Inverting Foundation Models of Brain Function with Simulation-Based Inference

Niels Bracher; Stefan T. Radev; Xavier Intes

arxiv: 2604.23865 · v2 · submitted 2026-04-26 · 💻 cs.LG · cs.AI· stat.ML

Inverting Foundation Models of Brain Function with Simulation-Based Inference

Niels Bracher , Xavier Intes , Stefan T. Radev This is my paper

Pith reviewed 2026-05-08 06:19 UTC · model grok-4.3

classification 💻 cs.LG cs.AIstat.ML

keywords foundation modelsbrain activitysimulation-based inferenceneural decodinginverse modelingstimulus generationlarge language models

0 comments

The pith

Synthetic brain maps from a foundation model can be inverted to recover the emotional properties of the stimuli that produced them.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether a foundation model of brain function can be run in reverse. It generates news headlines whose emotional tone is controlled by parameters for valence, arousal, and dominance, then feeds those headlines into the brain emulator to produce predicted activity patterns. Simulation-based inference is used to learn a mapping that turns those patterns back into the original parameters. A sympathetic reader would care because successful inversion shows that the emulator has captured enough structure in its neural encodings to support decoding and stimulus design. The work is framed as a proof-of-concept step toward using such models for in-silico neuroscience experiments.

Core claim

The authors show that latent linguistic parameters controlling headline generation can be recovered from the brain activity maps produced by the emulator. They demonstrate this recovery with simulation-based inference, which learns a probabilistic mapping from the predicted maps to the parameter values. The results are taken as validation that the model's neural encodings preserve information about stimulus properties. The same setup also shows that large language models can act as controllable generators of stimuli for simulated brain experiments.

What carries the argument

Simulation-based inference used to learn a probabilistic mapping from predicted brain maps back to the latent parameters that generated the stimuli.

If this is right

The recoverability of parameters validates the information content of the brain emulator's encodings.
Language models can serve as controllable stimulus generators for running simulated neuroscience experiments.
The approach supplies a concrete method for decoding and inverse design with foundation models of brain function.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the inversion holds on real data, the method could be extended to decode emotional tone or other stimulus features directly from measured brain activity.
Optimizing the parameters to produce target brain maps could allow systematic design of stimuli that drive specific neural responses.
The same simulation-based inversion pipeline might apply to other foundation models of brain function across different tasks or modalities.

Load-bearing premise

The brain emulator accurately mimics real human responses to the language-model stimuli and the learned mapping generalizes outside the simulated training data.

What would settle it

Recovery of the true generating parameters fails when the same inversion procedure is applied to held-out simulated brain maps or to real human brain recordings elicited by the same headlines.

Figures

Figures reproduced from arXiv: 2604.23865 by Niels Bracher, Stefan T. Radev, Xavier Intes.

**Figure 1.** Figure 1: Overview of our simulation-based inference (SBI) pipeline for decoding from foundation brain models view at source ↗

**Figure 2.** Figure 2: Sample synthetic activation sequence for one headline. view at source ↗

**Figure 3.** Figure 3: Cycle consistency (left) and Bayesian inversion results (right) view at source ↗

**Figure 4.** Figure 4: Real-world point estimate results. Inferred outlet-level latent score profiles for real-world headlines on shared AI-related topics reported by Associated Press (AP), Financial Times (FT), The Guardian, and Reuters. Small differences appear in valence, subjectivity, and formality, whereas arousal, dominance, and certainty appear to be encoded similarly. of four outlets: Associated Press (AP), Financial Ti… view at source ↗

**Figure 5.** Figure 5: Extended in-distribution SBI diagnostics. view at source ↗

**Figure 6.** Figure 6: Detailed qualitative results for real-world AI headlines. view at source ↗

read the original abstract

Foundation models of brain activity promise a new frontier for in silico neuroscience by emulating neural responses to complex stimuli across tasks and modalities. A natural next step is to ask whether these models can also be used in reverse. Can we recover a stimulus or its properties from synthetic brain activity? We study this question in a proof-of-concept setting using TRIBEv2. We pair the brain emulator with large language models (LLMs) that generate news headlines from linguistic parameters such as valence, arousal, and dominance. We then use simulation-based inference to learn a probabilistic mapping from brain maps to latent stimulus parameters. Our results show that these parameters can be recovered from predicted brain maps, validating the quality of neural encodings. They also show that LLMs can serve as controllable stimulus generators for simulated experiments. Together, these findings provide a step toward decoding and inverse design with foundation brain models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shows SBI can invert TRIBEv2 to recover stimulus parameters from its own simulated brain maps, but this only confirms internal invertibility and leaves the claim of validating neural encodings untested against real data.

