acceptodds
Under review as a conference paper at ICLR 2027

Predicting Unseen Scientific Interactions: Conditional Flow Models Hallucinate Non-Additive Effects

Abstract

Conditional generative models are increasingly used for scientific prediction beyond the direct training support. Yet scientific questions often concern whether one intervention changes the effect of another. Such interactions can be of primary interest while explaining little of the total variation, allowing models to appear accurate overall without recovering the relevant scientific effect. It remains unclear how these models learn interactions and when these successfully extrapolate to unseen conditions. We therefore study interaction recovery in flow-matching models using a physiological time-series simulator, drug dose-response data, and combinatorial cell perturbations, where we separate recovery from hallucinated interaction structure. In the controlled setting, individual effects are learned much earlier than interactions and models hallucinate interactions where none exist. Across settings we find qualitatively different generalization problems, with successful prediction depending on whether condition structure continues from observed to held-out targets. When it does, interactions are recoverable. If not, models can generate plausible responses while missing the interaction. We furthermore find that hallucinations increase as targets become less identifiable, and that successful interaction recovery in output space can fail under nonlinear readouts. Finally, we show that aggregate fidelity metrics can conceal whether models learn and extrapolate interactions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
generative models, flow matching, out-of-distribution, scientific prediction, scientific discovery, biology

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_7V2hOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper explores whether conditional flow models can correctly recover non-additive interactions when extrapolating to unseen conditions. Here, the authors argue that aggregate metrics can be misleading, as interaction effects may be much smaller yet highly scientifically relevant. Decomposing predicted interactions into true interaction recovery and an orthogonal hallucinated component, the authors study a simluated setting, drug response data, and interacting single-cell perturbations.

Critical strengths

Strength 1. The paper reads well and presents the problem along with good perspectives. Both the settings and results were communicated effectively.

Strength 2. The study fits timely well. Both flow matching and interactions are increasingly relevant in scientific modelling. This makes understanding the properties of the models important.

Critical weaknesses

Weakness 1. The overall significance is somewhat limited. There are only few interesting results in the paper. Several main observations follow fairly directly from the interaction and identifiability setup. The strongest evidence comes from the controlled simulator while real data provided less convincing examples. I find it unclear whether the paper currently provides enough substance required for acceptance.

Weakness 2. It is unclear how specific the results are to the particular flow-matching recipe considered here (the authors also acknowledge this in the ending paragraph). It would be interesting to see a more general setting.

Questions for authors

Question 1. Beyond evaluation, could the paper guide changes to model design or training?

For a concrete example -- but it is fine to think of this question more generally -- could we make changes in objectives or architectural biases to avoid hallucinated effects?

Question 2. When estimating the interaction targets from finite experimental samples, how sensitive are the metrics for interaction recovery and hallucination to the empirical interaction estimates? For example, the double perturbations seen in Norman et al range from 49 to 1030 cells per condition, so we might expect empirical estimates to greatly vary.

5 comments

We sincerely thank the reviewer for their evaluation. We have run new experiments for each of the raised points, which we address one by one below. First, we highlight the main additions:

  • New real-data setting: for molecular conformers (GEOM), the same molecules are scored under decreasing training support with an equivariant flow model, leading to increased interaction errors (EIE_I: .70 → .83).
  • Additional modeling recipes: endpoint and velocity flows, diffusion, a cVAE, and regressors
  • Mechanism: we provide an investigation showing that as generative models extrapolate away from the training support, mode proportions progressively break down while every sample stays sharp and realistic.
  • Fixing failures: we find that changing models, objectives, or samplers does not restore recovery, whereas new data in the held-out region does (α\alpha: .31 → .81), and find that the presented interaction framing can guide and improve data collection

 

 

Significance, and real data being less convincing (W1)

We agree that some observations follow from the setup: interactions are harder to learn than main effects, and less support makes extrapolation worse. However, two findings are not explained by distance from the training data alone:

  • Models predict interactions when none are present in the data. In the additive simulator (γ=0\gamma=0), both the flow and a matched regressor hallucinate (βflow=.33\beta_\text{flow}=.33, βreg=.18\beta_\text{reg}=.18).
  • Two equally held-out corners behave very differently: the HH interaction is recovered almost fully (α≈1\alpha \approx 1), while LL is not (α=.42\alpha=.42). However, the two corners differ in whether training data determines how the response continues into the corner. Choosing a simple continuation rule via cross-validation (training data only) reconstructs the HH interaction (cosine 1.00) but not LL (cosine .77).

