🤖 AI

Estimating the Empowerment of Language Model Agents

This paper introduces EELMA, an information-theoretic framework that estimates language model agent empowerment through multi-turn text interactions, demonstrating that this goal-agnostic metric effectively correlates with task performance and identifies pivotal moments for general capabilities across diverse environments.

Original authors: Jinyeop Song, Jeff Gore, Max Kleiman-Weiner

Published 2026-05-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jinyeop Song, Jeff Gore, Max Kleiman-Weiner

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 are trying to judge how good a new robot assistant is. Usually, you give it a specific list of chores—like "wash the dishes" or "find a recipe"—and you only count how many chores it finishes. If it fails the dishes but finds a great recipe, you might say it's "bad at dishes" but "good at cooking."

But what if you want to know if the robot is generally capable, even before you tell it what to do? What if you want to know if it has the potential to handle any future situation, not just the ones you planned?

This paper introduces a new way to measure that potential. The authors call it "Empowerment."

The Core Idea: The "Key to the Future"

Think of Empowerment as a measure of how many different "doors" an agent can open in the future.

Low Empowerment: Imagine a robot trapped in a small, empty room with only one door that leads back to where it started. No matter what it does, it can only go in circles. It has very little control over its future.
High Empowerment: Now imagine a robot in a huge mansion with many rooms, elevators, and secret passages. If it chooses to go left, it can reach the kitchen. If it goes right, it can reach the library. It has many options. It can influence where it ends up.

The paper argues that an agent (like a Large Language Model) that has high empowerment is generally smarter and more capable because it can navigate to a wider variety of future states. It's not just about finishing a specific task; it's about having the ability to reach many different places.

The Problem: The "Word Soup" Challenge

The authors wanted to measure this for AI agents that talk and read (Language Model Agents). But there's a catch:

Traditional methods for measuring empowerment work well in simple, grid-like video games where every state is a unique number (like "Square 1," "Square 2").
But AI agents deal with text. The same situation can be described in a million different ways.
- Example: "The agent is at (2, 3)" and "The robot stands at x=2, y=3" are the exact same state, but to a computer counting strings, they look completely different.
If you try to count these text variations like you count grid squares, the math breaks down. The computer gets confused by the "word soup" and can't tell how much control the agent actually has.

The Solution: EELMA (The "Translator")

To fix this, the authors created a tool called EELMA.

Think of EELMA as a smart translator and summarizer.

It reads the text: It takes the messy, variable text descriptions of the world (observations) and the actions the agent takes.
It compresses them: Instead of treating every sentence as unique, it translates them into a compact "summary code" (an embedding). It realizes that "The agent is at (2,3)" and "Robot stands at x=2, y=3" are actually the same summary code.
It counts the possibilities: Once the text is translated into these clean summaries, EELMA can finally count how many different future paths the agent can actually reach.

What They Found

The team tested EELMA in three different "worlds":

Text-based Games: Like a digital version of "The Wizard of Oz" or "Hanoi Tower" where you type commands.
Web Browsing: Agents trying to navigate real websites (like finding a product price or logging into a site).
Tool Use: Agents trying to use software tools (like booking a flight or managing a retail order).

The Results:

It Works: In the games, EELMA's measurements matched perfectly with the actual mathematical "true" empowerment.
It Predicts Success: They found a strong link: Agents with high empowerment scores generally got better at their tasks (like finding the right webpage or solving the puzzle), even though EELMA wasn't told what the specific goal was.
It Spots "Power Moves": In the web browsing tests, they looked at a moment where an agent had to log in with a password.
- When the agent typed the correct password, its empowerment score spiked. Why? Because logging in unlocked a whole new set of future possibilities (accessing admin panels, changing settings).
- When it typed the wrong password, the score stayed low.
- This shows EELMA can automatically detect "pivotal moments" where an agent gains significant control, without needing a human to say, "Hey, that was a good login!"