read the letter

The core result is that simulation-based inference recovers valence, arousal, and dominance parameters from brain maps generated by TRIBEv2 when the stimuli are news headlines produced by LLMs. It works inside their closed loop and demonstrates that the chosen mapping is invertible under the simulator's rules. They also show LLMs can act as controllable generators for those stimuli, which keeps the experimental setup scalable without hand-crafted examples. That combination is the actual new piece: a practical way to run inverse design experiments entirely in silico with a foundation brain model. The execution looks clean for a proof-of-concept; the SBI training uses data produced by the same forward model, so the recovery numbers are at least internally consistent. The LLM control adds a useful degree of freedom for varying linguistic properties systematically. The main limitation is that all training and test data come from TRIBEv2 itself. There is no comparison to real fMRI recordings for the same headlines, so the experiment cannot tell us whether the recovered parameters would hold up if the brain maps were measured rather than predicted. The validation statement therefore depends on an assumption about TRIBEv2's fidelity that this paper does not check. Minor gaps include the lack of reported error bars or ablation checks on the inference network in the abstract, though the full text may fill those in. This is aimed at computational neuroscientists already working with brain foundation models or SBI methods who want to explore decoding or inverse problems in simulation. A reader looking for a concrete example of how to close the loop between stimulus generation and parameter recovery will find usable ideas here. It is worth sending to peer review because the method is clearly described and the internal result is reproducible on its own terms; referees will likely focus on the missing external anchor rather than reject the framing outright.

Referee Report

2 major / 1 minor

Summary. The paper presents a proof-of-concept for inverting a foundation model of brain function (TRIBEv2) using simulation-based inference (SBI). LLMs generate news headlines from parameters such as valence, arousal, and dominance; TRIBEv2 produces corresponding synthetic brain maps; and an SBI network is trained to recover the original parameters from those maps. The central claim is that successful recovery demonstrates the quality of the model's neural encodings and shows LLMs can serve as controllable stimulus generators for simulated experiments.

Significance. If the invertibility result holds under the simulator, the work illustrates a technical route to decoding latent stimulus properties from synthetic brain activity and could support future in silico design of stimuli. Its broader significance for neuroscience is constrained by the absence of any anchor to empirical fMRI data, so the demonstration remains internal to the chosen forward model.

major comments (2)

[Abstract] Abstract: the statement that parameter recovery 'validating the quality of neural encodings' is not supported by the described experimental design. All training and test data are generated by the identical TRIBEv2 forward simulator with no external real-brain fMRI recordings for the same stimuli, so the experiment shows only that the chosen mapping is invertible inside the simulator rather than that the encodings match human responses.
[Abstract] Abstract and Results: the claim of successful recovery is presented without any quantitative metrics (e.g., posterior error, R², coverage probabilities), error bars, cross-validation splits, or ablation controls on the SBI network or TRIBEv2 parameters. This absence leaves the robustness of the inversion unassessable and makes the headline result difficult to evaluate.

minor comments (1)

The manuscript would benefit from an explicit statement in the introduction or methods clarifying that the current validation is simulator-internal and outlining the additional steps required to test generalization to real fMRI data.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We are grateful to the referee for highlighting important clarifications needed in our presentation. We address the major comments point by point below and have made revisions to the abstract as indicated.

read point-by-point responses

Referee: [Abstract] Abstract: the statement that parameter recovery 'validating the quality of neural encodings' is not supported by the described experimental design. All training and test data are generated by the identical TRIBEv2 forward simulator with no external real-brain fMRI recordings for the same stimuli, so the experiment shows only that the chosen mapping is invertible inside the simulator rather than that the encodings match human responses.

Authors: We agree that the current wording overstates the implications of our results. The experiment is conducted entirely within the TRIBEv2 forward model, and thus shows invertibility of the simulated neural encodings rather than their fidelity to human brain activity. We have revised the abstract to read: 'Our results show that these parameters can be recovered from predicted brain maps, demonstrating the invertibility of the neural encodings in the TRIBEv2 model.' This change removes the unsupported claim while preserving the core contribution of the proof-of-concept. revision: yes
Referee: [Abstract] Abstract and Results: the claim of successful recovery is presented without any quantitative metrics (e.g., posterior error, R², coverage probabilities), error bars, cross-validation splits, or ablation controls on the SBI network or TRIBEv2 parameters. This absence leaves the robustness of the inversion unassessable and makes the headline result difficult to evaluate.