 

For real data, we added a second system (molecular conformers) and a matched design for single-cell (Norman), in which the same held-out targets are scored under decreasing training support.

Real-data testChange in supportInteraction error increase95% CI
GEOM (molecular conformers, new): same 150 query moleculesneighbouring → parent → distant molecules onlyEIE_I: +.05 →\rightarrow +.13+[.04, .25] (distant - neighbour)
Norman: same held-out pairsmore than half of neighbouring training pairs removedEIE_I: +.14+[.07, .21]

In both settings, the interaction error rises as support is removed while the targets stay fixed. As only the training support changes between conditions, these provide matched tests on real data.

 

 

How specific are the results to this flow-matching recipe? (W2)

We varied the recipe in both the simulator and the single-cell setting:

SystemModelEIE_IRecovery α\alphaHallucination β\beta
PPG, LL cornerEndpoint flow (paper).76.42.48
Velocity flow.67.40.30
Diffusion.78.42.51
Regressor.83.34.50
NormanFlow (paper).92.65.78
Diffusion.89.62.71
cVAE.87.58.76
Direct-mean MLP.82.56.69
  • No recipe restores recovery in the LL corner, but they modestly change how much is hallucinated.
  • On Norman, all four models hallucinate at comparable levels.
  • In a new two-mode version of the simulator (kindly see our reply to reviewer f4vV), two flow variants, a diffusion model, and a regressor fail in the same region.

We therefore now present these failures as arising from training data that incompletely constrain the interaction, rather than from one particular recipe. We propose to keep flow models as the main subject of the paper, as they are what is deployed in the fields we study, and because they fail in a distinct way: unlike a regressor, they produce sharp, realistic samples whose interactions are wrong.

Can the paper guide model design or training? (Q1)

We appreciate this question and tested both modelling changes and new data, using the PPG LL corner:

ChangeEIE_IRecovery α\alphaHallucination β\beta
Endpoint flow (manuscript).76.42.48
Velocity flow.67.40.30
Diffusion.78.42.51
Regressor.83.34.50
Additional training and/or data:
Endpoint flow, trained further without new data.82.31.43
Endpoint flow + training on 16 new conditions, chosen at random.50.80.39
Endpoint flow + training on 16 new conditions, targeting response.39.75.30
Endpoint flow + training on 16 new conditions, targeting interaction.32.81.26
  • We find that modeling changes can lead to reductions in hallucination moreso than recovery. We attempted the requested objectives but were not able to achieve consistent gains: interaction-aware contrast and readout losses, and condition embeddings which process each factor separately. We will report these negative results in the appendix. We interpret these preliminary experiments to indicate that extrapolation is mainly determined by data.
  • Indeed, only new data in the held-out region raised recovery (α\alpha: .31 → .80). For the bottom row, we apply the paper's framing to data collection: at each step, we pick the condition that most reduces uncertainty about the interaction component, rather than the response (which yields EI=.39E_I=.39 instead of EI=.32E_I=.32). Here we use a Gaussian process over the two condition coordinates (only where training data exists). The α/β\alpha / \beta decomposition shows that the advantage over random placement is due to lower hallucination (β=.39→.26\beta=.39 \rightarrow .26).

 

 

Sensitivity to the empirical interaction estimates (Q2)

We thank the reviewer for this suggestion and agree this is an important control. We estimated the noise in each measured interaction from 200 random half-splits of that pair's cells, and provide a noise floor (via a model predicting the true means). We then split all 131 pairs into thirds by the number of measured cells:

Measured cells per doublePairsModel β\betaNoise-floor β\betaModel α\alphaNoise-floor α\alpha
49–20544.87.42.46.76
208–29143.90.35.60.86
292–103044.93.29.69.91
  • Hallucination does not come from the poorly measured pairs, as β\beta is fairly stable across each third, while more cells leads to a lower noise floor.
  • Recovery is lower for poorly measured pairs, which results from target noise diluting the measured interaction.
  • We find the number of generated cells to not matter strongly. Moving from 192 to 768 changes β\beta by .02 and α\alpha not at all (kindly see reply to reviewer x7eF for the full table).