Why This Matters

Currently, to test AI, we have to build expensive, hand-crafted tests for every single task (e.g., "Can it book a flight?" "Can it write code?"). This is slow and expensive.

This paper suggests we can use Empowerment as a "universal health check" for AI agents.

It doesn't care what the specific goal is.
It tells you if the agent has the capacity to control its environment.
It helps us spot when an agent is getting stuck in loops (low empowerment) or when it's finding powerful new ways to interact with the world (high empowerment).

In short, instead of just asking "Did it finish the job?", this method asks, "Does it have the keys to the whole building?"

Technical Summary: Estimating the Empowerment of Language Model Agents

Problem Statement

As Large Language Model (LM) agents become increasingly integrated into real-world systems for tasks like web browsing, tool use, and multi-agent interaction, evaluating their capabilities and safety remains a significant challenge. Current evaluation paradigms rely heavily on goal-centric benchmarks (e.g., WebArena, SWE-bench) where human-designed objectives serve as proxies for capability. This approach suffers from two primary limitations:

Labor Intensity: Designing large-scale, diverse evaluation tasks is resource-heavy.
Blind Spots: Traditional metrics restrict focus to specific end goals or milestones, potentially overlooking the open-ended nature of agentic interactions. This limits the ability to detect agents pursuing objectives outside the measured scope or to track capability growth that does not immediately translate to task success, posing risks for AI safety.

There is a critical need for a scalable, goal-agnostic evaluation framework that can assess an agent's general capabilities without relying on manually crafted rewards or specific task definitions.

Methodology: EELMA

The authors propose EELMA (Estimating Empowerment of Language Model Agents), a framework designed to estimate empowerment—an information-theoretic measure of an agent's influence on future states through its actions.

Theoretical Foundation

Empowerment is defined as the channel capacity between an agent's actions and its subsequent sensory inputs, capturing the maximal mutual information between actions and future states. Formally, for an agent with policy $\pi_{LM}$ , the effective empowerment $E(\pi_{LM})$ is the average conditional mutual information over states encountered:
$E(\pi_{LM}) \triangleq \mathbb{E}_{s_t} [I(a_t; s^* | s_t)]$
where $s^*$ is a future state sampled $\tau$ steps ahead. Under the assumption of uniformly distributed rewards, empowerment serves as a lower bound for the mean discounted reward, suggesting it acts as a proxy for general agentic capability.

The EELMA Algorithm

Classical empowerment estimators (based on counting transitions) fail in text-based environments due to:

Observation Redundancy: Multiple natural language strings can describe the same underlying state (e.g., "located at x=1" vs. "at position 1").
High Dimensionality: The observation space is too large for direct probability estimation.

EELMA addresses these by mapping textual observations and actions to compact embeddings and estimating mutual information via InfoNCE (Noise Contrastive Estimation). The architecture consists of:

Embedding Layer: A pretrained text encoder (e.g., E5, MiniLM) with a fine-tunable MLP projection (using LoRA) maps observations ( $o_t$ ), actions ( $a_t$ ), and future observations ( $o^*$ ) to latent vectors.
Neural Encoders: Two encoders, $\phi$ (for current state and state-action pairs) and $\psi$ (for future states), are trained jointly.
Loss Function: The model maximizes two InfoNCE objectives to approximate the conditional mutual information:
- $I(o_t; o^*)$ : Observations only.
- $I(o_t, a_t; o^*)$ : Observations and actions.

The final empowerment estimate $\hat{E}$ is derived by subtracting the log-density ratio of the observation-only term from the state-action term, effectively isolating the information the action provides about the future state.

