Behavior Foundation Models (BFMs)

• The models are specialized pretrained for generating, assessing, or adopting behaviors across various tasks or environments, drawing from diverse data sources and methodologies with a strong emphasis on breadth of behavior and efficiency in task adaptation.

Behavior Foundation Models (BFMs) are pretrained models intended to represent, generate, evaluate, or control reusable behavior across multiple tasks, contexts, agents, or environments. The term encompasses several related research traditions: reward-free reinforcement-learning models that infer policies from successor representations; humanoid-control models that learn reusable motion and whole-body behavior; human-behavior simulators adapted from LLMs; and structured relational models that represent behavior through interacting attributes. Despite their differences, BFMs share a central objective: amortize the acquisition of behavioral knowledge during broad pretraining so that downstream behavior can be selected, adapted, or generated with little or no task-specific retraining.

1. Conceptual scope and defining properties

BFMs are distinguished from ordinary task-specific policies by the intended breadth of their behavioral repertoire and adaptation interface. A conventional controller generally targets one task, reward, embodiment, or command distribution. A BFM instead seeks to learn reusable behavioral priors, latent skills, successor representations, or conditional action distributions that can be reused across downstream objectives.

The humanoid-control literature defines a BFM as a specialized foundation model for controlling agents in dynamical environments, pretrained on large-scale and diverse behavior data and capable of generating or adapting behavior across multiple downstream tasks (Yuan et al., 25 Jun 2025). Relevant pretraining data include motion-capture trajectories, demonstrations, interaction sequences, reward-free transitions, synthetic motion, and online or offline reinforcement-learning experience. The resulting model may be an actor–critic policy, transformer, variational model, diffusion model, adversarial motion prior, graph-based representation, or forward–backward representation.

Four properties recur across BFM formulations:

1. Broad behavioral coverage: the model represents multiple behaviors rather than one task policy.
2. Reusable behavioral priors: pretraining captures regularities such as natural motion, feasible transitions, balance, contact patterns, social scripts, or decision tendencies.
3. Task flexibility: behavior may be conditioned on rewards, goals, demonstrations, language, motion references, task tokens, latent vectors, or environmental context.
4. Efficient adaptation: new tasks may be solved zero-shot, through a small number of online interactions, by latent optimization, or through parameter-efficient adaptation.

The term is used in at least two major senses. In embodied reinforcement learning, BFMs are latent-conditioned or goal-conditioned policies for dynamical systems. In behavioral science, they are models that predict or simulate human decisions, infer subject characteristics, and model contextual effects. These uses are related by the emphasis on reusable behavioral structure but differ substantially in data, objectives, evaluation, and deployment requirements.

A BFM should not be equated with a model that merely produces human-like outputs. Psychological questionnaires, language completions, or motion trajectories may resemble human behavior without revealing a stable or transferable mechanism. RealBehavior therefore distinguishes behavioral resemblance from faithful behavioral characterization and evaluates reproducibility, internal consistency, external consistency, and generalizability (Zhou et al., 2023). Similarly, a model that matches a population distribution may still fail to reproduce individual consistency or causal responses to interventions.

2. Representational foundations

Successor representations and reward-conditioned behavior

A major BFM family is based on successor features, successor measures, and forward–backward representations. For a policy $\pi$, the successor measure describes the discounted distribution of future states:

$$M^\pi(X\mid s,a)=\sum_{t\geq 0}\gamma^t\Pr(s_{t+1}\in X\mid s,a,\pi).$$

For a state feature map $\phi:S\rightarrow\mathbb{R}^d$, successor features are

$$\psi^\pi(s,a)=\mathbb{E}_\pi\left[\sum_{t\geq 0}\gamma^t\phi(s_{t+1})\mid s,a\right].$$

If a reward is linear in the features, $r(s)=z^\top\phi(s)$, then

$Q_r^\pi(s,a)=z^\top\psi^\pi(s,a).$

A pretrained model can therefore infer a task vector from a new reward and query a corresponding policy without task-specific policy training. Forward–backward models factorize successor measures using a forward representation $F$ and a backward representation $B$:

$$M^{\pi_z}(s,a,ds')\approx F(s,a,z)^\top B(s')\,\rho(ds').$$

The task embedding can be obtained from the reward through

$z_r=\mathbb{E}_{s\sim\rho}[r(s)B(s)],$

after which the policy $$M^\pi(X\mid s,a)=\sum_{t\geq 0}\gamma^t\Pr(s_{t+1}\in X\mid s,a,\pi).$$0 is executed (Cetin et al., 2024).