Thanks, I actually think these are some interesting experiments. I however believe you need to work hard on making the additions more rigorous across recipes, as they don't yet feel completely like general takeaways yet. I believe it might be possible to do so for a camera ready, so I now lean towards weak accept.

Thanks for the effort, I hope you will continue working on refining it more!

Reviewer_x7eFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The study investigates the ability of conditional flow models to estimate interactions between two factors. Tests performed on synthetic signals, responses to drugs, and gene perturbations suggest that even good overall predictions may include false interactions. The process of interaction learning may differ from learning individual effects.

Critical strengths
  • When dealing with scientific applications, the relationship between two treatments can be more important than the effect itself. The paper shows several good examples of cases where the predictive score hides issues related to this interaction.

  • The controlled simulator is useful because it changes interaction strength while keeping individual effects fixed. False interactions in the additive setting are particularly interesting.

Critical weaknesses
  • The contribution is mainly experimental, but model comparisons are limited. Although regression is evaluated on drug-response data, comparable regression controls for the simulator and single-cell experiments, together with other generative models, would help establish how broadly the findings apply.

  • The single-cell comparison uses different pairs: 113 graph-supported versus 18 cold-start pairs. Differences in pair difficulty, interaction strength, or cell counts could explain the results. Testing the same held-out pairs with and without supporting training pairs would be more convincing. Seeing each gene in other pairs also does not, by itself, uniquely determine the interaction between two genes in an unseen combination.

Questions for authors
  • In addition to the drug-response baseline, do you have any regression or other generative-model results for the main findings, particularly the false interactions in the additive simulator setting?

  • Were any held-out pairs of genes tested both with and without other pairs of genes that they appear in during training? Otherwise, how is the impact of the training data distinguished from differences in the two sets of gene pairs?

  • How much of the single-cell “hallucination” can be attributed to variability in the average responses? Do your findings change if the number of generated single cells or uncertainty calculation changes?

4 comments

We sincerely thank the reviewer for their evaluation and their concrete suggestions. We have run new experiments for each of the raised points, which we address one by one below. First, we highlight the main additions:

  • New real-data setting: for molecular conformers (GEOM), the same molecules are scored under decreasing training support with an equivariant flow model, leading to increased interaction errors (EIE_I: .70 → .83).
  • Additional modeling recipes: endpoint and velocity flows, diffusion, a cVAE, and regressors
  • Mechanism: we provide an investigation showing that as generative models extrapolate away from the training support, mode proportions progressively break down while every sample stays sharp and realistic.
  • Fixing failures: we find that changing models, objectives, or samplers does not restore recovery, whereas new data in the held-out region does (α\alpha: .31 → .81), and find that the presented interaction framing can guide and improve data collection

 

 

Regression and other generative models, especially false interactions in the additive simulator (W1, Q1)

Comparison with regressor (trained using same model, data, and training budget)

  • For PPG, we find the regressor fails in a similar manner to generative models. It hallucinates in the additive regime (γ=0\gamma=0, with βreg=.18,βflow=.33\beta_\text{reg}=.18,\beta_\text{flow}=.33) and similarly recovers interactions only in one of the two held-out region of condition space (failed corner at αreg=.34,αflow=.42\alpha_\text{reg}=.34, \alpha_\text{flow}=.42). We note it both hallucinates and recovers modestly less, but these results confirm that false interactions do not require a generative model.
  • We also train a diffusion model and velocity (in addition to endpoint) flow matching, which behave broadly similarly (kindly see our reply to reviewer f4vV).
ModelFalse interaction β (γ = 0): LLHHRecovery α (γ = 1): LLHH
Endpoint flow.33 ± .07.07 ± .01.42 ± .021.00 ± .00
Velocity flow.11 ± .05.02 ± .00.40 ± .011.00 ± .00
Diffusion.18 ± .04.04 ± .00.26 ± .02.97 ± .00
Regressor.18 ± .02.07 ± .01.34 ± .041.00 ± .00

 

Single-cell perturbation: We train a direct-mean MLP, a diffusion model, and a cVAE (identical encoder, pairs, folds):

ModelEIE_IRecovery αHallucination βΔβ vs flow [95% CI]
Flow (paper).92 [.70, 1.38].65 [.27, .93].78 [.54, 1.17]
Direct-mean MLP.82 [.70, 1.21].56 [.36, .69].69 [.47, .97]−.09 [−.31, .04]
Diffusion.89 [.69, 1.22].62 [.27, .79].71 [.53, 1.08]−.07 [−.22, .06]
cVAE.87 [.70, 1.25].58 [.34, .70].76 [.51, 1.13]−.02 [−.19, .08]

We find all four models hallucinate at comparable levels, with the direct-mean MLP both recovering and hallucinating mildly less.

