Source-linked AI summary
The Dynamics of Continuous Mixture Collapse in Language Models
Ali Backour
TL;DR
Continuous latent reasoning aims to preserve multiple possible reasoning directions, but pretrained models often fail to preserve the underlying mixtures. The paper combines theoretical analysis with controlled experiments to identify three independent failure sources: architectural and training-induced distortion, plus softmax-feedback dynamics that either amplify small differences or erase distinctions. It further shows that preserving K-way mixtures generally requires context-dependent correction whose dimensionality can grow with K.
Problem
Pretrained language models often collapse or ignore continuous reasoning states, motivating analysis of why mixture information is not reliably preserved.
Method
The paper combines comparisons with matched random controls, idealized linear-transport analysis, recursive softmax dynamics, pretrained-model rollouts, and a K-component generalization.
Results
The paper identifies architectural distortion, training-amplified distortion, and softmax-feedback dynamics that either amplify mixture imbalances or contract distinct mixtures into indistinguishable states.
Takeaways & Limitations
Continuity alone does not preserve continuous states because representation, readout, and closed-loop dynamics must jointly preserve their assigned semantics.
Takeaways & Limitations
The empirical analysis is primarily binary and covers two model families, while the K-way result is theoretical and its full-rank condition is not directly measured.
Abstract
from arXiv · showhide
LLMs latent-state reasoning methods replace discrete intermediate tokens with continuous states, such as weighted mixtures of token embeddings, to retain multiple possible reasoning directions rather than committing to one. Yet pretrained language models often fail to preserve these mixtures. We study why through a combination of theoretical analysis and controlled empirical investigations on a variety of models. We identify three independent, distinct sources of failure. First, transformer architectures already distort mixture geometry, and training substantially amplifies this effect. Moreover, the failure can occur even if the model transports mixtures perfectly linearly: the softmax readout and autoregressive feedback form a dynamical system that either amplifies small differences until one component of the mixture dominates or contracts different mixtures until they become indistinguishable. We verify this theoretical prediction empirically: the observed transition between contraction and amplification occurs near the theoretical threshold derived by our analysis, and pretrained-model rollouts lie predominantly on the amplifying side. Finally, we generalize to mixtures of many components and show that exact preservation generally requires context-dependent correction, whose required dimensionality can grow with the number of components.
1 Introduction
Continuous reasoning replaces discrete token commitments with states that can preserve multiple reasoning directions, but pretrained models often collapse or ignore these mixtures. This paper investigates why, identifying architectural, learned, and dynamical sources of failure.
- Discrete chain-of-thought commits to one token at each step, discarding alternative candidates and potentially limiting abstract reasoning.
- Continuous reasoning represents intermediate states as mixtures or hidden states, enabling compressed computation and parallel propagation of multiple reasoning directions.
- Prior work finds pretrained models often let the highest-probability component dominate or collapse continuous states rather than preserving them.
- The paper identifies three independent failure sources: architectural distortion, training-amplified distortion, and softmax-feedback dynamics that amplify or erase mixture differences.
- For K components, exact preservation generally requires context-dependent correction whose required information can grow with the number of components.
2 Where Continuous Mixtures Lose Their Geometry
Controlled embedding-mixture experiments measure whether transformer hidden states preserve requested mixture geometry. Across models and depths, trained networks develop collapse-like responses, while matched untrained controls remain closer to faithful preservation.
- The experiment injects a convex combination of token embeddings into an interior slot and evaluates the next-token distribution after that slot.For component embeddings e_i and simplex weights p_i, the injected state is e(p) = Σ_i p_i e_i.
- A 50/25/25 mixture expected to yield 50% red, 25% green, and 25% blue instead produced 99.5% red on average.
- Figure 2 compares trained models with five same-architecture random controls across normalized depth, showing increasingly step-like trained responses while controls remain closer to the preservation diagonal.
- Across 325 three-way mixtures, both tested models map most interior points to near-pure components with abrupt transitions across narrow boundaries.
- CALIB equals 1 for exact preservation and 0 for a hard threshold at w = 0.5, providing the section’s summary measure of mixture fidelity.
3 Mixture Collapse as a Dynamical System
Even with perfect linear mixture transport, softmax readout and autoregressive feedback create recursive dynamics that either amplify an initial majority or erase distinctions between mixtures. The critical coupling is L = 2, and pretrained-model trajectories predominantly operate in the amplifying regime.
