Technical Summary: OmniSapiens: A Foundation Model for Social Behavior Processing via Heterogeneity-Aware Relative Policy Optimization

1. Problem Statement

Socially intelligent AI systems require the ability to reason across diverse human behavioral tasks and generalize to new social contexts. However, existing AI systems are predominantly specialized for single behavioral tasks (e.g., sentiment analysis or emotion recognition), lacking unified models that can effectively generalize across diverse behavioral contexts.

The primary challenge in developing unified social behavioral models is the inherent heterogeneity of behavioral data. This data spans distinct feature types (e.g., acoustic prosody, interpersonal interaction, text, vision) and prediction targets (e.g., emotion classification, social inference, pathological state detection). This heterogeneity produces uneven training signals across samples, leading to imbalanced learning dynamics.

Current unified models often employ Reinforcement Learning (RL) methods like Group Relative Policy Optimization (GRPO) to enhance reasoning. However, these methods aggregate policy updates across heterogeneous samples without explicitly regulating disparities in their learning dynamics. Consequently, optimization can become disproportionately dominated by specific behavioral tasks or samples with systematically elevated advantage magnitudes, hindering multitask performance and cross-context generalization.

2. Methodology

2.1 Model Architecture: OmniSapiens-7B 2.0

The authors introduce OMNISAPIENS-7B 2.0, a foundation model for unified social behavior analysis.

• Backbone: The model utilizes Qwen 2.5 Omni-7B as its multimodal base architecture, capable of processing text, vision, and audio inputs.
• Training Setup: It is trained on the Human Behavior Atlas benchmark, covering over 100,000 samples across 10 diverse behavioral tasks: Sentiment (SEN), Emotion Recognition (EMO), Social Reasoning (SOC), Intent Recognition (INT), Non-Verbal Communication (NVC), Humor (HUM), Sarcasm (SAR), Anxiety (ANX), Depression (DEP), and PTSD (PTSD).
• Objective: A single shared policy $\pi_\theta$ is trained to maximize performance across all tasks simultaneously.

2.2 Core Algorithm: Heterogeneity-Aware Relative Policy Optimization (HARPO)

To address the imbalance in learning signals, the paper proposes HARPO, a critic-free, on-policy RL method. HARPO modifies the standard GRPO objective by dynamically rebalancing learning signals across heterogeneous samples and tasks.

Key Mechanisms of HARPO:

1. Contribution Signal Approximation:
   Since policy gradients are scaled by advantage values, the magnitude of advantages serves as a proxy for a sample's contribution to the policy update. HARPO computes contribution signals at two levels:

   • Sample-level ($p^{(t)}_{(m,q)}$): The average absolute group-normalized advantage for a specific sample $q$ within task $m$.
   • Task-level ($p^{(t)}_m$): The average absolute group-normalized advantage across all samples for task $m$.

2. Geometrically Centered Modulation:
   To prevent any single task or sample from disproportionately influencing the update, HARPO scales advantages based on their deviation from a geometric mean reference.

   • A geometric mean reference ($\bar{p}^{(t)}_{ref,m}$ for samples, $\bar{p}^{(t)}_{ref,M}$ for tasks) is calculated.
   • Modulation factors ($s^{(t)}_{(m,q)}$ and $s^{(t)}_m$) are derived as the reciprocal of the ratio between the contribution signal and the geometric mean.
   • Effect: Samples or tasks with contribution signals above the reference receive factors $<1$ (downscaling), while those below receive factors $>1$ (upscaling).
   • Stability: Because the factors are constructed from a geometric mean, their geometric mean is exactly 1. This ensures that multiplicative upscaling is exactly compensated by downscaling, preventing unintended changes to the global step size.

3. Inertial Smoothing:
   To ensure stability and respond to persistent trends rather than stochastic on-policy fluctuations, the contribution signals and modulation factors are updated using inertial control:

   • Contribution signals use an Exponential Moving Average (EMA).
   • Modulation factors (being multiplicative ratios) are smoothed via multiplicative updates.