 

GEOM (molecular conformers, new): here we score the same 150 query molecules with decreasing support in the training set.

Training support for the query moleculesET-FlowRegressor
Neighbouring molecules.70 ± .01.82 ± .05
Parent molecules only.75 ± .02.83 ± .04
Distant molecules only.83 ± .011.48 ± .10

Both models get worse as support is removed, with the flow model degrading considerably less.

 

We therefore propose to present these failures in terms of training data incompletely constraining the interactions and which is exhibited across flows, diffusion, VAEs, and regressors. Nevertheless, we still believe studying these errors in generative models is important due to the following reasons:

  • Flow (and diffusion) models are increasingly deployed across the fields we investigate, so their failures are practically relevant.
  • Generative models can fail in the manner directly related to a central point of the paper: contrary to a regressor, they produce sharp, realistic samples whose interactions are wrong. Our new two-mode experiment shows this directly (please see the Model investigation section in our reply to reviewer f4vV).

The same held-out pairs with and without supporting pairs (W2, Q2)

We appreciate this suggestion and ran this on single-cell perturbation (Norman) as a crossover design. Specifically, in each fold, we split the held-out pairs into two halves by gene. We then compare two runs, with each run removing the double perturbations which share a gene with the one of the two halves. Each held-out pair is then scored once with and without its neighbouring pairs (3 seeds, 4 folds, 2 runs):

PairsnChange in EIE_I, anchored [95% CI]Change in EIE_I, fully generated [95% CI]
Lost any neighbouring training pair112+.08 [−.01, .15]+.04 [−.06, .12]
Lost more than half of them75+.14 [.07, .21]+.10 [.03, .16]
Lost none (control)19+.04 [−.13, .22]–

We find that removing most of a pair's neighbours leads to degraded interaction prediction, while only removing few neighbours does not yield a clearly detectable effect.

 

Further matched design Here, the addition of GEOM is also highly relevant, as the same 150 query molecules are scored under three levels of support, yielding EI:.70→.75→.83E_I: .70 \rightarrow .75 \rightarrow .83 from most to least support. Moreover, as in our original manuscript, our PPG simulator also keeps the held-out corner fixed and holds out 25, 50, or 75% of the total grid, yielding EI:.80→.96→.99E_I: .80 \rightarrow .96 \rightarrow .99.

Finally, we agree that seeing each gene in other pairs does not determine the interaction and propose to change the wording we use to "relationally constrained" instead of "relationally identified". We thank the reviewer for this correction.

   

How much of the hallucination is variability, and does it depend on the number of cells or the uncertainty calculation? (Q3)

We thank the reviewer for this suggestion, we agree this is an important control analysis.

Here we show how the metrics behave across the number of cells:

Generated cells per condition2496192 (paper)768
Anchored β\beta1.07.83.80.78
Fully generated β\beta.63.61.60.60
Fully generated α\alpha.39.38.39.39

Additionally, we check where the hallucination comes from:

Source (share of β2\beta^2, 31 repeatable pairs)AnchoredFully generated
Noise in the measured targets (200 half-splits of each condition's cells)8%10%
Finite number of generated cells2%0.4%
Variation between independently trained seeds28%43%
Systematic: shared by all seeds62%46%

To conclude:

  • beyond about 100 generated cells, hallucination only minimally changes and recovery not at all. The 192 cells used in the paper thus appear to lie in a stable sample-size regime.
  • Noise accounts for up to 10% of the hallucination.
  • About half of the hallucination is systematic, such that independently trained models make the same error. We interpret this part to indicate what the training data leave undetermined.

I thank the authors for the detailed reply and the new experiments which directly address my questions about the baselines and the variability. My main concerns however are still there.