- Recursive dynamics: Softmax converts linear logit interpolation into a weighted geometric probability mixture, whose autoregressive feedback creates a recursive dynamical system.The model evaluates pure components separately, interpolates their logits, applies softmax, and feeds the resulting distribution into the next step.
- Recursive dynamics: The recursion tracks mixture imbalance with coupling L_t and field b_t, where L_t measures differing branch preferences and b_t their common log-odds offset.The next log-odds is ℓ_t+1 = b_t + L_tu_t under exact linear logit transport.
- Critical threshold: For constant coupling, the balanced state changes stability at L = 2: it is stable below the threshold and unstable above it.Above the threshold, two additional stable polarized fixed points appear.
- Empirical verification: Because language-model contexts evolve, L_t is time-varying rather than fixed, making the dynamics non-autonomous relative to the classical constant-coupling analysis.The theoretical results therefore characterize time-varying dynamics using uniform supercritical or subcritical conditions.
- Critical threshold: Above the critical value, the initial majority is preserved and amplified toward a polarized state, while subcritical coupling contracts distinct mixtures toward indistinguishability.For example, the polarized fixed point is approximately 0.96 at L = 4 and 0.9993 at L = 8; under subcritical coupling with b_t = 0, u_t approaches zero independently of initialization.
4 Recursive Softmax in Pretrained Language Models
Recursive softmax feedback can destroy a mixture even when transport into logit space is exactly linear. Pretrained trajectories predominantly operate above the critical coupling, where small initial differences are amplified toward opposite branches, while lower coupling makes mixtures contract and lose distinctions.
- Experimental setup: The experiment removes nonlinear transformer processing from the mixture path, leaving softmax as the mixture’s only nonlinearity.The two branches evolve normally, while only their endpoint logits are interpolated before feedback.
- Empirical dynamics: Across both model families, L_t remains predominantly above the critical value L = 2, placing pretrained rollouts in the majority-amplifying regime.The measured field b_t is much smaller than the coupling.
- Empirical dynamics: Above the threshold, mixtures initialized just above and below 1/2 rapidly diverge toward opposite branches.Reducing the feedback gain reverses this behavior: trajectories merge below the threshold.
- Threshold behavior: The transition occurs around the theoretical threshold L_eff = 2: large effective coupling separates trajectories, whereas small coupling merges them and erases their initial distinction.The effective coupling is swept by changing feedback temperature without changing the model, items, or branch trajectories.
5 The Cost of Correcting K-Way Mixtures
For K-component mixtures, preserving one target state requires correcting context-dependent distortion in the model’s continuous dynamics. Even this weak local requirement can require context-dependent information whose dimensionality grows with the number of components.
- K-way generalization: The K-way analysis represents mixture weights p in the interior of the simplex and uses log-ratio coordinates to generalize the two-component analysis.The two-token case is treated as a special case of mixtures with K components.
- Context-dependent dynamics: One decoding step induces a context-dependent map a_t+1 = Φ(x_t, a_t), where both contextual representations and candidate identities may change during rollout.The binary linear-logit reduction recovers the earlier dynamical equation.
- Correction problem: Exact preservation requires leaving mixture proportions unchanged, but the paper finds that Φ(x_t, a_t) ≠ a_t and that the distortion depends on context.A fixed transformation of the mixture therefore cannot generally undo the distortion.
- Correction cost: Even preserving one interior target state under nearby contextual perturbations can require a number of context-dependent quantities proportional to K.This is a weak requirement, not full preservation of every mixture state.
- Correction cost: Under full-rank contextual variation, local output changes can span all K − 1 mixture directions, making faithful correction require substantial context-dependent information.The Jacobian rank need not be small because contextual changes can affect the final hidden state along many independent directions.
6 Discussion and Implications
The paper presents a unified account in which architecture, training, and recursive softmax feedback independently undermine continuous-mixture preservation. It argues that robust correction must track context and scale with mixture complexity, while proposed interventions address only particular failure sources.
- Discussion and Implications: The experiments and theory identify three independent failure sources: architectural distortion, training-amplified distortion, and softmax feedback that either amplifies or erases component differences.The rollout experiments reproduce the dynamical behavior predicted by the analysis.
