Source-linked AI summary
Behavioral Foundation Models for Quality Diversity
Nazim Bendib, Nicolas Perrin-Gilbert, Olivier Sigaud
TL;DR
BFMs have typically been used to extract one task-optimal behavior, leaving the potential for diverse, high-performing repertoires in their latent spaces underexplored. BFM-QD searches that space and outperforms parameter-space QD baselines across task types, with especially strong gains in sparse and deceptive settings.
Problem
Prior BFM work extracts a single optimal behavior, leaving open whether its latent space can produce a large repertoire of diverse, high-performing behaviors.
Source: §1 · p. 2Method
BFM-QD runs quality-diversity search in a pretrained BFM’s latent space and uses closed-form Backward Inference as a gradient-free policy-improvement operator.
Source: §1 · p. 2Results
BFM-QD consistently outperforms parameter-space baselines across dense locomotion, sparse navigation, and contact-rich manipulation, with stark gains in sparse and deceptive settings.
Source: §1 · p. 2, §7 · p. 10Takeaways & Limitations
BFM latent-space exploration supports quality-diversity optimization beyond single-task policy extraction, positioning BFMs as general-purpose backbones for diverse repertoires.
Source: §7 · p. 10BFM-QD requires a behaviorally diverse offline dataset and a sufficiently pretrained BFM, and its frozen model cannot express behaviors outside its pretrained latent space.
Source: §6 · p. 10
Abstract
from arXiv · showhide
Behavioral Foundation Models (BFMs) are an emerging paradigm in reinforcement learning, playing a role analogous to large language models in natural language processing: they have shown remarkable versatility, enabling zero-shot performance, fast imitation, and online adaptation, all by exploiting the structure of a latent space. In this work, we investigate whether the latent behavioral space induced by BFMs can serve as an effective search space to discover large repertoires of behaviorally diverse and high-performing policies through Quality-Diversity (QD) methods. While QD methods generally search directly in high-dimensional policy parameter space, in this paper, we present BFM-QD, a framework that performs QD search in the compact latent space of a BFM. We further show that the BFM-QD framework provides a closed-form, gradient-free policy improvement operator that approximates a policy gradient update, but requires no critic training and no backpropagation. Across continuous-control benchmarks spanning dense locomotion, sparse navigation, and contact-rich manipulation, BFM-QD consistently outperforms parameter-space baselines, with particularly stark gains in sparse and deceptive settings, where all tested parameter-space QD methods collapse to near-zero performance. These results show the effectiveness of the BFM-QD framework, benefiting from the synergy between dimensionality reduction of the search space and offline pretraining from diverse behavioral data. This positions BFMs as a general-purpose backbone for QD optimization, extending their utility beyond zero-shot task solving to the discovery of diverse behavioral repertoires.
1 Introduction
BFM-QD searches a pretrained BFM’s latent behavioral space to build repertoires, extending BFMs beyond extracting a single policy. Across task types, it outperforms parameter-space QD baselines, especially in sparse and deceptive settings.
1 Introduction
- Existing BFM work extracts one optimal behavior, leaving open whether its latent space can yield a large repertoire of diverse, high-performing behaviors.QD is motivated as a way to find repertoires useful for damage recovery, sim-to-real transfer, and skill libraries.Source: §1 · p. 2
- Across dense locomotion, sparse navigation, and contact-rich manipulation, BFM-QD outperforms parameter-space baselines, with especially stark gains where those methods fail.On easy, dense tasks the margin is modest, but convergence is faster.Source: §1 · p. 2
- BFM-QD relocates QD search from policy parameters to a pretrained BFM’s latent space, while leaving the QD objective unchanged.The framework is presented as the first to use a BFM latent space for QD optimization.Source: §1 · p. 2
- BFMs provide a closed-form, gradient-free policy-improvement operator matching a policy-gradient variation operator, without critic training or backpropagation.This is presented as a contribution of the framework.Source: §1 · p. 2
- Random latent vectors decode to diverse, efficient behaviors, and interpolation between solutions produces meaningful offspring.These properties characterize behavioral richness and structure in the BFM latent space.Source: §1 · p. 2
- QD performance saturates early in BFM pretraining, before zero-shot performance converges, so a fully trained BFM is not required.This finding concerns the pretraining budget needed for QD performance, not full zero-shot convergence.Source: §1 · p. 2
2 Related Work
Prior BFM work uses pretrained behavioral structure for task solving and adaptation, while this work turns that latent space toward repertoire search. QD methods vary in how they guide search, but BFM-QD instead searches an offline-pretrained policy-generating model’s latent space.