  • First of all, the new results actually weaken the paper's central claim. The regressors, the diffusion models and the cVAE all show the same false interactions and 28-43% of the single cell hallucination is due to seed variation. It's therefore a result of the training data not constraining the interaction rather than anything to do with the flow models or their inductive bias. That in itself necessitates a very significant rewrite of the title, abstract and contributions which I can't assess from the rebuttal. The new comparison tables also don't have seeds or confidence intervals so it's not clear what is actually specific to the flow models.
  • Secondly the crossover experiment provides very weak evidence. The overall effect has an interval that includes zero, only the 'lost more than half' subgroup is clearly positive and the control group (n=19) is too small to be of any use. The paper's headline figure is the drop in recovery from 0.42 to 0.17, though recovery and hallucination are not reported separately in this instance.
  • Thirdly, some of the numbers for the paper’s own model appear to be at odds with the manuscript (low-low PPG false interaction 0.19 v .33 and recovery 0.31 v .42; single cell recovery 0.42 v .65). Can you clarify what has changed? It's also not at all clear how a two factor interaction is defined for GEOM.

We thank the reviewer for their continued engagement. We'd like to address the three points in order.

 

Q1: We appreciate this feedback. The extra analyses indeed confirm the failures are not specific to flow matching. While it was a weakness of our initial submission that it did not investigate this, it also did not claim any causality. As a consequence, the listed contributions (interaction diagnostics, hallucination, continuation determines extrapolation, aggregate metrics can be inappropriate for interactions) need minor edits to refer to extrapolating models generally. We'd naturally be receptive to feedback if the reviewer recommends otherwise. We also propose a generalization of the title: "Predicting Unseen Scientific Interactions: Hallucinated Interactions Beyond the Training Support".

On seeds and intervals, we apologize for reporting only means initially, and are adding these to the tables in our previous reply to the reviewer.

 

Q2: On consideration, we acknowledge the weakness of the subgroup split alone. We provide two additions:

  • Without needing to rely on splitting, the share of a pair's neighbouring training pairs that is removed correlates with the increased interaction error (Spearman .36, 95% CI [.06, .60], permutation p < .001).
  • Recovery and hallucination separately (change when neighbours are removed, 95% CI's resampling genes):
PairsnΔαΔβ
Lost more than half of their neighbours75−.09 [−.18, .00]+.10 [.04, .16]
Lost any112−.04 [−.13, .05]+.06 [−.01, .11]
Lost none (control)19−.01 [−.14, .10]+.04 [−.13, .20]

Removing most of a pair's neighbours mainly increases hallucination.

We also check the interval for the the reported .42 to .17 drop, but find its interval to include zero (Δα −.25 [−.53, .18]), so we'll make sure to include this and mark it as descriptive in the revision.

As such, we rely on the matched tests as the PPG setting (holding out 25%, 50% and 75% of the grid), the crossover presented here above, and GEOM (same 150 query molecules scored under 3 levels of support; see below).

 

Q3: Thank you for raising this. These differ in terms of training budget or estimate:

  • PPG: The paper reports the model trained to 1M steps, while the comparison tables use runs trained for 100k steps (with a completed lr schedule) so that all models receive the same amount of steps. We'll explicitly note training budgets in the revision.
  • Single cell: We apologize for this mixup. The paper's 0.42 belongs to the fully generated estimate on supported pairs, while the 0.65 from the reply is the anchored recovery all pairs. For completeness: fully generated is 0.42 (supported) and 0.39 (all), while anchored gives 0.66 and 0.65. This probably results from the paper stating the perturbation results use the anchored estimate, but this revealed this is incorrect for the supported vs cold-start comparison. We'll make sure to clarify this in the revision.

GEOM was added after the initial submission and apologize for being too brief in our initial reply. Each query is a "square" of a molecule sharing a scaffold with two attachment sites. The two factors are the substituents at the two sites: the parent molecule (00), the two single edits (01 and 10), and the double edit (11). Then, the response is the mean heavy-atom radius of gyration over the conformer ensemble, and the interaction is is μ11−μ10−μ01+μ00\mu_{11} - \mu_{10} - \mu_{01} + \mu_{00}. We are now uploading an updated version of the manuscript which has the full description of the GEOM data, methods, and its results.