This mechanism provides a precise sense in which a BFM separates task-independent behavior dynamics from task-dependent rewards. Its limitations are equally fundamental: exact zero-shot optimality is restricted to rewards represented in the feature span, and the quality of inference depends on the state distribution used to estimate the task vector.

Nonlinear task representations

Standard successor-feature and FB methods encode rewards linearly. “Finer Behavioral Foundation Models via Auto-Regressive Features and Advantage Weighting” addresses this restriction with auto-regressive backward features. The task vector is divided into blocks:

$$M^\pi(X\mid s,a)=\sum_{t\geq 0}\gamma^t\Pr(s_{t+1}\in X\mid s,a,\pi).$$1

with later blocks conditioned on earlier ones:

$$M^\pi(X\mid s,a)=\sum_{t\geq 0}\gamma^t\Pr(s_{t+1}\in X\mid s,a,\pi).$$2

This creates a nonlinear task manifold rather than a fixed linear subspace. The method, FB-AWARE, combines auto-regressive features with advantage-weighted offline policy optimization, evaluation-based action selection, uncertainty handling through forward ensembles, and offline policy-improvement techniques (Cetin et al., 2024). Its empirical gains are concentrated on spatially precise tasks and tasks requiring generalization beyond the behaviors represented in the training set.

Relational behavioral structure

Vector representations may become inadequate when behavior depends on relations among attributes rather than on attribute values independently. The Behavioral Molecular Structure model represents behavioral attributes as nodes, relations as edges, and complete behaviors as graphs:

$$M^\pi(X\mid s,a)=\sum_{t\geq 0}\gamma^t\Pr(s_{t+1}\in X\mid s,a,\pi).$$3

Its pipeline is

$$M^\pi(X\mid s,a)=\sum_{t\geq 0}\gamma^t\Pr(s_{t+1}\in X\mid s,a,\pi).$$4

BMS uses relational graph convolution, behavior-level pooling, and, for graph generation, GraphVAE. Experiments on crime detection, recommendation, and fraudulent-transaction generation suggest that relational structure becomes more useful as attribute dimensionality and interaction complexity increase, although structural expressiveness does not guarantee better downstream accuracy (Wang et al., 2023).

Self-supervised latent dynamics

Regularized Latent Dynamics Prediction provides a simpler alternative to complex Bellman-based representation objectives. It predicts future encoded states from observed transitions:

$$M^\pi(X\mid s,a)=\sum_{t\geq 0}\gamma^t\Pr(s_{t+1}\in X\mid s,a,\pi).$$5

Because prediction alone can make state features increasingly similar, RLDP adds an orthogonality or diversity penalty:

$$M^\pi(X\mid s,a)=\sum_{t\geq 0}\gamma^t\Pr(s_{t+1}\in X\mid s,a,\pi).$$6

The method uses spherical feature normalization and learns representations before successor-measure and policy learning. It is particularly effective in low-coverage offline datasets, where Bellman-based objectives may extrapolate to unsupported actions (Jajoo et al., 16 Mar 2026). RLDP demonstrates that BFM representation learning need not always be tightly coupled to policy improvement, although its zero-shot capabilities remain constrained by reward linearity, dataset coverage, and feature sufficiency.