Objective Function:
HARPO retains the PPO-style trust-region objective of GRPO but substitutes the standard group-normalized advantage $\hat{A}$ with the HARPO-modulated advantage $A^H$:
$$A^H_{(m,q,i)} = s^{(t)}_{(m,q)} \cdot s^{(t)}_m \cdot \hat{A}^{(t)}_{(m,q,i)}$$

2.3 Reward Design

The model is trained with a composite reward function $r$:
$$r = (1 - w_{fmt}) r_{task} + w_{fmt} r_{fmt} + \lambda_{len} r_{len}$$

• $r_{task}$: Binary reward for exact label match (classification) or cosine similarity for free-text QA.
• $r_{fmt}$: Binary reward for adhering to the required output structure (reasoning trace followed by prediction).
• $r_{len}$: Penalty for excessive response length.

3. Key Contributions

1. OmniSapiens-7B 2.0: The release of a unified foundation model capable of jointly tackling a wide range of behavioral analysis tasks spanning affective, cognitive, pathological, and social dimensions.
2. HARPO: The introduction of a new RL method designed specifically to learn from heterogeneous behavioral data by rebalancing learning signals without a critic network.
3. Interpretability: The model produces more interpretable reasoning traces, improving transparency and reliability for real-world applications.

4. Experimental Results

4.1 Multitask Performance

Evaluated across 10 behavioral tasks, OmniSapiens-7B 2.0 achieved the best and most consistent performance compared to:

• General-purpose LLMs: (e.g., Gemma-4, Qwen-2.5-Omni).
• Specialized Behavioral Models: (e.g., HumanOmniV2, previous OmniSapiens variants).
• RL Algorithms: Outperformed GRPO, RLOO, RE++, GPG, and EMAGRPO.

Key Metrics:

• Achieved the highest average task performance rank (1.90) among all models.
• Improved performance by up to +12.02% over prior state-of-the-art models.
• Demonstrated robustness where other methods (like GRPO) showed performance collapse on specific tasks (e.g., Sarcasm detection).

4.2 Zero-Shot Generalization

The model was evaluated on five held-out behavioral benchmarks (Autism Detection, Speech Emotion Recognition, In-the-wild Depression Recognition, Social Media Sentiment, Sarcasm Incongruity) without fine-tuning.

• OmniSapiens-7B 2.0 outperformed all baselines on all five held-out tasks.
• Gains reached up to +9.37% compared to prior models.
• This suggests the model learns transferable behavioral features that generalize beyond the training distribution.

4.3 Reasoning Trace Interpretability

The paper evaluates the quality of reasoning traces using XAI desiderata:

• Robustness: Achieved the highest consistency (87.7%) across independently sampled traces and paraphrase variations (88.4%).
• Efficiency: Generated significantly more concise reasoning traces (19.86 tokens) compared to baselines (e.g., 57.69 tokens for OmniSapiens-7B RL).
• Human Evaluation: In pairwise comparisons, OmniSapiens-7B 2.0 was preferred or tied with baselines in 99.2% of cases for concision, 88.3% for coherence, and 78.9% for specificity.

4.4 Ablation Studies

• Structured Modulation: Removing the geometric reference (using simple reciprocal scaling) degraded performance (Avg. Rank 2.00 vs 1.90), confirming the necessity of geometric centering to prevent global step size distortion.
• Inertial Control: Removing inertial smoothing significantly degraded performance (Avg. Rank 2.70), particularly for tasks with fluctuating contribution signals (e.g., Sarcasm, Sentiment).
• Sample-Level Modulation: Removing sample-level modulation (task-level only) degraded performance (Avg. Rank 2.60), indicating the need to balance signals at the individual sample level to suppress extreme advantage values.

5. Significance and Claims

The paper claims that OmniSapiens-7B 2.0 represents a significant step toward unified social behavior analysis. By explicitly modeling heterogeneous learning signals through HARPO, the model overcomes the limitations of existing unified approaches that are constrained by imbalanced learning dynamics.

The authors assert that their work:

• Unlocks potential applications in affective computing, behavioral health, and socially intelligent AI systems, particularly in settings where behavioral data is scarce and model interpretability is critical.
• Demonstrates that protecting weaker, structurally meaningful signals (via HARPO) allows foundation models to encode latent social and behavioral structures more effectively.
• Provides a reliable framework for real-world behavioral applications by generating reasoning traces that are not only accurate but also robust, efficient, and interpretable.

The work suggests that future social behavioral foundation models may benefit from explicitly addressing the heterogeneity of learning signals rather than treating all tasks and samples uniformly.