Reviewer_f4vVOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how conditional flow matching models learn and generalize non-additive interactions under unseen experimental conditions. The authors introduce an evaluation framework that decomposes predicted interactions into recovery and hallucination components, and conduct experiments on three scientific domains: physiological time-series simulation, drug dose-response prediction, and single-cell perturbations. The main findings suggest that interaction learning can behave differently from overall prediction accuracy, and that interaction recovery depends on the structure of the observed conditions.

Critical strengths
  1. The paper addresses a practically relevant problem, as accurately predicting interactions between experimental factors is important in many scientific applications.

  2. The evaluation across three different scientific domains provides a relatively broad empirical perspective.

  3. The paper is generally well-organized, and the distinction between overall prediction accuracy and interaction recovery is clearly motivated.

Critical weaknesses
  1. Limited methodological novelty. The main methodological contribution is the decomposition of predicted interactions into recovery and hallucination components. However, this decomposition is essentially a standard orthogonal projection, and the resulting error identity follows directly from the Pythagorean theorem. While the proposed metrics may be useful for evaluation, they do not introduce a fundamentally new methodology or theoretical insight.

  2. The empirical findings are largely expected and offer limited new insights. Several central observations, such as interaction effects being harder to learn than dominant main effects, extrapolation becoming more difficult when training data provides weaker constraints, and nonlinear transformations changing prediction errors, are relatively intuitive. The experiments mainly confirm these observations without sufficiently explaining the underlying mechanisms or revealing suprising behaviors specific to flow matching models.

  3. Insufficient investigation of flow matching-specific behaviors. Although the paper focuses on conditional flow matching models, most of its findings appear applicable to general predictive models. For example, accurate endpoint predictions do not necessarily imply accurate interaction recovery, regardless of the underlying model architecture. The paper lacks systematic comparisons with alternative generative or discriminative models to establish whether the observed interaction hallucinations are particulary associated with flow matching.

  4. Limited practical implications. The paper identifies several failure modes but provides limited guidance on how to address them. It remains unclear how the proposed analysis could inform better training objectives, model architectures, or generalization strategies. As a result, the practical contribution is primarly diagnostic.

Questions for authors
  1. What are the specific properties of flow matching models that contribute to interaction hallucinations? Would similar behaviors be observed with standard regression models or other generative approaches?

  2. Since the recovery–hallucination decomposition is based on standard vector projections, what additional insights does it provide beyond existing error metrics, cosine similarity, and interaction estimation methods?

  3. Could the authors provide more mechanistic explanations for the observed differences in interaction generalization, rather than primarily attributing them to the available training support?

  4. Do the authors envision any concrete strategies for improving interaction recovery based on their findings, such as interaction-aware training objectives or regularization?

4 comments

We sincerely thank the reviewer for their evaluation. We have run new experiments for each of the raised points, which we address one by one below. First, we highlight the main additions:

  • New real-data setting: for molecular conformers (GEOM), the same molecules are scored under decreasing training support with an equivariant flow model, leading to increased interaction errors (EIE_I: .70 → .83).
  • Additional modeling recipes: endpoint and velocity flows, diffusion, a cVAE, and regressors
  • Mechanism: we provide an investigation showing that as generative models extrapolate away from the training support, mode proportions progressively break down while every sample stays sharp and realistic.
  • Fixing failures: we find that changing models, objectives, or samplers does not restore recovery, whereas new data in the held-out region does (α\alpha: .31 → .81), and find that the presented interaction framing can guide and improve data collection

   

Is this specific to flow matching? (W3, Q1)

While we use, and propose to keep, flow models as the main subject since they are what is deployed in our studied fields, we now show and no longer suggest that it is the generative aspects that cause the broad observed failures.

  • For PPG, we fit a matched regressor (i.e. same network, data, training budget) and find it fails in a similar manner. It hallucinates in the additive regime (γ=0\gamma=0, with βreg=.18,βflow=.33\beta_\text{reg}=.18,\beta_\text{flow}=.33 and recovers interactions only in one of the two held-out region of condition space (failed corner at αreg=.34,αflow=.42\alpha_\text{reg}=.34, \alpha_\text{flow}=.42.
  • For single cells, we fit a direct-mean MLP, a diffusion model, and a cVAE (same encoder) and find that all models hallucinate at a similar level (median β=.69−.78\beta=.69-.78).