3. Training paradigms and model families

Goal-conditioned learning and motion priors

Goal-conditioned BFMs learn policies of the form

$$M^\pi(X\mid s,a)=\sum_{t\geq 0}\gamma^t\Pr(s_{t+1}\in X\mid s,a,\pi).$$7

where $$M^\pi(X\mid s,a)=\sum_{t\geq 0}\gamma^t\Pr(s_{t+1}\in X\mid s,a,\pi).$$8 may be a target state, motion reference, future pose, velocity command, language instruction, object-interaction goal, or latent skill. Humanoid systems use motion capture, demonstrations, synthetic reference motions, teleoperation, and simulation.

Examples include DeepMimic-style motion tracking, AMP, ASE, CALM, CASE, NPMP, ModSkill, MoConVQ, MaskedMimic, HOVER, and related systems. Their learned priors encode coordination, balance, contact transitions, and physically plausible motion. They may support locomotion, reaching, jumping, manipulation, interaction, recovery, and expressive whole-body behaviors (Yuan et al., 25 Jun 2025).

“Behavior Foundation Model: Towards Next-Generation Whole-Body Control System of Humanoid Robots” classifies BFM pretraining into goal-conditioned learning, intrinsic reward-driven learning, and forward–backward representation learning. It identifies Motivo, based on FB learning with conditional policy regularization, as an early zero-shot humanoid BFM and reports motion-tracking success of 83%, 61% of top-line reward-optimization performance, approximately 12 seconds per 300-step episode, and motion diversity of $$M^\pi(X\mid s,a)=\sum_{t\geq 0}\gamma^t\Pr(s_{t+1}\in X\mid s,a,\pi).$$9 (Yuan et al., 25 Jun 2025).

Intrinsic reward-driven learning

Intrinsic-reward methods learn without an externally supplied task objective. Curiosity, random-network distillation, mutual-information maximization, novelty, and state-coverage objectives encourage diverse behavioral exploration. DIAYN maximizes mutual information between latent skills and visited states, while other methods promote novelty or diversity across trajectories.

This approach can provide broad behavioral coverage but may discover unsafe, irrelevant, or physically implausible behaviors. In humanoid systems, interaction requirements are substantial, and broad exploration does not guarantee task usefulness. Intrinsic pretraining is therefore often combined with demonstrations, motion priors, goal conditioning, or safety regularization.

Human-behavior LLMs

Be.FM applies supervised fine-tuning and LoRA to open-source Llama 3.1 models using behavioral-science literature, economic-game records, and survey data. Be.FM 8B and Be.FM 70B are trained to simulate decisions, infer subject and population characteristics, infer contextual factors, and apply behavioral-science knowledge (Xie et al., 29 May 2025).

Its conceptual framework represents behavior as

$\phi:S\rightarrow\mathbb{R}^d$0

where $\phi:S\rightarrow\mathbb{R}^d$1 is an outcome, $\phi:S\rightarrow\mathbb{R}^d$2 denotes subject characteristics, $\phi:S\rightarrow\mathbb{R}^d$3 is context, and $\phi:S\rightarrow\mathbb{R}^d$4 is behavioral knowledge. The training corpus includes American Economic Review workflows, MobLab game decisions, and Big Five survey responses. Evaluations include economic-game distribution simulation, personality prediction, age inference, context inference, research-workflow generation, and International Economics Olympiad problem solving.

OdysSim adopts a related but distinct definition of BFMs as LLMs trained to simulate human behavior at scale. Its SOUL framework organizes 62 datasets and 23 benchmark tasks across conversation and interaction dynamics, social skills, cognition, role and persona, and human-like evaluation. The OdysSim corpus contains approximately 21.4 million interactions and 10 billion tokens, with socially grounded context retrofitted to dialogue data (Zhou et al., 12 Jun 2026).

Generative humanoid behavior models

“Behavior Foundation Model for Humanoid Robots” combines a Conditional Variational Autoencoder with masked online distillation from a privileged simulation proxy. The BFM receives proprioception, goal states, and masks specifying active root, link, or joint targets. It produces target joint positions for a low-level PD controller (Zeng et al., 17 Sep 2025).

