Escaping Mode Collapse in LLM Generation via Geometric Regulation

Abstract: Mode collapse is a persistent challenge in generative modeling and appears in autoregressive text generation as behaviors ranging from explicit looping to gradual loss of diversity and premature trajectory convergence. We take a dynamical-systems view and reinterpret mode collapse as reduced state-space accessibility caused by geometric collapse: during generation, the model's internal trajectory becomes confined to a low-dimensional region of its representation space. This implies mode collapse is not purely a token-level phenomenon and cannot be reliably solved by symbolic constraints or probability-only decoding heuristics. Guided by this perspective, we propose Reinforced Mode Regulation (RMR), a lightweight, online state-space intervention that regulates dominant self-reinforcing directions in the Transformer value cache (implemented as low-rank damping). Across multiple LLMs, RMR substantially reduces mode collapse and enables stable, high-quality generation at extremely low entropy rates (down to 0.8 nats/step), whereas standard decoding typically collapses near 2.0 nats/step.

• The paper reframes mode collapse as geometric collapse, showing that shrinking correlation dimension in next-token probability trajectories can precede looping, template repetition, and semantic stalling more reliably than entropy or Distinct-n.
• The paper introduces Reinforced Mode Regulation (RMR), which detects highly persistent directions in the Transformer value cache through a bounded-spectrum generalized eigenvalue problem and applies targeted low-rank damping during decoding.
• The paper reports that RMR increases non-collapse rates from 8% to 56% at temperature 0.7 and from about 5% to 33% at an entropy target of 1.0 nats per step, while preserving measured text quality in tested settings.

Overview

This paper reframes mode collapse in autoregressive text generation—ranging from explicit token looping to subtler degeneration such as template repetition and semantic stalling—as a geometric phenomenon rather than a purely symbolic one. The authors argue that during decoding, the model's internal trajectory becomes confined to a low-dimensional region of its representation space, a condition they call geometric collapse, which reduces state-space accessibility. Building on this diagnosis, they propose Reinforced Mode Regulation (RMR), an inference-time intervention that identifies temporally persistent directions in the Transformer value cache via a bounded-spectrum generalized eigenvalue problem and applies low-rank damping to those directions. The central empirical claim is strong: RMR raises non-collapse rates from 8% to 56% at temperature 0.7, and from 5% to 33% at an entropy target of 1.0 nats/step, where standard decoding nearly always collapses.

A dynamical-systems account of collapse

The paper's conceptual foundation is a minimal stochastic dynamical model: a state-dependent iterated function system (IFS) with $2m$ contraction maps split into two groups, whose map-selection distribution depends on a history-averaged bias term $m_t$. This construction reproduces a temperature-induced phase transition analogous to the mean-field Ising model: above a critical inverse temperature the trajectory admits a unique ergodic invariant measure filling the ambient space; below it, positive feedback traps the trajectory in one of two low-dimensional attractors. Crucially, applying weak damping $m_t \leftarrow (1-\eta)m_t$ with $\eta = 10^{-4}$ restores full state-space accessibility even at low temperature. The implication is that suppressing slow, self-reinforcing variables—a principle grounded in Mori–Zwanzig projection theory, geometric singular perturbation theory, and synergetics—can prevent accessibility collapse without raising randomness. The authors acknowledge that LLMs do not satisfy strict contraction assumptions; the IFS is illustrative of the regime near looping, not a literal model of decoding.

Correlation dimension as a diagnostic

To measure accessibility empirically, the paper uses the correlation dimension—the scaling exponent $d$ in $C_t(\varepsilon) \propto \varepsilon^d$, estimated by the Grassberger–Procaccia method over the sequence of next-token log-probability vectors (projected to 10,000 dimensions via vocabulary binning). A finite-time variant computed online in $O(t)$ per step enables trajectory-level monitoring. The state choice follows the authors' earlier work on correlation dimension of autoregressive LLMs [(2605.00435) builds on dt2025corrdim], though they note the scaling range must be widened relative to that prior setting for stable temporal monitoring—an assumption-dependent detail worth noting.