Why then study generative models?

  • Flow (and diffusion) models are increasingly deployed across the fields we investigate, so their failures are practically relevant.
  • Generative models can fail in the manner directly related to a central point of the paper: contrary to a regressor, they produce sharp, realistic samples whose interactions are wrong. Our new two-mode experiment (please see "Mechanism beyond training support") shows this directly.

   

Mechanism beyond "training support" (W2, Q3)

We agree that some of our findings are expected: interactions are harder to predict than main effects, and less support makes extrapolation worse. However, we present three findings which are not explained by support alone:

  • models predict interactions where none exist (γ=0\gamma=0; see above);
  • two equally held-out corners behave considerably differently (the PPG corner comparison);
  • new data in the held-out region matter far more than any modelling choice (see Practical implications).

Model investigation. We appreciate the suggestion for further mechanistic explanations.

We now include a two-mode version of the PPG simulator in which one physiological parameter (BV3) is hidden from the model and takes one of two values at random. True outputs therefore form two clusters with equal probability. We investigate model error as a function of distance to training data:

Distance from training data (grid steps)Endpoint flowVelocity flowDiffusionRegressor
0–799% / 8%97% / 9%99% / 9%0% / –
8–1068% / 31%77% / 30%71% / 40%0% / –
11–1423% / 85%46% / 61%22% / 83%0% / –
15+0% / –5% / –0% / –0% / –

Table shows "share of samples on a true mode / share of conditions with wrong proportions" (with sampling noise yielding ~10% proportion error).

  • We find that all tested models show progressive mode collapse as targets lie further from the training support. Furthest away, nearly no samples land near the correct modes, instead clustering in the wrong place. Importantly, each sample is still a sharp, realistic waveform, without individual samples signaling the error. We note this failure mode is specific to generative models, as regressors predict averages which lie between the modes.
  • As the probability of each mode is 0.5 everywhere, the mode probabilities have no interaction. However, we find that all generative models invent interaction. For the held-out joint conditions, the interaction of mode probabilities has an RMS of 0.12-0.13 (sd across seeds ≤\leq .03), compared to the exact sampler with RMS$\approx$0.01.
  • Changes to the sampler (Langevin corrections, noise schedules, step counts) changes errors by a maximum of .08, which indicates that errors arise predominantly from what models have learned, rather than in how they are sampled.

What the α/β\alpha/\beta decomposition adds (W1, Q2)

We acknowledge it is an orthogonal projection and avoid claiming mathematical novelty. However, we do believe our conceptual arguments and diagnostics are useful to the field as generative models are increasingly used in scientific domains, which we hope to have strengthened with these revisions. The use of the decomposition is that both standard metrics are functions of the same two numbers: EI2=(1−α)2+β2E_I^2 = (1-\alpha)^2 + \beta^2 and cos⁡=α/α2+β2\cos = \alpha / \sqrt{\alpha^2 + \beta^2}. As EIE_I and cosine mix α\alpha and β\beta, this changes conclusions for our results, for example:

ComparisonEIE_Icosineα\alpha (recovery)β\beta (hallucination)
PPG: endpoint → velocity flow.759 → .669.65 → .80.42 → .40.48 → .30
Single cell: cVAE vs diffusion.935 vs .951.47 vs .54.30 vs .51.64 vs .73

For the PPG, both E1E_1 and cosine indicate that velocity flow is better. However, we find recovery to be very similar, but with considerably less hallucinated non-additivity. For single cell, E1E_1 favours the cVAE while cosine favours diffusion. Here α\alpha and β\beta help show why they disagree, with diffusion recovering considerably more of the interaction, yet also hallucinating more.

Finally, we note that when there is no true interaction, neither EIE_I nor cosine is defined. β\beta still measures the false interaction as used in our manuscript.

   

Practical implications (W4, Q4)

We appreciate the suggestion to provide more concrete strategies, which we have now investigated. Overall, we find that data matters far more than modelling choices. Using the PPG LL corner prediction (3 seeds):

ChangeEIE_IRecovery α\alphaHallucination β\beta
Endpoint flow (manuscript).76.42.48
Velocity flow.67.40.30
Diffusion.78.42.51
Regressor.83.34.50
Additional training and/or data:
Endpoint flow, trained further without new data.82.31.43
Endpoint flow + training on 16 new conditions, chosen at random.50.80.39
Endpoint flow + training on 16 new conditions, targeting response.39.75.30
Endpoint flow + training on 16 new conditions, targeting interaction.32.81.26

We show that model choice (flow, diffusion, regression) has at most a modest impact on performance, as all recipes share poor recovery with significant hallucination.