2.1 Behavioral Foundation Models
- BFMs support zero-shot task solving, fast imitation, online adaptation, and humanoid control; this work uses their latent space to discover behavior repertoires.Earlier approaches typically retrieve a single task-optimal behavior.Source: §2.1 · p. 3
2.2 Quality-Diversity algorithms
- QD methods maintain diverse, high-performing solutions; variants add covariance adaptation, critic-guided policy gradients, or descriptor conditioning to guide search.Latent-space illumination targets content generation, while Policy Manifold Search and DDE-Elites learn policy-parameter latents online from archive solutions.Source: §2.2 · p. 3
- Unlike archive-trained policy latents, BFM-QD searches the latent space of a policy-generating model pretrained offline without rewards.It complements methods that improve descriptors or critics by relocating the search space.Source: §2.2 · p. 3
3 Background
BFMs parameterize families of policies by latent codes and use learned successor features to infer policies for reward functions. QD maintains an archive of high-fitness solutions across behavioral-descriptor cells through repeated variation and evaluation.
3.1 Behavioral Foundation Models
- A BFM learns latent-indexed policies and successor features from offline transitions, representing a range of behaviors in a pretrained model.Forward-Backward and TD-JEPA are described as alternative ways to learn the underlying features and successor representations.Source: §3.1 · pp. 3–4
- At test time, a reward function is projected onto learned features to infer a latent code, and its corresponding policy is deployed zero-shot.The inferred code is projected to a hypersphere of radius √dz, where dz is the latent dimension.Source: §3.1 · p. 4
3.2 Quality-Diversity algorithms
- QD fills descriptor-space cells with high-fitness solutions, keeping at most one elite per cell and replacing it only with a fitter candidate.ME selects archive parents, varies them, evaluates fitness and descriptor, and inserts offspring into their cells.Source: §3.2 · p. 4
- Coverage measures populated cells, QD-score sums fitness across filled cells, and maximum fitness records the best single solution.Together, the metrics capture repertoire coverage, aggregate quality and diversity, and peak performance.Source: §3.2 · p. 4
4 Method
BFM-QD runs quality-diversity search over a frozen pretrained BFM’s compact, behaviorally rich latent space. Its Backward Inference operator steers archive solutions toward higher fitness using closed-form inference, without critic training or backpropagation.
4.1 Latent Space Quality-Diversity
- BFM-QD replaces parameter-space exploration with search over a pretrained BFM’s low-dimensional, behaviorally rich latent space.The BFM is trained once on offline data and kept frozen during QD runs.Source: §4.1 · p. 4, §4 · p. 4
- Archive latents are varied, decoded into policies for rollout evaluation, and reprojected onto the BFM’s latent sphere before archive competition.The archive stores codes rather than policy weights and seeds the next generation.Source: §4.1 · p. 5
4.2 Closed-Form Policy Improvement via Backward Inference
- Backward Inference estimates a reward-optimal latent code from evaluated transitions, then interpolates each selected archive parent toward it.The operator requires no gradient computation, critic training, or additional environment interactions beyond archive evaluation.Source: §4.2 · p. 5
- The interpolation is a Newton step on a local quadratic surrogate, with approximation error proportional to the reward projection residual.The BFM makes the step practical by providing the inferred optimum in closed form.Source: §4.2 · p. 5
5 Experiments
The experiments test whether BFM-QD works across task types and reward densities, what properties of the FB latent space support it, whether backward inference can replace policy gradients, and how much pretraining it needs. Across these analyses, BFM-QD performs strongly across varied tasks, with its advantage widening on harder tasks and its latent space supporting effective search and improvement.