The model is trained using AMASS-derived retargeted human motion, PPO-trained proxy behavior, online DAgger-style distillation, KL alignment between privileged and deployable latent distributions, domain randomization, and hard-negative mining. Its control modes include whole-body motion tracking, VR teleoperation, and locomotion. It supports latent interpolation, latent modulation, and residual learning for novel behaviors such as side salto.

Perceptive and terrain-grounded BFMs

Motion-centric BFMs may reproduce a human reference while ignoring whether it is compatible with the robot’s local terrain. The Perceptive Behavior Foundation Model addresses this distinction by preserving raw human motion as the behavioral command while using terrain perception to modify footholds, posture, timing, and contact realization (Wang et al., 6 Jun 2026).

Its Terrain-Conformal Reference Synthesis pipeline constructs terrain-consistent supervision through contact-aware foothold construction, foot-geometry-aware swing optimization, support-aware root reconstruction, collision repair, and multi-point inverse kinematics. A blind adapted-reference teacher is distilled into an identity-gated Transformer student whose terrain pathway initially acts as an identity and learns residual corrections. This separates behavioral intent from environment-specific physical realization.

4. Adaptation and task inference

Zero-shot reward inference

Zero-shot BFM inference supplies a new reward, goal, demonstration, or motion reference after pretraining. In successor-feature models, the reward is projected into the learned feature space and used to select a policy. In motion models, a target state may be mapped through the backward representation, while in masked humanoid models a control-mode mask determines which target components are active.

Zero-shot inference is computationally efficient but can be suboptimal because of representation error, imperfect policy optimization, distribution mismatch, and insufficient state coverage. “Fast Adaptation with Behavioral Foundation Models” shows that the latent policy family may contain better behaviors than the policy selected by standard reward projection (Sikchi et al., 10 Apr 2025).

Latent adaptation

ReLA and LoLA adapt only the latent task representation while freezing the pretrained BFM. ReLA learns a residual critic for the reward component not captured by the pretrained features and updates the latent with off-policy actor–critic learning. LoLA samples nearby latent vectors, evaluates short true-reward rollouts, bootstraps with the frozen BFM value estimate, and updates a latent distribution using a variance-reduced policy-gradient estimator.

Across four BFMs and 64 navigation and locomotion tasks, the methods improve zero-shot returns by approximately 10–40% within a few tens of episodes. LoLA is generally more stable and computationally faster, while ReLA can initially degrade because of critic distribution shift.

Task Tokens adapts a frozen MaskedMimic humanoid BFM by learning a task encoder that maps task observations to a 512-dimensional token. The token is concatenated with prompt and state tokens and passed through the frozen transformer. Only the task encoder and an auxiliary PPO critic are updated (Vainshtein et al., 28 Mar 2025). The reported adaptation cost is approximately 200,000 trainable parameters per task, compared with approximately 9.3 million for PULSE and approximately 25 million for full MaskedMimic fine-tuning.

Dynamics-aware adaptation

Standard FB models can interfere when trained across incompatible transition dynamics. Belief-FB adds a transformer-based dynamics encoder that infers a context embedding from a short reward-free trajectory. Rotation-FB additionally clusters task or policy directions around context embeddings using a von Mises–Fisher distribution (Bobrin et al., 19 May 2025).

These models preserve zero-shot policy extraction while adapting to hidden layouts, friction, wind, and other context changes through in-context inference rather than parameter updates. In randomized Four-Rooms, Pointmass, and AntWind environments, BFB and RFB improve performance on both seen and unseen dynamics, with reported gains of up to approximately two times over relevant baselines.