Key Contributions

Goal-Agnostic Metric: The paper proposes empowerment as a metric for evaluating LM-agents that does not require hand-labeled goals or task-specific rewards, complementing traditional task-success measures.
EELMA Framework: The development of the first empowerment estimator specifically designed for text-based LM-agents, capable of handling linguistic variability and high-dimensional observation spaces.
Empirical Validation: Extensive experiments demonstrating that empowerment strongly correlates with average task performance across textual games (Gridworld, Tower of Hanoi) and realistic environments (WebArena, $\tau$ -bench).
Behavioral Analysis: Characterization of how agent design choices (prompts, memory, model scale) modulate empowerment, and the identification of high-empowerment actions as pivotal moments (e.g., authentication) without manual annotation.

Results

1. Accuracy in Controlled Environments

In Gridworld and Tower of Hanoi tasks, EELMA's estimates closely aligned with ground-truth direct estimation (Spearman correlation $R_s = 0.80\text{--}0.94$ ).

Robustness to Linguistic Variability: Unlike direct estimation, which collapsed under paraphrased natural language observations, EELMA maintained high accuracy, demonstrating its ability to generalize across semantically equivalent but syntactically diverse descriptions.
Complexity Scaling: Effective empowerment decreased as environmental complexity increased (e.g., more boxes in Gridworld, more disks in ToH), mirroring the decline in mean discounted reward.

2. Correlation with Task Performance

Across various ablations (Chain-of-Thought prompting, memory context length, model scale), effective empowerment showed strong positive correlations with mean discounted reward:

CoT: Removing Chain-of-Thought reasoning drastically reduced empowerment (e.g., 99% drop in Gridworld).
Memory: Increasing memory context (0 to 3 steps) progressively increased empowerment.
Model Scale: Larger models (e.g., GPT-4o, Claude-3.5-Sonnet, Qwen2.5-32B) generally exhibited higher empowerment than smaller counterparts.

3. Real-World Environments (WebArena & $\tau$ -bench)

WebArena: Empowerment scores correlated strongly with discounted reward in GitLab, Reddit, and Shopping Admin domains ( $R_s = 0.83\text{--}0.94$ $R_{s} = 0.83 - 0.94$ ).
- Anomaly: In the Shopping domain, empowerment was high but task performance varied, suggesting reasoning bottlenecks (numerical price analysis) rather than navigation control limited performance.
- Jailbreaking Detection: The Qwen model showed high empowerment in Reddit tasks but low task success. Analysis revealed this was due to "jailbreaking" behavior (navigating to real external websites), which artificially inflated state diversity. This highlights a boundary condition where high empowerment reflects output diversity rather than task-relevant control.
$\tau$ -bench (Tool Use): In the airline domain, empowerment scaled monotonically with model capability. In the retail domain (where all models performed near floor level), empowerment inverted, as smaller models generated diverse but invalid tool calls, inflating mutual information without task relevance.

4. Pivotal Moment Detection

In a modified WebArena task requiring manual authentication, EELMA successfully detected the pivotal moment of successful login. Valid authentication attempts (entering correct credentials) resulted in significantly higher empowerment scores compared to invalid attempts, demonstrating the metric's utility for monitoring power-seeking behaviors without explicit reward signals.

Significance and Claims

The paper claims that empowerment serves as a robust, goal-agnostic proxy for general agentic capability. Its primary significance lies in:

Scalability: It offers a method to evaluate agents without the prohibitive cost of designing new benchmarks for every new capability.
Safety Monitoring: It can detect "pivotal moments" where an agent gains significant control over its environment (e.g., accessing admin privileges), potentially flagging power-seeking behaviors before they manifest in specific harmful goals.
Complementarity: It is not a replacement for task-success metrics but a complementary signal that captures the potential for control and influence, which is crucial for understanding agents in open-ended, reward-sparse environments.

The authors acknowledge limitations: empowerment measures control over the environment, not necessarily task success. In scenarios where the bottleneck is reasoning rather than state reachability, or when agents exhibit high diversity in task-irrelevant behaviors, empowerment may not predict success. However, it remains a valuable tool for characterizing the landscape of agent capabilities.