5.1 Experimental Setup
- The evaluation spans 18 tasks in four environments, comparing FB and TD-JEPA QD variants with five parameter-space baselines using matched MLP architectures.Runs use a frozen, 50-dimensional pretrained FB, a 50 × 50 archive, 200,000 evaluations, and five random seeds.Source: §5 · p. 5, §5.1 · pp. 5–6
5.2 Main Results
- BFM-QD variants lead in QD-score across tasks and widen their advantage on AntMaze-medium and Cube-single, where parameter-space baselines collapse to near-zero QD-score and low coverage.The advantage is modest on dense locomotion, where BFM-QD variants converge faster; they operate in 50 latent dimensions versus about 20,000 policy parameters.Source: Figure 2 · p. 6, §5.2 · pp. 6–7
5.3.1 The latent space is behaviorally rich by construction
- Without optimization, random trained-FB codes yield higher coverage and QD-score than random MLP policies, while an untrained FB performs significantly worse.The gap grows with task difficulty, and the trained-versus-untrained comparison attributes behavioral richness to offline pretraining rather than architecture alone.Source: §5.3.1 · pp. 7–8
5.4 Can Backward Inference in BFMs replace policy gradient updates?
- FB-BI matches FB-PGA-ME across environments without critic training, while parameter-space PGA-ME fails on AntMaze-medium and Cube-single.The reported failure occurs where sparse, deceptive rewards make it difficult to train an accurate critic.Source: Figure 4 · p. 8, §5.4 · p. 9
5.3.2 The latent space structure makes interpolation meaningful
- Interpolation improves FB latent-space variants, producing behaviorally intermediate policies that support descriptor-space exploration, but yields no consistent improvement in parameter space.The comparison tests ME and CMA-ME with and without the isoline operator in both spaces.Source: §5.3.2 · p. 8
5.5 BFM-QD does not require a large pretraining budget
- QD-score plateaus early in FB training, before zero-shot performance converges, and a partially trained BFM already enables strong QD performance.The authors distinguish early-emerging behavioral diversity, which QD needs, from value precision, which continues improving later.Source: Figure 5 · p. 9, §5.5 · p. 9
5.6 Additional Analyses
- Additional analyses find stronger BFMs matter most on hard tasks, diverse QD populations improve maximum fitness, and exploration data outperforms expert data for pretraining.The reported proxy between zero-shot and QD performance is reliable across five BFMs; restricted behavioral coverage in offline data bounds latent-space diversity.Source: §5.6 · pp. 9–10
- Other analyses find combined-pretraining and learned-latent baselines fall short of FB-ME, BI preserves diversity better than sampling around z*, and RND-pretrained BFMs outperform QD-data pretraining.FB-BI is stable for α ∈[0.005, 0.1] and degrades at larger steps that pull offspring toward z* and collapse diversity.Source: §5.6 · p. 10
6 Limitations
BFM-QD depends on diverse offline data and a sufficiently pretrained, environment-specific BFM, whose frozen latent space limits the behaviors it can discover.
6 Limitations
- BFM-QD needs behaviorally diverse offline data and a sufficiently pretrained BFM before search, although full BFM convergence is unnecessary.A single BFM can be reused across subsequent tasks and QD runs in the same environment; narrower datasets and QD-collected pretraining data degrade performance.Source: §6 · p. 10
- A frozen BFM imposes a hard expressivity ceiling: behaviors outside its pretrained latent space remain inaccessible.Finetuning during QD is identified as future work.Source: §6 · p. 10
- Each BFM is trained for one environment, so environment-agnostic QD remains open.As in standard QD, the method assumes a given reward and behavioral descriptor, and produces only behaviors captured by learned features ϕ(s).Source: §6 · p. 10
- The paper does not measure how far discovered behaviors lie outside the pretraining data.Source: §6 · p. 10
7 Conclusion
The conclusion presents BFM-QD as a latent-space alternative to parameter-space QD and reports consistent benchmark gains, especially in sparse and deceptive settings.
7 Conclusion
- BFM-QD consistently outperforms parameter-space baselines across dense locomotion, sparse navigation, and contact-rich manipulation, with stark gains in sparse and deceptive settings.All tested parameter-space methods collapse in those settings.Source: §7 · p. 10
- BFM-QD replaces parameter-space exploration with exploration in the latent space of a pretrained BFM.Source: §7 · p. 10
- The authors position BFM latent-space exploration as a foundation for QD algorithms and BFMs as general-purpose backbones for diversity-seeking optimization beyond single-task policy extraction.Source: §7 · p. 10
A Proof of the Backward Inference Operator
The appendix relates Backward Inference to latent-space policy improvement and reviews the BFM representations and training approaches that provide its context.