Active task inference

OpTI-BFM addresses the cost of evaluating a new reward on a large inference dataset. It treats task inference as an uncertainty-aware linear-bandit problem. Given reward-feature observations, it maintains a least-squares estimate $\phi:S\rightarrow\mathbb{R}^d$5 and confidence ellipsoid $\phi:S\rightarrow\mathbb{R}^d$6, then selects the BFM embedding maximizing an optimistic successor-feature value:

$\phi:S\rightarrow\mathbb{R}^d$7

The method actively selects behaviors that both appear rewarding and reduce uncertainty. Under idealized assumptions, its episodic regret is $\phi:S\rightarrow\mathbb{R}^d$8 (Rupf et al., 23 Oct 2025). Empirically, OpTI-BFM reaches Oracle-level performance within approximately five episodes and supports selective reward labeling and nonstationary-reward tracking.

Latent sequence optimization

BFMTrack extends latent adaptation from stationary tasks to time-varying motion tracking. Latent Sequence Optimization initializes one latent per target timestep, executes sampled latent sequences in simulation, and updates the entire latent trajectory using a policy-gradient estimator. Temporally correlated noise imposes smoothness during exploration.

For dense motion tracking, pink-noise LSO reduces EMD from 1.52 to 1.17, DTW from 4.66 to 3.77, MPJPE from 4.7 cm to 3.4 cm, and MMPJPE from 40.6 cm to 32.6 cm relative to the strongest moving-window baseline (Rupf et al., 23 Jun 2026). The method also supports sparse keyframes, but requires minutes of optimization per trajectory and is not currently suitable for interactive or real-time deployment.

5. Humanoid whole-body control

Humanoid BFMs address high-dimensional, underactuated systems with floating-base dynamics, intermittent contact, balance constraints, and simultaneous locomotion and manipulation. Their intended applications include locomotion, navigation, reaching, jumping, striking, sitting, standing, recovery, teleoperation, motion tracking, human–object interaction, and loco-manipulation.

Unified control interfaces

Masking is a central mechanism for combining control modalities. A model may receive root targets, link positions, joint angles, motion references, task tokens, or subsets of these signals. Masked policies infer unspecified body components while preserving a shared behavioral representation.

The 2025 humanoid BFM based on CVAE distillation uses binary masks over root, kinematic-link, and joint-angle controls. The scaling work instead defines eight curated modes, including root, bimanual, root-and-hand, end-effector, upper-body, and whole-body control (Zeng et al., 16 Jul 2026). These formulations aim to make locomotion, teleoperation, motion tracking, and loco-manipulation different interfaces to a common behavioral model.

Scaling behavior models

“Scaling Behavior Foundation Model for Humanoid Robots” identifies three coordinated scaling factors: the learning paradigm, behavioral data, and architecture (Zeng et al., 16 Jul 2026). It reformulates diverse control problems as integrated global-frame motion tracking, increases on-policy PPO rollout quantity, expands reference-motion diversity, and introduces the Humanoid Transformer.

The reported reference corpus contains approximately 102 million frames at 50 Hz from LAFAN, AMASS, OMOMO, GRAB, SnapMoGen, FineDance, BONES-SEED, and Embody3D. The study distinguishes homogeneous scaling, which adds more examples of already represented behaviors, from heterogeneous scaling, which increases behavioral coverage. Heterogeneous scaling produces the strongest gains on previously unseen behavior classes.

The paper reports more than 10% lower local MPKPE and an 82% reduction in global MPKPE relative to existing humanoid controllers on the reported comparison. These results support a conditional scaling principle: additional parameters, frames, or rollout quantity are not sufficient without relevant behavioral coverage and effective on-policy learning.

Perceptive terrain grounding

Perceptive BFM extends motion tracking to terrain-aware realization. It uses a local $\phi:S\rightarrow\mathbb{R}^d$9 height map over a $$\psi^\pi(s,a)=\mathbb{E}_\pi\left[\sum_{t\geq 0}\gamma^t\phi(s_{t+1})\mid s,a\right].$$0 region, a 10-step proprioceptive history, a 21-step reference window, and residual intent and action pathways initialized to zero (Wang et al., 6 Jun 2026). On terrain-conformal reference synthesis, penetration depth decreases from 5.48 cm for Z-offset projection to 2.38 cm, while clearance violation decreases from 14.3% for cubic interpolation plus IK to 7.4%.

