-on-outcome and does not support the causal claim about hallucination types. In 'What Types of Hallucinations Are Most Useful?', beneficial hallucinations are defined as instances that lead to correct predictions with higher confidence than any baseline; Figure 4 then reports that 85% of these beneficial cases are structural misdescriptions. Because the set is selected on the outcome, the type distribution among beneficial cases cannot show that structural misdescriptions are 'the most impactful type' unless it is compared with the type distribution among all generated descriptions or among non-beneficial ones. The annotation quality check is also limited: Fleiss' kappa of 0.57 on 50 instances is moderate, and no reliability is reported for the large-scale Deepseek-R1 annotations. This section should be reworded as a descriptive analysis of selected cases, with appropriate comparisons and uncertainty.","section":"What Types of Hallucinations Are Most Useful? / Figure 4"}],"minor_comments":[{"comment":"The scoring model name is inconsistent: 'HHM-2.1-Open' appears in the Hallucination Generation section and Table 6, while Figure 1 uses 'HHEM-2.1-Open.' The correct name should be used consistently.","section":"Hallucination Generation / Table 6"},{"comment":"The prompt templates render '[SM ILES]' and '[HALLU CIN AT ION]' with inserted spaces; these appear to be LaTeX line-breaking artifacts and should be corrected.","section":"Appendix: Prompt Template"},{"comment":"The HuggingFace link for Llama-3-8B points to the Llama-3.1-8B-Instruct repository; the link should be verified and corrected.","section":"Table 4"},{"comment":"The text says that hallucination scores increase as temperature increases, but Table 12 is non-monotonic and the 0.9 setting has the lowest HHEM score (6.82). Please reconcile Figure 6 with Table 12 or define explicitly whether 'hallucination score' increases with factual consistency or with divergence.","section":"Effect of Generation Temperature"},{"comment":"No code, generated descriptions, or data artifacts are made available. Given that the entire analysis depends on prompt templates and generated text, releasing these would substantially improve reproducibility.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The reader's and skeptic's concerns are well founded, and I agree that statistical reliability is the most load-bearing issue rather than the hallucination-proxy debate alone. The paper is a reasonable fit for the journal and the central question is interesting, but the headline claim currently rests on selected single-run point estimates. A major revision that adds uncertainty quantification, reports the full results without outcome-based curation, and validates or reframes the hallucination measurement would make the contribution publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a genuinely interesting idea — conditioning a property-prediction LLM on free-text molecule descriptions, even when those descriptions are factually wrong, can move ROC-AUC for some model/generator pairs. That is worth a serious look. But the paper's own tables undercut its headline claims, and the evidence as presented is not statistically solid.\n\nWhat's new: as far as I know this is the first systematic study of hallucinated text as prompt augmentation for molecular property prediction, and the four-category typology (structural misdescription, functional hallucination, analogical hallucination, generic fluff) is a useful vocabulary. Table 7 is a real contribution: it gives the full 7x7 matrix of generators and predictors plus baselines, so a reader can see at a glance that gains are model-specific. The ablation on model size and temperature is a nice addition.\n\nSoft spots. The most load-bearing problem is statistical. Every reported AUC is a single run with no confidence interval or significance test. Three of the five test sets are tiny: BBBP 205, Clintox 148 (11 positives), SIDER 143. On Clintox, a difference between 68.08 and 35.30 can come from a handful of re-ranked examples; the standard error is large. The abstract says 'Falcon3-Mamba-7B outperforms all baselines when hallucinated text is included,' but Table 7 shows that with Ministral-generated text Falcon3 averages 44.50, below PubChem's 45.47. That overstatement needs fixing. The 'beneficial hallucination' analysis is circular in a mild but real way: cases are selected because they improved predictions, and then the distribution of hallucination types in those cases is reported as though it explains the improvement, without comparing to the distribution in non-improving cases. Finally, the hallucination measurement is divergence from MolT5 reference descriptions, not from ground-truth chemistry — so 'hallucination' is really 'disagrees with MolT5.' The authors acknowledge the proxy verbally but never verify descriptions against actual structure.\n\nNone of these are fatal. The core phenomenon may well be real. But the paper currently supports a modest claim — some models benefit from some hallucination sources — and the big numbers in the abstract are not yet supported.