Authors: We acknowledge the need for more explicit quantitative support in the abstract. The results section includes visualizations of parameter recovery and SBI posterior distributions, but we have now added specific metrics to the abstract, such as the mean R² across parameters and posterior coverage rates. We have also included a brief mention of the cross-validation approach used for the SBI network. These additions make the robustness of the inversion more readily evaluable without altering the underlying experiments. revision: yes

Circularity Check

0 steps flagged

No circularity: parameter recovery is a standard invertibility test inside the simulator

full rationale

The paper generates stimuli from LLMs using independently chosen valence/arousal/dominance parameters, feeds them through the fixed TRIBEv2 forward emulator to produce brain maps, and then trains an SBI network to map those maps back to the original parameters. Because the target parameters are external inputs to the simulator (not outputs or fits derived from the maps), successful recovery on held-out simulations demonstrates only that the chosen forward mapping is learnably invertible; it does not reduce any claimed result to a tautology, self-definition, or fitted-input prediction. No self-citation load-bearing steps, uniqueness theorems, or ansatzes appear in the derivation chain. The procedure is therefore a self-contained internal consistency check rather than a circular validation.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated.

pith-pipeline@v0.9.0 · 5451 in / 929 out tokens · 34804 ms · 2026-05-08T06:19:14.697794+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references · 13 canonical work pages

[1]

OpenAI product announcement

URL https://openai.com/index/ introducing-gpt-5-4-mini-and-nano/ . OpenAI product announcement. Pavlick, E. and Tetreault, J. An empirical analysis of for- mality in online communication.Transactions of the association for computational linguistics, 4:61–74, 2016. Ponce, C. R., Xiao, W., Schade, P. F., Hartmann, T. S., Kreiman, G., and Livingstone, M. S. ...

work page 2016
[2]

- High (1.0): Use positive-charge nouns/adjectives (Benefit, Success, Growth)

Valence (Sentiment): - Low (0.0): Use negative-charge nouns/adjectives (Crisis, Failure, Threat). - High (1.0): Use positive-charge nouns/adjectives (Benefit, Success, Growth)

work page
[3]

- High (1.0): Use high-velocity, high-impact verbs (Blasts, Surges, Explodes, Slams, Sparks)

Arousal (Energy): - Low (0.0): Use static, low-impact verbs (Remains, Exists, Continues) and avoid verbs of change, escalation, or sudden events. - High (1.0): Use high-velocity, high-impact verbs (Blasts, Surges, Explodes, Slams, Sparks)

work page
[4]

System is hit by

Dominance (Agency): - Low (0.0): Use passive voice / patient-as-subject (e.g., "System is hit by...") and avoid naming or implying a controlling agent. - High (1.0): Use active voice / agent-as-subject (e.g., "Agency Enforces...")

work page
[5]

Shocking,

Subjectivity (Interpretation): - Low (0.0): Use concrete nouns and numbers; no evaluative adjectives. - High (1.0): Use heavy modifiers and interpretive framing (e.g., "Shocking," " Stunning")

work page
[6]

- High (1.0): Use absolute indicatives (e.g., Is, Will, Must, Confirmed)

Certainty (Modality): - Low (0.0): Use epistemic hedges (e.g., Might, Could, Possible, Rumored). - High (1.0): Use absolute indicatives (e.g., Is, Will, Must, Confirmed)

work page
[7]

Administrative collapse is observed

Formality (Register): - Low (0.0): Use monosyllabic, common Germanic words (e.g., Fix, Whack, Way) and clearly everyday, non-institutional phrasing. - High (1.0): Use multisyllabic, technical Latinate words (e.g., Implement, Methodology, Mitigation). Orthogonality Instruction: These 6 knobs are independent. You can have High Formality with Low Dominance (...

work page
[8]

- High (1.0): Uses positive-charge nouns/adjectives (Benefit, Success, Growth)

Valence (Sentiment): - Low (0.0): Uses negative-charge nouns/adjectives (Crisis, Failure, Threat). - High (1.0): Uses positive-charge nouns/adjectives (Benefit, Success, Growth)

work page
[9]

- High (1.0): Uses high-velocity, high-impact verbs (Blasts, Surges, Explodes, Slams, Sparks)

Arousal (Energy): - Low (0.0): Uses static, low-impact verbs (Remains, Exists, Continues). - High (1.0): Uses high-velocity, high-impact verbs (Blasts, Surges, Explodes, Slams, Sparks)

work page
[10]

System is hit by

Dominance (Agency): - Low (0.0): Uses passive voice / patient-as-subject (e.g., "System is hit by..."). - High (1.0): Uses active voice / agent-as-subject (e.g., "Agency Enforces...")

work page
[11]

Shocking,

Subjectivity (Interpretation): - Low (0.0): Uses concrete nouns and numbers; no evaluative adjectives. - High (1.0): Uses heavy modifiers and interpretive framing (e.g., "Shocking," " Stunning")

work page
[12]

- High (1.0): Uses absolute indicatives (e.g., Is, Will, Must, Confirmed)