The method is distinctive because the raw human reference remains the user-facing behavioral command. Terrain perception changes the physical realization rather than replacing the command with an autonomous terrain-navigation objective.

6. Evaluation, fidelity, and benchmark design

Behavioral fidelity

RealBehavior proposes that psychological measurement of foundation models must test more than questionnaire scores. Its four criteria are reproducibility, internal consistency, external consistency, and generalizability. In its Big Five experiments, agreeableness and conscientiousness are relatively robust, whereas extraversion exhibits severe failures, including negative internal consistency and poor cross-instrument agreement (Zhou et al., 2023).

The results caution against interpreting human-like scores as evidence of human personality. Model responses may reflect item semantics, social desirability, memorized questionnaires, prompt-induced personas, refusal policies, or alignment objectives. A safer interpretation is that models exhibit partially stable, alignment-shaped behavioral dispositions.

Individual and distributional validity

Be.FM and BehaviorBench emphasize that a behavioral model must be evaluated both individually and distributionally. BehaviorBench spans behavior prediction and simulation, strategic decision-making, subject-trait inference, and behavioral knowledge application across 12 task families (Huang et al., 23 Jun 2026).

Individual-level metrics include MAE, accuracy, win rate, and BLEURT. Distributional evaluation uses Wasserstein distance. This distinction is essential because a model that predicts the modal response for nearly everyone may achieve reasonable individual accuracy while eliminating behavioral heterogeneity.

Be.FM-1.5 is reported to achieve substantially stronger distributional alignment than general-purpose models while remaining competitive on individual-level metrics. In Beauty Contest play against humans, it reaches win rates of 48% for the 4B model and 50% for the 70B model, compared with 17% for GPT-4.1 and Claude Sonnet 4.6. On some knowledge-intensive tasks, however, general-purpose frontier models remain substantially stronger.

Human simulation and social interaction

OdysSim’s SOUL framework evaluates conversation, social skills, cognition, role and persona, and human-like judgment across 23 tasks. Its 8B OSim model reaches an overall SOUL score of 64.6, ranks first or tied-first on 8 of 23 tasks, and nearly matches real users on reaction alignment in $$\psi^\pi(s,a)=\mathbb{E}_\pi\left[\sum_{t\geq 0}\gamma^t\phi(s_{t+1})\mid s,a\right].$$1-bench, with scores of 93.2 and 93.5 respectively (Zhou et al., 12 Jun 2026).

These results do not establish complete human realism. The benchmark includes LLM-judge metrics, synthetic or retrofitted social contexts, English-language data, and limited naturalistic behavioral traces. Reward hacking is a major concern: judge-targeting language and extremely short responses can increase evaluation scores without improving underlying behavioral fidelity.

Embodied evaluation

Humanoid BFM evaluation includes motion-tracking error, task success, robustness, naturalness, diversity, sim-to-sim transfer, and real-robot deployment. Task Tokens reports high success across Reach, Direction, Steering, Long Jump, and Strike, but full fine-tuning, PULSE, or PPO can achieve higher success on individual tasks. Its primary advantage is the combination of high success, approximately 200,000 trainable parameters per task, faster convergence, and better preservation of human-like motion (Vainshtein et al., 28 Mar 2025).

BFMTrack reports improvements in dense and sparse tracking but requires offline simulation optimization. Perceptive BFM demonstrates qualitative transfer to a Unitree G1 over stairs, blocks, sparse supports, grass, and mixed obstacle layouts, but does not report repeated real-robot completion rates or standardized collision statistics (Wang et al., 6 Jun 2026).

7. Limitations, safety, and future directions

BFMs remain constrained by data, representation, embodiment, evaluation, and safety limitations.

Representation and coverage limits

Successor-feature BFMs cannot represent rewards outside the learned feature span. Relational models depend on manually designed schemas and may propagate spurious correlations. RLDP preserves feature diversity but does not guarantee semantic sufficiency. Auto-regressive features increase task expressivity but cannot create behaviors absent from the data.

