Technical Summary: Reasoning Emerges from Constrained Inference Manifolds in Large Language Models

Problem Statement

Current evaluations of reasoning in Large Language Models (LLMs) rely predominantly on labeled benchmarks and task accuracy. This approach conflates internal inference quality with dataset alignment and prompting strategies, treating reasoning as an opaque input–output mapping. Consequently, these metrics offer limited insight into how reasoning is internally realized, why models with similar accuracy differ in robustness, or the specific dynamical processes that underlie successful inference. There is a need for a framework that characterizes reasoning as an intrinsic dynamical process unfolding within the model's representation space, independent of task-specific supervision.

Methodology

The authors analyze reasoning as a dynamical process by examining the evolution of internal representations during inference. The study utilizes a diverse set of contemporary autoregressive LLMs (Qwen2.5, Qwen3, Gemma 3, and DeepSeek-R1-Distill-Qwen) ranging from 0.6B to 72B parameters.

Data Collection and Stimuli:

• Stimuli: Generic cognitive stimuli are derived from the "Other" subset of the MMLU benchmark, covering heterogeneous domains (e.g., ethics, law, science) without relying on specific task labels or correctness signals during analysis.
• Trajectory Extraction: For each stimulus, the hidden state of the last token is recorded at every transformer layer and inference step. These sequences form discrete trajectories embedded in the model's high-dimensional representation space.

Geometric and Information-Theoretic Metrics:

1. Intrinsic Dimensionality (ID): Estimated using the Tight Local Intrinsic Dimensionality (TLE) estimator. Two distinct IDs are measured:
   • $D_{stim}$: The ID of stimulus-induced inference trajectories (geometric organization).
   • $D_{world}$: The ID of static vocabulary embeddings (representational expressivity).
2. Information Volume ($V$): A metric quantifying the structured variation preserved within the reasoning manifold, calculated via the determinant of the covariance matrix of centered trajectory states.
3. Stimulus Expansion: To test stability, the stimulus set is progressively expanded by cumulatively adding disjoint question types from MMLU-Other, tracking how $D_{stim}$ evolves with increasing conceptual diversity.

Unified Diagnostic:
A label-free reasoning health score ($H$) is formulated to integrate these structural constraints:
$$H = \log(D_{world}) \cdot \frac{V}{\exp(\varepsilon D_{stim})}$$
where $\varepsilon$ is a fixed penalty parameter. This score is computed solely from internal dynamics and compared against downstream benchmark performance via Spearman rank correlation.

Key Contributions and Results

1. Self-Organization into Low-Dimensional Manifolds

The study finds that inference-time reasoning dynamics consistently self-organize into extremely low-dimensional manifolds, despite residing in high-dimensional embedding spaces.

• Observation: As inference proceeds through network layers, the intrinsic dimensionality of representations rapidly decreases, often stabilizing at fewer than ten degrees of freedom.
• Distinction: This collapse is stimulus-induced and dynamic; static vocabulary embeddings retain near-full dimensionality, indicating that the models possess high representational capacity, but reasoning specifically constrains this space.

2. The Insufficiency of Dimensionality Alone

Low dimensionality is necessary but not sufficient for robust reasoning.

• Finding: Models with similarly compressed trajectories exhibit sharply different inference behaviors. Overly aggressive compression can lead to brittle or information-poor dynamics.
• Information-Geometry Trade-off: Effective reasoning requires a specific regime where geometric compression is accompanied by an increase in information volume ($V$) as depth increases. Deeper layers suppress noise (reducing ID) while amplifying task-relevant conceptual variations (increasing $V$).

3. The Role of Expressive Capacity

Robustness under heterogeneous inputs depends on the model's expressive capacity ($D_{world}$), defined as the intrinsic dimensionality of its static vocabulary.

• Stability Mechanism: Models with higher expressive capacity exhibit slower growth in stimulus-induced dimensionality ($D_{stim}$) as conceptual diversity increases.
• Constraint: Models with insufficient expressive capacity must recruit additional dimensions to encode diverse concepts, leading to trajectory dispersion and structural instability.

4. The Unified Structural Diagnostic ($H$)

The authors introduce the reasoning health score $H$, which integrates geometric organization ($D_{stim}$), information content ($V$), and expressive capacity ($D_{world}$).

• Performance: $H$ shows strong monotonic agreement (Spearman $\rho > 0.9$) with downstream reasoning performance across multiple independent benchmarks (e.g., AIME, GPQA, LiveBench).
• Label-Free Nature: Crucially, $H$ is computed without access to task labels, correctness signals, or evaluation datasets, yet it successfully predicts robustness and generalization.

Significance and Claims

The paper reframes reasoning evaluation by shifting focus from task accuracy to the internal geometry and information flow of inference dynamics.

• Admissible Regime: The authors propose that healthy reasoning occupies a narrow "admissible regime" characterized by three conditions:
  1. Adequate Representational Expressivity: A high-capacity substrate ($D_{world}$) to support diverse concepts.
  2. Spontaneous Manifold Compression: The self-organization of inference trajectories into low-dimensional structures ($D_{stim}$).
  3. Non-Degenerate Information Volume: The preservation of structured variation ($V$) within the compressed manifold.
• Pathological Dynamics: Models violating these constraints exhibit characteristic failure modes, such as diffuse exploration, representational crowding, or degenerate collapse.
• Complementary Framework: The proposed diagnostic does not replace benchmark-centric assessment but offers a complementary, principled lens to isolate intrinsic properties of reasoning from dataset-specific effects. It suggests that reasoning is fundamentally governed by geometric and informational constraints, providing a basis for diagnosing model health and understanding the structural conditions that enable robust inference.

The authors maintain that while this framework delineates a structural window for robust reasoning, it does not predict monotonic scaling laws, and the applicability to non-autoregressive or multimodal systems remains an open question.