🤖 machine learning

Implicit Neural Representations of Individual Behavior

This paper introduces Behavioral INR, a self-supervised generative model that adapts implicit neural representations to learn policy identities from unlabeled, heterogeneous behavioral data by representing policies as state-action functions modulated by latent variables, thereby improving policy identifiability in complex continuous settings and handling variable episode lengths and out-of-distribution shifts.

Original authors: Andrew Kang, Priya Narasimhan

Published 2026-06-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Andrew Kang, Priya Narasimhan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you walk into a massive, chaotic dance studio. Inside, hundreds of people are dancing simultaneously. Some are doing ballet, some are breakdancing, and others are just shuffling around. The problem? There are no name tags, no music cues, and no choreography sheets. You just see a blur of movement.

Your goal is to figure out: Who is dancing what? And more importantly, can you tell that "Ballet Dancer A" is the same person even if they are dancing in a different part of the room or at a different speed?

This is exactly the problem Andrew Kang and Priya Narasimhan tackle in their paper, "Implicit Neural Representations of Individual Behavior." They introduce a new tool called Behavioral INR to solve this puzzle.

Here is the breakdown in simple terms:

1. The Old Way: The "Photo Album" Approach

Previous methods tried to solve this by taking a "photo album" of the dance. They would look at a sequence of moves (Step 1, Step 2, Step 3) and try to summarize the whole thing into a single label.

The Flaw: If the dancer changes their speed, the room size, or the lighting, the "photo album" looks totally different. The computer gets confused and thinks it's a new dancer, even though it's the same person. It relies on shortcuts, like "Oh, this dancer always starts with a spin," rather than understanding the style of the dance itself.

2. The New Way: The "Recipe" Approach (Behavioral INR)

The authors realized that a dancer's style isn't just a list of moves; it's a rulebook or a recipe.

The Analogy: Think of a recipe for chocolate cake.
- The Ingredients (State): The current situation (e.g., "The batter is runny" or "The oven is hot").
- The Action: What you do next (e.g., "Add more flour" or "Turn down the heat").
- The Chef (The Policy): The specific person making the cake.

The old methods tried to memorize the sequence of cakes made. The new method, Behavioral INR, tries to learn the Chef's unique recipe. It asks: "Given this specific situation, what is this specific chef likely to do next?"

They use a mathematical trick called an Implicit Neural Representation (INR).

In Vision: Usually, INRs are used to turn a set of coordinates (x, y) into a color (RGB). It's like a magic function that knows exactly what color every pixel in an image should be, no matter how much you zoom in.
In Behavior: They flip this. Instead of coordinates to colors, they map States to Actions. The model learns a continuous "function" that represents a specific policy (a specific agent's behavior).

3. The Secret Sauce: The "Latent Code"

To make this work for many different dancers (or policies), the model uses a Latent Code.

Think of this as a unique ID card or a DNA strand for each dancer.
When the model looks at a new sequence of moves, it tries to find the right ID card that, when plugged into the "recipe," perfectly explains why the dancer moved the way they did.
Crucially, the model doesn't need to be told "This is Dancer A." It figures it out on its own by seeing which ID card fits the data best.

4. Why This is a Big Deal (The "Zoom" Test)

The paper tests this in some very tough scenarios:

Variable Length: Just like you can zoom in or out on a high-resolution image without it getting pixelated, this model can handle a dance that is 10 steps long or 1,000 steps long. It doesn't care about the length; it cares about the underlying rule.
The "Out-of-Distribution" Test: Imagine you train the model on a dancer in a small room, then test them in a giant stadium.
- Old methods fail because the "photo" of the room looks different.
- Behavioral INR succeeds because it learned the recipe. It knows, "If the room is big, this chef still adds the same amount of salt." It identifies the person, not the environment.

5. Where It Works Best

The authors found that this "Recipe" approach shines when the behavior is complex and the environment is tricky.

It wins in complex, continuous worlds (like Formula 1 racing or robotic arms) where the "rules" of the behavior are subtle and hard to guess just by looking at a few moves.
It ties or loses in simpler worlds (like Chess or simple robot tasks) where the behavior is so obvious or repetitive that a simple "photo album" (looking at past moves) is enough to guess who is playing.

Summary

The paper introduces Behavioral INR, a tool that treats a person's behavior not as a list of past actions, but as a continuous, invisible function (a recipe) that dictates how they react to the world.

By learning this "recipe," the AI can identify who is acting, even if they are acting in a new environment, at a different speed, or doing a different number of steps. It's like recognizing a chef's cooking style by their unique way of seasoning, rather than just memorizing the list of dishes they've made.

Technical Summary: Implicit Neural Representations of Individual Behavior

Problem Statement

The paper addresses the challenge of policy representation learning from unlabeled multi-policy behavioral data. In many real-world domains—such as robotics play, sports, racing, and game logs—datasets contain trajectories generated by multiple distinct policies (e.g., different players, drivers, or demonstrators), but the specific policy responsible for each episode is unknown.

The core objective is to infer a latent representation $z_i$ for each episode that recovers the underlying policy identity without supervision. The authors assume that within a single episode, the policy is fixed, but no policy labels, pairwise constraints, or environment parameters are available during training. The representation must be inferred offline and self-supervisedly solely from the episode's state-action structure.