A Proof of the Backward Inference Operator
- Backward Inference approximates a second-order Newton step for latent-space policy improvement, using a reward projection to infer z* without critic training or backpropagation.The approximation error depends on reward projection quality; it holds exactly when the reward lies in the feature span and approximately otherwise.Source: A Proof of the Backward Inference… · pp. 15–16
B Background on Behavioral Foundation Models
- BFMs are pretrained from reward-free interactions to produce policies for downstream rewards, balancing compact representations of long-term dynamics with expressivity and efficient zero-shot extraction.Source: B Background on Behavioral Foundation… · p. 17
B.1 Successor Features
- Successor Features linearly decompose rewards into shared features and task-specific weights, enabling task adaptation by refitting weights and zero-shot policy retrieval, but expressivity is bounded by the feature span.The feature map may be hand-coded or learned, making its quality a bottleneck.Source: §B.1 · p. 17
B.2 Forward-Backward Representations
- Forward-Backward representations factorize the successor measure into forward and backward maps, allowing reward-conditioned task inference and zero-shot policy retrieval without further environment interaction.Source: §B.2 · p. 18
B.3 TD-JEPA
- TD-JEPA trains an encoder, task encoder, policy-conditioned predictor, and policies to capture multi-step latent dynamics, then extracts reward-conditioned zero-shot policies.It replaces explicit successor-measure factorization; FB lacks an explicit state encoder, hindering scaling to raw pixels and feature reuse for transfer.Source: §B.3 · p. 19
B.4 BTD-FB
- BTD-FB replaces FB's uniform task prior, which can dilute learning signals in high-dimensional latent spaces, with a data-fitted distribution of behavioral tasks.It derives task vectors from offline-data subtrajectories and fits a Gaussian Mixture Model to them.Source: §B.4 · p. 19
C Background on Quality-Diversity Methods
Quality-Diversity methods maintain archives of high-performing solutions across behavioral cells, but differ in how they generate offspring. The section compares random mutation, adaptive and gradient-based operators, learned latent variation, and offline initialization.
C Background on Quality-Diversity Methods
- QD maintains an archive of at most one elite per behavioral cell, replacing it only when a candidate in that cell has higher fitness.Coverage measures populated cells, while QD-score sums their elite fitnesses.Source: C Background on Quality-Diversity… · p. 20
C.1 MAP-Elites
- MAP-Elites samples archive parents, applies isotropic Gaussian mutation, and inserts offspring when they improve the elite in their behavioral cell.Its isotropic mutation does not use fitness-landscape structure to guide search.Source: §C.1 · p. 20
C.2 CMA-ME
- CMA-ME replaces isotropic mutation with an adaptive covariance based on archive improvements, but still uses undirected sampling without task-specific fitness gradients.The covariance is estimated from directions that previously improved the archive.Source: §C.2 · p. 20
C.3 PGA-ME
- PGA-ME combines Gaussian mutation with gradient ascent on a learned critic, but its fitness-only gradient steps do not target underpopulated behavioral regions.The critic is trained on transitions collected during archive evaluation.Source: §C.3 · pp. 20–21
C.4 DCRL-ME
- DCRL-ME conditions its actor and critic on a target descriptor so offspring can be steered toward a chosen archive region, but both networks are trained from scratch on online evaluation data.The conditioned actor can also generate candidates directly for target cells.Source: §C.4 · p. 21
C.5 DDE-Elites
- DDE-Elites learns a compact VAE encoding from archive elites and combines isometric mutation, line mutation, and reconstructive crossover selected by a UCB1 bandit.Because the VAE trains only on existing archive solutions, its latent space need not reflect task-relevant structure.Source: §C.5 · p. 21
C.6 ME+Pretrain
- ME+Pretrain isolates offline initialization from behavioral-latent structure by warm-starting MLP policies with offline TD3 on the BFM dataset before standard MAP-Elites search.The pretraining objective uses task fitness as reward.Source: §C.6 · p. 21
D.1 Environments and Tasks
The evaluation spans four environments, from dense locomotion to long-horizon navigation and contact-rich manipulation. Baselines share the BFM-QD environment setup, archive configuration, training budget, and random seeds.
