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, it offers limited insight into how reasoning is internally realized, why models with similar accuracy differ in robustness, and what structural properties distinguish healthy reasoning from pathological dynamics. This work addresses the need to study reasoning as an intrinsic dynamical process unfolding within the representation space, independent of task-specific supervision.

Methodology

The authors analyze inference-time representation trajectories elicited by generic cognitive stimuli (derived from the "Other" subset of the MMLU benchmark) across a diverse set of contemporary autoregressive LLMs (Qwen2.5, Qwen3, Gemma 3, and DeepSeek-R1-Distill-Qwen families, ranging from 0.5B to 72B parameters).

The core methodology involves:

1. Trajectory Extraction: For each stimulus, the hidden state of the last token is recorded at every transformer layer and inference step, forming a discrete trajectory $T_\ell(x)$ in the high-dimensional representation space.
2. Geometric Analysis:
   • Intrinsic Dimensionality (ID): Estimated using the Tight Local Intrinsic Dimensionality (TLE) estimator to quantify the effective degrees of freedom of the trajectories. Two distinct IDs are measured: $D_{world}$ (from static vocabulary embeddings) and $D_{stim}$ (from inference-time trajectories).
   • Information Volume ($V$): Calculated via the log-determinant of the covariance matrix of centered trajectory points to quantify the amount of structured variation preserved within the manifold.
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.
4. Unified Diagnostic: A scalar metric, the Reasoning Health Score ($H$), is formulated to integrate geometric organization, information content, and expressive capacity without using task labels or correctness signals.

Key Contributions

The paper identifies three structural constraints that jointly govern healthy reasoning dynamics and introduces a label-free diagnostic based on these constraints:

1. Spontaneous Manifold Compression: Inference-time representations spontaneously collapse onto low-dimensional manifolds embedded within high-dimensional spaces. This dimensional collapse is stimulus-induced, reproducible, and distinct from the high expressivity of the static vocabulary embeddings.
2. Non-Degenerate Information Flow: Low dimensionality alone is insufficient. Effective reasoning requires that the compressed manifold maintains high information volume. The authors observe that while ID decreases with network depth, information volume increases, indicating a focusing mechanism that suppresses noise while amplifying task-relevant variations.
3. Expressive Capacity as a Stabilizer: The intrinsic dimensionality of static vocabulary embeddings ($D_{world}$) acts as a background structural condition. Models with higher expressive capacity can accommodate increasing conceptual diversity without requiring a substantial expansion of inference-time degrees of freedom, thereby stabilizing low-dimensional reasoning under heterogeneous stimuli.

Results

• Dimensional Collapse: Across all model families and scales, reasoning trajectories consistently stabilize at intrinsic dimensionalities far below the ambient embedding dimension (often converging to fewer than ten degrees of freedom).
• Decoupling of Compression and Performance: Stronger geometric compression (lower ID) does not linearly correlate with better reasoning performance. Models with similarly low ID can exhibit sharply different inference behaviors, including brittle or information-poor dynamics.
• The Admissible Regime: High-performing models concentrate in a narrow region of a three-dimensional space defined by low $D_{stim}$, high $V$, and sufficient $D_{world}$. Models outside this regime exhibit pathological dynamics such as diffuse exploration, representational crowding, or degenerate collapse.
• Diagnostic Performance: The unified reasoning health score ($H$) shows strong monotonic agreement (Spearman correlation coefficients > 0.9) with downstream reasoning performance across diverse benchmarks (e.g., AIME, GPQA, LiveBench). This correlation holds despite $H$ being entirely task-agnostic and computed solely from internal dynamics.

Significance and Claims

The paper reframes reasoning evaluation by shifting focus from task outcomes to the internal geometry and information flow of inference dynamics. The authors claim that reasoning in LLMs is fundamentally governed by geometric and informational constraints.

• Complementary Framework: The proposed structural diagnostic does not replace benchmark-centric assessment but offers a complementary lens. It isolates intrinsic properties of inference dynamics from dataset-specific effects, allowing for the comparison of reasoning behavior across models and architectures without reliance on labeled data.
• Mechanistic Insight: The findings suggest that robust reasoning emerges not from a single mechanism but from the coordinated interaction of low-dimensional organization, sufficient information preservation, and a rich representational substrate.
• Diagnostic Utility: The work provides a principled means to identify pathological regimes (e.g., over-compression or information starvation) and offers a structural reference for evaluating the effects of future training approaches or alignment techniques on internal computation.

The authors explicitly note limitations: the diagnostic is descriptive and correlational, does not establish causal mechanisms, and relies on last-token hidden states as a proxy for distributed computation. Furthermore, the applicability of these admissible regimes to non-autoregressive, multimodal, or embodied systems remains an open question.