This is in stark contrast to sampling additional conditions, for which we compare random and targeted sampling. The latter applies the paper's framing to data collection: at each step, we pick the condition that most reduces uncertainty about the interaction component, rather than the response (which yields EI=.39E_I=.39 instead of EI=.32E_I=.32). Here we use a Gaussian process over the two condition coordinates (only where training data exists). The α/β\alpha / \beta decomposition shows that the advantage over random placement is due to lower hallucination (β=.39→.26\beta=.39 \rightarrow .26).

In the appendix, we will also include preliminary experiments using interaction-aware contrast and readout losses, as examples of the requested interaction-aware training objectives. However, we find these to not provide consistent gains, which is congruent with the interpretation that the degree to which training support determines the continued interaction structure is most important (i.e., more so than standard model and training choices).

Thank you for the detailed response. Much appreciated the new comparisons with additional model families.

But I still have some concern about the novelty and depth of the scientific insights. The recovery/hallucination decomposition is certainly a useful set of diagnostics, though it's all based on fairly standard geometric quantities. The two-mode experiment in terms of characterising distributional failure under extrapolation is interesting, but again I don't think we have a real understanding of the underlying mechanisms.

Overall I think the additional experiments have actually strengthened the paper and have gone some way to addressing a number of my concerns. I'll be raising my rating to borderline, though I still have some reservations about the overall contribution.

Thank you kindly for the continued engagement and for raising your score.

Regarding mechanisms, we have indeed been trying to understand this better. While we unfortunately have not been able to properly write this up in time, we will include an additional analysis on this in the final version, of which we summarize the main findings here:

To understand where in sampling the multimodal failures arise, we compare the network's predicted clean sample with the exact posterior at every step. We present two observations when extrapolating outside the training support:

  1. As expected, before a sample commits to a mode, generative models predict the average across modes. But this average carries a highly similar extrapolation error as a regressor (rank correlation of 0.81-0.84 across held-out conditions). Where this biased average sits relative to the modes, determines which mode gets oversampled (i.e. the closer one).
  2. Models commit to a mode much later than the exact posterior indicates, which progressively worsens away from the training support, with models' mode 'decisions' occuring at signal-to-noise ratios higher than it should: ~4x near support, ~30x at intermediate distance, and ~400x far away. As a result, ~20% of samples end up at a different mode than assigned by the exact posterior. We find this second effect to partly cancel the first on average, but individual conditions still end up with with incorrectly weighted modes.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for the AC and reviewers

We thank all reviewers for their constructive engagment and helping us in substantially improving the manuscript. We'd like to summarize what the final version will contain:

 

Reframed contribution: New baselines confirm that interaction failures are not flow-matching specific, as diffusion, cVAE, and regressor models can fail similarly when training data do not fully determine the interaction. We propose minor revisions to the title ("Predicting Unseen Scientific Interactions: Hallucinated Interactions Beyond the Training Support"), abstract, and contributions to reflect this generalized conclusion. Generative models and especially flow models remain the main subject: they are what's being deployed in these fields and fail in distinct ways by producing sharp, realistic samples with invented interactions across mode proportions, which we now analyze in a more detailed manner.

Commited for the final version:

  1. Standardized recipes across modalities with equal tuning effort.
  2. Matched tests holding targets fixed while only varying training support: GEOM conformers, Norman cross-over (incl. dose-response correlation), and PPG held-out region sizes.
  3. Mechanism: The two-mode experiment and sampling-trajectory analysis yields two findings: the predicted mean of generative models before mode-commitment inherits the regressor's extrapolation bias, and models commit to modes increasingly late further from the training support.
  4. Practical guidance: We show that new data matter for more than modeling choices, and that interaction-targeted data collection reduces hallucination more than a variety of alternatives. We include negative results for interaction-aware losses and alternate embedding schemes in the appendix.

 

Kind regards,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.