D.1 Environments and Tasks
- The four benchmarks progress from dense locomotion in HalfCheetah_uni and Walker2d_uni to long-horizon AntMaze-medium navigation and contact-rich Cube-single manipulation.The environments increase in difficulty from left to right in Figure 6.Source: Figure 6 · p. 22, HalfCheetah_uni · p. 22, Walker2D_uni · p. 22
- The evaluation defines environment-specific behavioral descriptors and fitness functions, and aggregates task results using per-task min-max normalization of QD-score and maximum-fitness curves.Scores are normalized across methods and seeds before averaging across tasks within each environment.Source: AntMaze-medium · pp. 22–23
D.2 Baselines Implementation Details
- Baselines share BFM-QD’s environment setup, archive configuration, training budget, and seeds, while differing in search operators and initialization.The implementations include covariance adaptation, critic-guided and descriptor-conditioned variation, VAE-based variation, and offline TD3 warm-starting.Source: §D.2 · pp. 23–26
D.3 BFM Implementation Details
BFM-QD keeps the pretrained policy frozen and performs archive search over latent codes. Its implementations combine model-specific architectures with Gaussian mutation and either backward-inference or interpolation variation.
D.3 BFM Implementation Details
- The FB model uses parallel observation-action and observation-latent streams feeding a shared trunk, while TD-JEPA uses state and task encoders with forward and backward predictors.TD-JEPA maintains target networks for its encoders and predictors using exponential moving averages.Source: §D.3 · p. 27
- The backward-inference operator uses sampled state-reward pairs to infer a latent target, then interpolates parents toward it; half of offspring use this operator and half use Gaussian mutation.Offspring are projected back onto the latent sphere.Source: §D.3 · pp. 28–29
- The interpolation operator combines two archive-parent latent codes with a uniformly sampled coefficient, alongside Gaussian mutation, and projects offspring onto the latent sphere.Each operator produces half of the generation’s offspring through interpolation or mutation.Source: §D.3 · pp. 28, 30
- For HalfCheetah_uni and Walker2D_uni, the authors collect the BFM pretraining dataset using Random Network Distillation.The same official implementation and hyperparameters are used.Source: §D.3 · p. 28
- BFM-QD stores latent codes rather than policy weights, applies variation in the latent space, and evaluates candidates with the frozen BFM policy.The archive is initialized and updated using evaluated fitness and behavioral descriptors.Source: §D.3 · p. 28
E.1 Expanded main results
BFM-QD variants lead across tasks, with larger advantages over parameter-space methods on harder tasks. The additional analyses relate archive coverage and maximum fitness to the BFM latent space, zero-shot performance, and diversity pressure.
E.1 Expanded main results
- BFM-QD variants consistently lead per-task QD-score results, with the gap over parameter-space baselines widening as reward density falls and task complexity rises.Figure 7 reports unnormalized curves across 18 tasks and four environments.Source: Figure 7 · p. 31, §E.1 · p. 31
E.2 Archive Visualization
- FB-CMA-ME achieves dense, broad descriptor-space coverage on AntMaze-medium and Cube-single, while parameter-space methods leave large regions unpopulated.CMA-ME occupies only a small cluster near the maze origin and has an almost empty Cube-single archive apart from moving the cube on the table surface.Source: Figure 8 · p. 32, §E.2 · p. 32
E.3 The latent space is behaviorally rich by construction
- Modest zero-shot returns can still yield high QD-scores, suggesting latent-space diversity matters more than zero-shot accuracy in that comparison.The figure also reports FB-CMA-ME as highest in average QD-score, closely followed by BTD-FB-CMA-ME and TD-JEPA-CMA-ME.Source: Figure 10 · p. 33
E.4 Effect of the Behavioral Foundation Model
- Zero-shot performance positively correlates with QD-score overall, but even weaker BFMs can achieve strong QD-scores on easy tasks.FB has the highest average QD-score, followed closely by BTD-FB and TD-JEPA.Source: §E.4 · p. 33
E.5 Diversity as an Optimization Strategy
- FB-CMA-ME outperforms FB-CMA-ES on maximum fitness, and the authors attribute this to diversity pressure helping escape local optima in latent space.FB-CMA-ME also consistently outperforms FB-CMA-ES and FB Zero-shot across tasks.Source: Figure 11 · p. 34, §E.5 · p. 34
E.6 Sensitivity to the Offline Dataset Quality
The supplementary analyses examine how offline dataset composition, alternative latent-space baselines, and sampling around the inferred optimum affect BFM-QD. They report that behavioral coverage and preserving each solution’s distinct behavior matter for QD performance.