\n\nWho is this for? Researchers working on prompt-based LLM applications in chemistry and anyone interested in the 'useful hallucination' debate. It deserves a serious referee, but a referee should demand multiple seeds, error bars, and a rewrite of the abstract so it matches Table 7.","headline":"Interesting and transparent, but the headline gains are single-run point estimates on tiny test sets and the abstract overstates what the paper's own tables show.","tokens_in":22689,"tokens_out":2582,"would_cite":true,"duration_ms":21842,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that LLM hallucinations can be harnessed: adding hallucinated molecule descriptions to prompts improves molecular property prediction, sometimes beating factual chemical databases like PubChem.","keywords":["hallucination","molecule property prediction","drug discovery","SMILES","large language models","implicit counterfactuals","ROC-AUC","MolT5"],"falsifier":"Re-run the five benchmark tasks after replacing each hallucinated description with either randomly sampled chemistry-flavored text or with verified descriptions from a trusted database; if the gains persist with random text, the effect is not about hallucination content but merely about adding extra tokens, and the implicit-counterfactual explanation would be unsupported. Alternatively, verify the reported structural misdescriptions against computed molecular structure, and check whether they are actually incorrect in most cases.","tokens_in":21566,"feed_emoji":"🧪","tokens_out":5215,"duration_ms":42948,"temperature":0.7,"pith_summary":"This paper tries to overturn the usual assumption that hallucinations are always bad: it asks whether the plausible-but-wrong molecule descriptions that LLMs generate can actually help the same models predict molecular properties, such as whether a compound inhibits HIV or crosses the blood-brain barrier. The authors prompt seven instruction-tuned LLMs to describe SMILES strings, then feed those descriptions back into the models as extra context for binary property prediction across five MoleculeNet datasets. They report that hallucinated context often beats two clean baselines—SMILES-only prompts and MolT5 reference descriptions—and in some cases even beats structured PubChem metadata. The practical stake is that hallucinated text, treated as a kind of implicit counterfactual, could be a cheap source of inductive bias for early-stage drug discovery where structured knowledge is sparse.","feed_headline":"Hallucinated molecule text can boost LLM drug predictions","feed_subtitle":"Adding hallucinated molecule descriptions to prompts beats factual baselines on several drug-discovery benchmarks.","key_machinery":"The load-bearing mechanism is the prompt template [SMILES][Description][Instruct], where a binary yes/no prediction is decoded by the highest-probability token over {Yes, No}. Hallucination is measured by the HHEM-2.1-Open model, which scores alignment between each LLM-generated description and a MolT5 reference description; low scores indicate high hallucination. A four-category typology—structural misdescription, functional hallucination, analogical hallucination, and generic fluff—is used to label the 18,000+ beneficial hallucinations, and structural misdescription is identified as the most impactful type.","core_discovery":"The central discovery is that hallucinated descriptions, interpreted as implicit counterfactuals, can serve as a useful inductive bias for LLM molecular property prediction. Concretely, Falcon3-Mamba-7B achieves an average ROC-AUC 8.22% higher than the PubChem baseline when given GPT-4o-generated descriptions, and Llama-3.1-8B improves over SMILES by 15.80% and over MolT5 by 11.23% when given Llama-3 descriptions. The authors further collect 18,872 cases where hallucinations improve predictions over all baselines for Falcon3-Mamba-7B and find that 85% of these beneficial hallucinations are structural misdescriptions—incorrect but chemically plausible statements about atoms, bonds, or substructures—rather than functional, analogical, or vague text. Larger models extract more benefit from hallucinated context, while sampling temperature affects factuality but has only a small, non-monotonic effect on downstream performance.","pith_inferences":["The reported gains might not require genuine hallucination: a controlled experiment substituting random chemistry-flavored text for the hallucinated descriptions would reveal whether it is the content or merely the presence of extra text that helps.","Because the paper measures hallucination as divergence from MolT5 rather than from verified structure, the practical recipe may transfer to any domain where a reference description model exists, not just chemistry.","The dominance of structural misdescriptions suggests a testable extension: deliberately perturbing SMILES strings (for example, swapping atoms or bonds) to generate targeted counterfactual descriptions could reproduce the benefit in a more controlled and interpretable