- Discussion and Implications: A globally reweighted or softened mixture is unlikely to suffice because the required correction changes with context and can require at least K − 1 context-dependent degrees of freedom.The K-way result applies under full-rank context variation.
- Discussion and Implications: Continuity alone does not preserve continuous states; preservation depends on whether representation, readout, and closed-loop dynamics jointly maintain the intended semantics.This connects prior positive results for trained latent reasoning with empirical collapse in pretrained models.
- Future directions: Future interventions target different causes: mixed-state training addresses geometry, while bypassing vocabulary readout addresses recursive softmax feedback.Coconut exemplifies the latter route, but its gains remain task-dependent, so removing softmax feedback alone is not sufficient for robust continuous reasoning.
- Future directions: A general K-way solution likely needs a context-conditioned controller whose capacity grows with the number of components.The paper identifies learning and testing such neutral dynamics as a next step.
7 Limitations and Broader Impact
The work is primarily diagnostic, and its empirical scope has important boundaries. Controlled embedding mixtures may not capture every semantic continuous state, while the study does not identify which training factors cause distortion or evaluate downstream performance in a fully trained system.
- Controlled embedding mixtures may not capture every form of semantic continuous state in continuous-reasoning methods.
- The empirical analysis is primarily binary and covers two model families, while the K-way result is theoretical and its full-rank condition is not directly measured.
- Matched random controls isolate effects associated with learned weights but do not identify which aspects of training cause the distortion.
- The study examines mixture preservation rather than downstream performance in a fully trained continuous-reasoning system.
- The work is primarily diagnostic, and stronger latent reasoning could make intermediate computation less human-readable, increasing the importance of monitoring and interpretability methods.
A Benchmark Construction
The benchmark tests whether models preserve controlled mixtures of two token embeddings by measuring the relative probability of corresponding answer words. It uses 1000 binary items across seven semantic categories, with endpoint checks ensuring well-posed questions across model families.
- The benchmark contains 1000 binary mixture items spanning seven semantic categories.The categories are color, temperature, size, speed, hardness, weight, and brightness.
- The benchmark measures the model’s relative probability assigned to two answer words at the next-token position.
- At p = 1 and p = 0, the mixture recovers the two ordinary-token endpoints; intermediate p values test mixture preservation.
- Each item inserts a convex combination p e(first) + (1 −p) e(second) into one prompt slot rather than a real token.
- Items require single-token components and answers under both model families, and trained models must assign at least 0.5 probability to the corresponding endpoint answer.
- The chat template closes the reasoning block before the answer position so next-token probability targets the requested one-word answer.
B Proofs
The proof establishes that under uniformly supercritical coupling, the dynamics preserve sign and are bounded below by a constant-coupling system. That comparison system converges to a unique attracting positive fixed point.
- For u > 0, F_L(u) = tanh(Lu/2) increases with both u and L, is odd, and preserves the sign of the initial state.
- Replacing varying couplings with the smallest encountered coupling yields a constant-coupling sequence v_t that remains below u_t by monotonicity and induction.
- When L_min > 2, the constant-coupling map has exactly one positive fixed point u*(L_min), attracting every initial v_0 in (0, 1).
- Therefore, under uniformly supercritical coupling, lim inf_t u_t is at least u*(L_min).
Proof of Theorem 2 (subcritical coupling washes out the mixture)
The subcritical analysis shows that mixture information decays geometrically, while a persistent context-dependent field can sustain a bounded residual state. Thus, the field may prevent convergence to balance, but its influence remains controlled below the critical threshold.
- Iterating the contraction inequality bounds the state independently of the initial mixture.
- A persistent context-dependent field can sustain a nonzero residual state even when the initial mixture’s contribution decays geometrically.
- The residual state satisfies |u_T| ≤ b_max 2(1 − ρ) = b_max 2 − L_max.
- A persistent field can prevent convergence to the balanced state, but under subcritical coupling its influence remains bounded relative to the distance from the critical threshold.
Proof of Theorem 3 (context-dimension lower bound)
The proof establishes a lower bound on the dimensionality of context-dependent correction by differentiating the correction map at an anchor point and bounding the resulting Jacobian rank.
- The chain rule sums contributions because contextual representation x enters Φ through both of its arguments.
- The Jacobian factors through D_xr, whose rank is at most m because r takes values in R^m.
- If D_xΦ has full rank K −1, then the context dimension must satisfy m ≥K −1.