E.6 Sensitivity to the Offline Dataset Quality
- On AntMaze-medium, QD performance degrades monotonically as expert data replaces random data, while even the spatially biased dataset substantially outperforms CMA-ME.The comparison varies six dataset compositions and averages QD-score across AntMaze-medium tasks; the results attribute the trend to expert trajectories covering a narrower set of optimal paths.Source: §E.6 · pp. 34–35, Figure 12 · p. 34
E.7 Additional Latent-Space Search Baselines
- FB-ME outperforms all baselines on Walker and Cube-single; on Cube-single it scores 0.58, versus 0.15 for DDE-Elites+Pretrain (BC) and 0.14 for DDE-Elites+Pretrain (TD3).On Walker, DDE-Elites+Pretrain scores 0.61 (BC) and 0.76 (TD3), versus 0.74 for DDE-Elites; no DDE-Elites or PoMS variant reaches ME. PoMS scores 0.63 on Walker and 0.07 on Cube-single.Source: §E.7 · p. 35, Table 13 · p. 35
E.8 Backward Inference vs. Sampling Around the Inferred Optimum
- FB-BI performs far better than sampling around a fixed inferred optimum, which confines search to nearby behaviors rather than exploring the broader behavioral space.The authors attribute the difference to BI moving each parent a small step toward the inferred optimum while retaining its distinct behavior.Source: §E.8 · p. 36, Table 14 · p. 36
E.9 Pretraining Data Collected by QD
This section documents analyses and supporting details on BFM pretraining data coverage, FB-BI step sizes, compute costs, final scores, limitations, and reproducibility. It also reports disclosures about crowdsourcing, human subjects, and the role of LLMs in the research.
E.9 Pretraining Data Collected by QD
- QD-collected pretraining data reduces performance because search refines elites instead of broadly covering transitions, whereas RND explores states with high prediction error.The takeaway is that broad dynamics coverage matters more than diversity along test descriptor axes; exploration data beats goal-directed data.Source: §E.9 · p. 36, Table 15 · p. 36
E.10 Sensitivity to the Backward Inference Step Size
- FB-BI performance is stable for α ∈[0.005, 0.1], including the default α = 0.02, but larger steps degrade performance by collapsing diversity toward z∗.The experiment evaluates Walker over eight step sizes across three seeds; α = 0 recovers FB-ME and α = 1 replaces each solution with z∗.Source: §E.10 · p. 36, Figure 13 · p. 37
E.11 Compute Resources
- Experiments ran on NVIDIA H100 GPUs, and because BFM pretraining is a one-time cost reused across tasks in an environment, QD search dominates amortized per-run cost.Table 16 reports wall-clock time per run.Source: §E.11 · p. 37, Table 16 · p. 37
E.12 Results Table
- Tables 17–20 report final unnormalized QD-scores, averaged over five seeds, for all methods on Walker2D, HalfCheetah, AntMaze-medium, and Cube-single.The tables mark best and second-best scores; Table 19 also identifies the dense-reward variant with a superscript D.Source: §E.12 · p. 38, Table 17 · p. 38, Table 18 · p. 38
2. Limitations
- The paper identifies three limitations: reliance on a diverse offline dataset, degradation with narrower datasets, and a frozen BFM’s hard ceiling on behavioral expressivity.All datasets are public, and the paper reports sufficient Appendix D detail for reimplementation, but does not provide open access to code.Source: §2 · pp. 40, 42
NeurIPS Paper Checklist
- The checklist says Appendix D provides the experimental details needed for reproducibility, including environments, datasets, archive configuration, evaluation budget, model architecture, latent dimension, hyperparameters, seeds, and variation operators.Results average five independent seeds, or three in two additional experiments, with standard deviations shown as learning-curve shading or bar-plot error bars.Source: §2 · pp. 41–42
2. Limitations
- The paper marks [N/A] for whether it involved crowdsourcing or research with human subjects.The checklist defines [N/A] as indicating no crowdsourcing or human-subject research.Source: §2 · p. 45
- The paper reports that no LLMs were needed in the research and that LLMs were not an important, original, or non-standard core-method component.The authors mark the LLM-use declaration [N/A].Source: §2 · p. 46