way.","If the effect is robust, hallucination could be treated as a tunable source of synthetic data for scientific prediction tasks, with temperature and model size as knobs rather than things to suppress."],"forward_implications":["Adding LLM-generated description text to SMILES prompts can raise ROC-AUC by double digits for some models, beating the SMILES-only and MolT5 baselines.","The source of the hallucination matters: GPT-4o-generated descriptions give the most consistent gains across all seven predictor models.","Among hallucination types, structural misdescriptions drive most of the benefit, suggesting that factually wrong but chemically plausible structural claims help LLM decision-making.","Larger models (up to 8B parameters) extract more signal from hallucinated text, while generation temperature between 0.1 and 0.9 has limited influence on downstream accuracy.","Hallucinated context can, in some configurations, outperform a gold-standard rule-based baseline built from PubChem metadata."],"supporting_citations":[{"why":"Provides the MoleculeNet benchmark and the five datasets (HIV, BBBP, Clintox, SIDER, Tox21) on which all property predictions are evaluated.","marker":"Wu et al. 2018"},{"why":"Supplies the MolT5 molecule-to-text model, both as the reference for measuring hallucination and as one of the three baselines.","marker":"Edwards et al. 2022"},{"why":"Supplies the PubChem database used to construct the rule-based factual baseline descriptions.","marker":"Kim et al. 2025"},{"why":"Provides the HHEM-2.1-Open model used to score factual consistency between generated descriptions and MolT5 references, the paper's operational definition of hallucination.","marker":"Bao et al. 2024"},{"why":"Defines the SMILES representation that is the input format for both description generation and property prediction.","marker":"Weininger 1988"},{"why":"Introduces the Llama-3 model family, used in the model-size ablation and as generator/predictor in the main experiments.","marker":"Dubey et al. 2024"},{"why":"Provides DeepSeek-R1, the external LLM used to annotate hallucination types on 18,000+ beneficial hallucination instances.","marker":"Guo et al. 2025"},{"why":"Shows that textual molecule descriptions can improve generalization, the basis for conditioning property prediction on natural language descriptions.","marker":"Liu et al. 2023"}],"fun_headline_variants":["Hallucinations boost LLM drug discovery predictions","LLM hallucinations improve molecule property prediction","Falcon3-Mamba beats baselines with hallucinated text","Structural misdescriptions help LLMs predict drug properties","Hallucinated molecule descriptions lift LLM accuracy by 8%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper equates hallucination with divergence from MolT5 reference descriptions as scored by HHEM-2.1-Open, but MolT5 is itself a generated model rather than verified ground-truth chemistry, so what is called hallucination is really dissimilarity to one particular reference text.","fun_headline_variants_meta":{"raw":{"variants":["Hallucinations boost LLM drug discovery predictions","LLM hallucinations improve molecule property prediction","Falcon3-Mamba beats baselines with hallucinated text","Structural misdescriptions help LLMs predict drug properties","Hallucinated molecule descriptions lift LLM accuracy by 8%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00028,"raw_usage":{"total_tokens":1659,"prompt_tokens":945,"completion_tokens":714,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":561,"completion_tokens_details":{"reasoning_tokens":636}},"tokens_in":561,"tokens_out":714,"duration_ms":5945,"temperature":1.0,"reasoning_tokens":636,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:34:09.089395+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the five benchmark tasks after replacing each hallucinated description with either randomly sampled chemistry-flavored text or with verified descriptions from a trusted database; if the gains persist with random text, the effect is not about hallucination content but merely about adding extra tokens, and the implicit-counterfactual explanation would be unsupported. Alternatively, verify the reported structural misdescriptions against computed molecular structure, and check whether they are actually incorrect in most cases.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MolT5 molecule-to-text model, both as the reference for measuring hallucination and as one of the three baselines."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the PubChem database used to construct the rule-based factual baseline descriptions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the HHEM-2.1-Open model used to score factual consistency between generated descriptions and MolT5 references, the paper's operational definition of hallucination."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that textual molecule descriptions can improve generalization, the basis for conditioning property prediction on natural language descriptions."}],"review_version":1}