Certainty (Modality): - Low (0.0): Uses epistemic hedges (e.g., Might, Could, Possible, Rumored). - High (1.0): Uses absolute indicatives (e.g., Is, Will, Must, Confirmed)

work page
[13]

type": "json_schema

Formality (Register): - Low (0.0): Uses monosyllabic, common Germanic words (e.g., Fix, Whack, Way) and clearly everyday, non-institutional phrasing. - High (1.0): Uses multisyllabic, technical Latinate words (e.g., Implement, Methodology, Mitigation). Scoring Strategy: - Infer one shared score per dimension for the headline set as a whole. - Base each sc...

work page 2023

[1] [1]

OpenAI product announcement

URL https://openai.com/index/ introducing-gpt-5-4-mini-and-nano/ . OpenAI product announcement. Pavlick, E. and Tetreault, J. An empirical analysis of for- mality in online communication.Transactions of the association for computational linguistics, 4:61–74, 2016. Ponce, C. R., Xiao, W., Schade, P. F., Hartmann, T. S., Kreiman, G., and Livingstone, M. S. ...

work page 2016

[2] [2]

- High (1.0): Use positive-charge nouns/adjectives (Benefit, Success, Growth)

Valence (Sentiment): - Low (0.0): Use negative-charge nouns/adjectives (Crisis, Failure, Threat). - High (1.0): Use positive-charge nouns/adjectives (Benefit, Success, Growth)

work page

[3] [3]

- High (1.0): Use high-velocity, high-impact verbs (Blasts, Surges, Explodes, Slams, Sparks)

Arousal (Energy): - Low (0.0): Use static, low-impact verbs (Remains, Exists, Continues) and avoid verbs of change, escalation, or sudden events. - High (1.0): Use high-velocity, high-impact verbs (Blasts, Surges, Explodes, Slams, Sparks)

work page

[4] [4]

System is hit by

Dominance (Agency): - Low (0.0): Use passive voice / patient-as-subject (e.g., "System is hit by...") and avoid naming or implying a controlling agent. - High (1.0): Use active voice / agent-as-subject (e.g., "Agency Enforces...")

work page

[5] [5]

Shocking,

Subjectivity (Interpretation): - Low (0.0): Use concrete nouns and numbers; no evaluative adjectives. - High (1.0): Use heavy modifiers and interpretive framing (e.g., "Shocking," " Stunning")

work page

[6] [6]

- High (1.0): Use absolute indicatives (e.g., Is, Will, Must, Confirmed)

Certainty (Modality): - Low (0.0): Use epistemic hedges (e.g., Might, Could, Possible, Rumored). - High (1.0): Use absolute indicatives (e.g., Is, Will, Must, Confirmed)

work page

[7] [7]

Administrative collapse is observed

Formality (Register): - Low (0.0): Use monosyllabic, common Germanic words (e.g., Fix, Whack, Way) and clearly everyday, non-institutional phrasing. - High (1.0): Use multisyllabic, technical Latinate words (e.g., Implement, Methodology, Mitigation). Orthogonality Instruction: These 6 knobs are independent. You can have High Formality with Low Dominance (...

work page

[8] [8]

- High (1.0): Uses positive-charge nouns/adjectives (Benefit, Success, Growth)

Valence (Sentiment): - Low (0.0): Uses negative-charge nouns/adjectives (Crisis, Failure, Threat). - High (1.0): Uses positive-charge nouns/adjectives (Benefit, Success, Growth)

work page

[9] [9]

- High (1.0): Uses high-velocity, high-impact verbs (Blasts, Surges, Explodes, Slams, Sparks)

Arousal (Energy): - Low (0.0): Uses static, low-impact verbs (Remains, Exists, Continues). - High (1.0): Uses high-velocity, high-impact verbs (Blasts, Surges, Explodes, Slams, Sparks)

work page

[10] [10]

System is hit by

Dominance (Agency): - Low (0.0): Uses passive voice / patient-as-subject (e.g., "System is hit by..."). - High (1.0): Uses active voice / agent-as-subject (e.g., "Agency Enforces...")

work page

[11] [11]

Shocking,

Subjectivity (Interpretation): - Low (0.0): Uses concrete nouns and numbers; no evaluative adjectives. - High (1.0): Uses heavy modifiers and interpretive framing (e.g., "Shocking," " Stunning")

work page

[12] [12]

- High (1.0): Uses absolute indicatives (e.g., Is, Will, Must, Confirmed)

Certainty (Modality): - Low (0.0): Uses epistemic hedges (e.g., Might, Could, Possible, Rumored). - High (1.0): Uses absolute indicatives (e.g., Is, Will, Must, Confirmed)

work page

[13] [13]

type": "json_schema

Formality (Register): - Low (0.0): Uses monosyllabic, common Germanic words (e.g., Fix, Whack, Way) and clearly everyday, non-institutional phrasing. - High (1.0): Uses multisyllabic, technical Latinate words (e.g., Implement, Methodology, Mitigation). Scoring Strategy: - Infer one shared score per dimension for the headline set as a whole. - Base each sc...

work page 2023