Offline models are also limited by dataset support. Advantage weighting is effective when high-quality trajectories exist in the data but can be harmful on broad exploratory datasets. Latent adaptation can recover only behaviors represented by the pretrained policy family. Dynamics-aware BFMs assume sufficient transition coverage and often assume that environmental context remains static within an episode.

Simulation and embodiment transfer

Most embodied BFM experiments remain simulation-focused. Real-world deployment is affected by actuator latency, friction mismatch, sensor noise, calibration error, unmodeled compliance, contact uncertainty, observation delay, and embodiment changes. Human-motion data do not directly specify robot contact forces, actuator torques, balance margins, or morphology-specific feasibility.

A BFM intended for real humanoids requires safety filters, collision constraints, torque and thermal limits, fall recovery, contact-force regulation, uncertainty estimation, and supervisory fallbacks. Physical plausibility cannot be inferred solely from visual motion similarity.

Evaluation validity

Behavioral scores can be inflated by memorization, prompt sensitivity, synthetic labels, judge-specific optimization, output-format compliance, or distributional shortcuts. RealBehavior’s occasion-based behavioral consistency is not a human behavioral prediction test because synthetic descriptions and model-trained classifiers provide the criterion. Be.FM context inference is qualitative and does not establish causal intervention validity. BehaviorBench’s Wasserstein metric captures distributional differences but not subgroup correlations, multimodality, causal response, or longitudinal consistency.

Future evaluation should combine individual accuracy, population distributions, subgroup calibration, temporal consistency, strategic adaptation, causal intervention tests, cross-cultural validity, uncertainty calibration, and real human comparisons.

Alignment and behavioral diversity

Alignment objectives may produce restricted behavioral profiles. RealBehavior associates later GPT-series models with increased agreeableness and conscientiousness, while noting lower or unstable extraversion and emotional stability (Zhou et al., 2023). In human simulation, helpful-assistant post-training can produce excessive agreement, verbosity, formatting, and task completion, creating a behavioral Sim2Real gap (Zhou et al., 12 Jun 2026).

Diversifying alignment objectives does not imply removing safety constraints. It means separating safety from a single narrow behavioral style and allowing controlled variation in expressiveness, social engagement, emotional tone, and interaction strategy.

Privacy and manipulation

Human-behavior BFMs can infer age, personality, preferences, vulnerabilities, or demographic attributes from behavior. Such capabilities create risks of profiling, discrimination, surveillance, coercive persuasion, and political or commercial manipulation. Behavioral datasets may contain sensitive information, and synthetic participants should not be treated as substitutes for real human subjects without validation.

Future research directions

Important directions include:

• multimodal BFMs combining vision, language, proprioception, tactile sensing, audio, and contact information;
• cross-embodiment and cross-environment transfer;
• temporally hierarchical behavior representations;
• automatic and uncertainty-aware relation discovery;
• richer behavioral datasets with real human trajectories and culturally diverse populations;
• causal and counterfactual behavioral evaluation;
• stable online adaptation under changing dynamics;
• active task inference with selective labeling;
• safety-preserving task-token and latent modulation;
• formal analysis of feature span, representation collapse, and behavioral interference;
• efficient latent sequence optimization;
• standardized benchmarks for zero-shot adaptation, compositional generalization, robustness, naturalness, and sim-to-real performance.

BFMs are therefore best understood as a family of architectures and training paradigms rather than a single model class. Their common ambition is to transform behavior from a collection of separately engineered policies into a reusable substrate that can be selected, adapted, composed, and evaluated across tasks. The strongest evidence so far concerns reward-conditioned policy reuse, latent adaptation, human-motion control, population-level behavioral simulation, and structured behavior representation. The central unresolved question is whether these systems can achieve reliable, interpretable, diverse, and safe behavioral generalization beyond the distributions, embodiments, cultures, and evaluation protocols represented during pretraining.