Technical Summary: When Pooling Fails

Problem Statement

The paper addresses a structural failure mode in planning agents that rely on identity-agnostic pooling (e.g., mean-pooling) to aggregate object embeddings into a single global state vector. While this architecture is standard in model-based reinforcement learning (e.g., DreamerV2/V3), tool-using agents, and robotic planners due to its compactness and differentiability, it imposes an irreversible representational loss.

The core problem is that once object embeddings are pooled, individual object identities are mathematically destroyed. Consequently, an agent may learn perfect planning sequences and achieve high aggregate rewards on benchmarks while remaining structurally "blind" to which specific object it is acting upon. This leads to object grounding failures, where the agent cannot distinguish between homogeneous objects (e.g., selecting the correct file among many in a directory). The paper argues that current engineering responses—scaling model capacity, increasing depth, or adding more training data—are ineffective because they do not address the architectural root cause.

Methodology and Theoretical Framework

1. Formal Definition of Identity Loss

The authors define Identity-Agnostic Pooling (Definition 3.1) as any permutation-invariant function $f: (\mathbb{R}^d)^N \to \mathbb{R}^d$. Following Zaheer et al. [2017], they establish that such functions cannot distinguish between permutations of inputs.

• Lemma 3.2 proves that no deterministic downstream network $h$ can recover the identity of a specific object $i$ from the pooled state $\bar{e}$ with probability 1, provided the embeddings share support. This is a structural impossibility, not a learning limitation.

2. The Information-Theoretic Ceiling

The paper models the pooled state $\bar{e}$ as a Gaussian channel where the signal is the embedding of a single object and the noise is the sum of all other $N-1$ objects plus environmental noise.

• Theorem 3.3 derives a Mutual Information Ceiling:
  $$I(e_i; \bar{e}) \leq \frac{d}{2} \log \left( 1 + \frac{\sigma^2}{\sigma_n^2 + \frac{(N-1)\sigma^2}{N}} \right)$$
  For low noise, this simplifies to $I(e_i; \bar{e}) \approx \frac{d}{2(N-1)}$. This shows that the information available to recover an object's identity decreases monotonically as the number of objects $N$ increases, regardless of model capacity.

3. The Failure Threshold ($N^*$)

By equating the mutual information to the entropy required to distinguish $N$ objects ($\log N$), the authors derive a Failure Threshold (Corollary 3.4):
$$\frac{d}{2(N^*-1)} = \log N^*$$
If the number of objects in a homogeneous category $N$ exceeds $N^*$, grounding is structurally irrecoverable.

4. Orthogonality Results

The paper proves that the ceiling is orthogonal to standard scaling efforts:

• Capacity Orthogonality (Theorem 3.7): Increasing model size cannot increase mutual information beyond the limit set by the pooled state $\bar{e}$ (Data Processing Inequality).
• Depth Orthogonality (Theorem 3.8): Adding layers after pooling cannot reduce the "Grounding Debt" ($D_G$).
• Training Invariance (Corollary 3.9): Training cannot alter the bound, as it depends only on embedding dimension $d$ and object count $N$.

5. Statistical Gravity vs. Architectural Ceiling

The authors distinguish between two failure modes:

1. Statistical Gravity: A training issue where the agent converges to high-frequency object-action pairs due to data imbalance. This is fixable with diverse data.
2. Architectural Ceiling: The fundamental limit derived above. This is not fixable by data diversity or training.

Experimental Validation

The authors validate these claims using a Controlled Filesystem Environment where objects are files and actions are object-indexed (Read, Write, Move, etc.).

Key Findings:

1. Baseline Failure (Config A): An agent using internal commitment (Variant B) achieved 100% success on action-dominated tasks (Log Generation) but 0% on object-identity tasks (Summary Generation), even with minimal object counts ($N \approx 4$). This confirms Lemma 1.
2. Entropy Sensitivity (Config B): Increasing environment density from $N \approx 9$ to $N \approx 55$ caused identity-dependent tasks to collapse to 0% success, while planning tasks remained unaffected. This validates the "Archive Dichotomy."
3. Capacity Scaling (Config D): Scaling the model from 343K to 75M parameters did not improve grounding performance (remained 0%). However, balanced scaling (depth + width) recovered planning capabilities, proving that planning and grounding are orthogonal competencies.
4. Dimension Scaling (Config 6): Varying embedding dimension $d$ (64, 128, 256) shifted the observed collapse threshold ($N^*$) in alignment with theoretical predictions. For $d=64$, collapse occurred at $N \approx 12-15$; for $d=256$, it shifted to $N \approx 40-50$.
5. Cross-System Validation:
   • DreamerV3: Theoretical predictions of $N^*$ align with reported "object vanishing" failures in multi-object manipulation tasks.
   • LLMs: The paper posits that when attention weights flatten over homogeneous tokens, LLMs functionally approximate mean pooling, explaining "lost in the middle" phenomena in large contexts.

Key Contributions

1. Quantifiable Ceiling: The paper provides a closed-form, computable bound ($N^*$) on the maximum number of objects a pooling-based agent can ground, determined solely by embedding dimension $d$ and object count $N$.
2. Proof of Orthogonality: It rigorously demonstrates that model capacity, depth, and training data cannot overcome the information loss inherent in identity-agnostic pooling.
3. Pre-Deployment Diagnostic: A practical framework allowing engineers to calculate $N^*$ before building a system. If $N > N^*$, the architecture is provably insufficient, and the only solutions are increasing $d$ or switching to per-object representations (e.g., Slot Attention, Graph Networks).
4. Differentiation of Failure Modes: It separates "Statistical Gravity" (fixable via data) from the "Architectural Ceiling" (fixable only via architecture), preventing wasted engineering effort on scaling models that are fundamentally incapable of the task.

Significance and Claims

The paper claims that the field has overlooked a fundamental architectural limitation in favor of scaling heuristics. Its significance lies in shifting the paradigm from "train harder" to "measure first."

• Benchmark Limitations: Standard benchmarks measuring task completion or aggregate reward are insufficient, as they mask structural blindness to object identity.
• Architectural Necessity: The paper argues that attention mechanisms are not universally required; rather, a principled decision must be made based on the ratio of object count to embedding dimension.
• Predictive Power: The derived threshold accurately predicts failure points in both controlled experiments and published results from real-world systems (Dreamer, LLMs), suggesting that many observed failures are not bugs but inevitable consequences of the pooling architecture.

The authors conclude that for any system satisfying the properties of object-indexed actions, categorical membership, irreversibility, and global aggregation, the grounding debt is strictly positive when $N > N^*$, rendering the agent structurally incapable of reliable object selection regardless of training.