A critical challenge identified is policy-level Out-of-Distribution (OOD) shifts. Standard behavioral OOD settings often vary agents or environments. However, in multi-policy imitation data, two additional axes of shift arise:

State-distribution shift: The same policy is evaluated on different state distributions.
Action-distribution shift: Marginal action statistics change across train/test splits.

The paper argues that many existing methods fail under these shifts because they rely on "shortcut learning"—identifying policies based on marginal state distributions $p(s)$ or action distributions $p(a)$ rather than the true conditional policy map $\pi(a|s)$ .

Methodology: Behavioral INR

The authors introduce Behavioral INR (Implicit Neural Representation), a self-supervised generative model that adapts the concept of INRs from computer vision to behavioral data.

Core Concept

In vision, INRs represent an image as a continuous function mapping spatial coordinates to RGB values. Behavioral INR represents a policy as a function mapping states to actions.

Function Representation: An episode is treated as a set of samples from an underlying state-action function.
Latent Modulation: A shared state-to-action network (an MLP) is modulated by an episode-level latent code $z_i$ via FiLM (Feature-wise Linear Modulation) layers.
Objective: The latent code $z_i$ must explain how actions vary as a function of states, rather than merely summarizing the trajectory history.

Architecture and Training

Decoder: A 3-block FiLM-modulated MLP that takes a state as input and predicts the subsequent action.
Latent Inference: Unlike amortized encoders that map history directly to a latent, Behavioral INR treats representation extraction as function fitting. At test time, the shared INR weights are frozen, and the episode latent $z_i$ is optimized via gradient descent on the observed state-action pairs of that specific episode to minimize prediction error.
Training: The model is trained on batches of state-action history, current states, and next actions using regression loss (continuous) or cross-entropy (discrete). No policy labels are used during training.

Baselines

The paper compares Behavioral INR against several faithful adaptations of existing representation mechanisms that do not require policy labels:

CVAE-Transf./RNN: Trajectory autoencoders using Transformers or RNNs to encode history and decode actions.
VQ-VAE: Discrete trajectory coding.
History-conditioned INR: An INR decoder where the latent is still produced by a standard amortized history encoder (separating the benefit of the INR decoder from the benefit of the fitted latent).
Diffusion Models: Conditional DDPM-style action predictors.
Behavior Embedding Maps (BEM): Non-parametric clustering of state-action pairs.

Key Contributions

Behavioral INR Framework: A novel formulation of policy representation as an implicit neural function, where the latent code modulates a shared state-action mapping.
Policy-Level OOD Definition: A formalization of OOD shifts specific to multi-policy datasets (state-distribution and action-distribution shifts) designed to test whether models learn the conditional policy map rather than marginal shortcuts.
Empirical Analysis of Shortcuts: A systematic evaluation showing that when policy identity can be recovered from state/action marginals, symbolic repetition, or low-dimensional statistics, amortized history encoders remain competitive. However, Behavioral INR significantly outperforms these baselines in "hard" regimes where the conditional state-action function must be recovered.

Results

The authors evaluate on synthetic Gaussian Random Field (GRF) data, MuJoCo (Minari and augmented Hopper), DM Lab Seek-Avoid, Lichess (chess), DROID (robotics), and FastF1 (Formula 1).

Synthetic & MuJoCo Scaling: In controlled settings with longer episodes, more policies, and OOD splits (e.g., Hopper 10x/20x), Behavioral INR consistently achieves higher linear probe accuracy and kNN accuracy for policy identity recovery compared to amortized baselines. While baselines often degrade as episode length and OOD difficulty increase, Behavioral INR maintains near-perfect probe accuracy.
Real-World Domains:
- FastF1: Behavioral INR is the strongest performer, demonstrating that fitted functional representations scale better when the number of policies increases and shortcuts are weak.
- DROID: Results are competitive, though action marginals provide strong shortcuts in this low-dimensional setting.
- Lichess: Performance is lower than baselines. The authors attribute this to the domain's structure: symbolic board states, repeated openings, and sparse move distributions allow other methods to exploit shortcuts that Behavioral INR does not rely on.
OOD Robustness: Behavioral INR is most effective when the task requires inferring the complex conditional map $\pi(a|s)$ rather than relying on $p(s)$ or $p(a)$ . In settings where policies overlap in states or actions, Behavioral INR successfully recovers identity where other methods fail.

Significance and Claims

The paper claims that Behavioral INR provides a more robust representation for policy identity in continuous, complex state-action spaces, particularly when data is unlabeled and distribution shifts are present.

The authors emphasize that their results do not suggest Behavioral INR dominates in every domain. Instead, they identify a specific regime:

When it works best: When policy identity must be inferred from a complex, non-linear state-conditioned function, and marginal shortcuts (state or action frequencies) are insufficient or misleading.
When it is competitive: When policy identity can be recovered from state/action shortcuts, symbolic repetition, or low-dimensional action statistics, amortized history encoders remain effective.

The work suggests that treating behavior as a function to be fitted, rather than a sequence to be summarized, offers a distinct advantage for understanding heterogeneous datasets in robotics, gaming, and sports, provided the underlying policy structure is sufficiently complex to prevent shortcut learning.