The diagnostic results are consistent across temperature-locked and entropy-locked decoding: as randomness decreases, explicit looping rates rise while mean correlation dimension falls monotonically. Two findings carry particular weight against symbolic explanations:

• Under entropy-locked decoding, collapse still occurs once the entropy constraint falls below a critical value, showing that collapse cannot be reduced to entropy decay alone.
• Conversely, at high constant entropy (e.g., 2.0), where exact looping is rare, geometric collapse still occurs in implicit forms—template repetition or semantic looping without progression. Token-level proxies behave differently here: next-token entropy fluctuates well before looping onset, and Distinct-$n$ drops only after repetition becomes visible, whereas correlation dimension remains stable during healthy generation and drops sharply at collapse onset.

This positions correlation dimension as a leading indicator of both hard and soft degeneration, distinct from entropy or Distinct-$n$.

Reinforced Mode Regulation

RMR treats the value cache $\mathbf{V}_t \in \mathbb{R}^{t \times D}$ as the generation state (keys are deliberately left untouched, since modifying them would perturb attention weights). The method solves the generalized eigenvalue problem $m_t$0, where $m_t$1 is the temporal covariance of value vectors and $m_t$2 their symmetrized one-step cross-covariance. Under stationarity the spectrum is bounded ($m_t$3), so eigenvalues are interpretable as directional autocorrelation coefficients—a property the authors exploit for principled thresholding: only directions with $m_t$4 are damped, with strength $m_t$5, implemented as $m_t$6 with exponentially weaker damping on older timesteps ($m_t$7). Statistics are maintained online via exponentially weighted moving averages, and top eigenvectors are computed by block power iteration at cost $m_t$8 with $m_t$9. Damping applied once every 10 steps suffices.

The spectral analysis supports the mechanism: without RMR, near collapse onset ($m_t \leftarrow (1-\eta)m_t$0) the leading eigenvalue rises toward 1 while the second drops toward 0—a widening spectral gap indicating dominance by a single persistent mode. With RMR, the leading eigenvalue is held below threshold across layers. Notably, even in non-collapse runs the dominant eigenvalue stays near 1 in most layers, which motivates the 0.8 threshold as targeting only the most persistent direction(s) while preserving moderately persistent components such as topic and style. Because the spectrum is bounded, thresholds transfer meaningfully across models, layers, and timesteps; the authors also note the formulation extends naturally to architectures like Mamba.

Empirical results

Non-collapse is defined as maintaining correlation dimension above 8 over 1,000 generated tokens (Heidegger prompt from SEP, Qwen3-4B-Base unless stated otherwise, with top-$m_t \leftarrow (1-\eta)m_t$1, top-$m_t \leftarrow (1-\eta)m_t$2 filtering). Key comparisons:

Setting	Standard	Typical sampling	Random regulation	RMR
Temp 0.7 (non-collapse rate)	8%	—	—	56%
Entropy 1.0 (non-collapse rate)	~5%	—	—	33%

Two controls strengthen the causal claim. Typical sampling fails to prevent collapse in low-randomness regimes and can underperform standard decoding, consistent with its tendency to reduce effective sampling uncertainty. Random regulation of a same-rank, same-strength random subspace also falls short—indicating the benefit derives from targeting the specific persistent subspace, not generic cache perturbation.

On text quality, the paper evaluates only non-collapse completions at temperatures where standard decoding is stable (1.0–1.2), using LLM-as-a-Judge scores on coherence, syntax, and information progression. Score distributions between standard decoding and RMR are statistically indistinguishable (Kolmogorov–Smirnov $m_t \leftarrow (1-\eta)m_t$3-values well above 0.05 across models and criteria), including for 4,000-token generations where RMR achieves a 99% non-collapse rate versus 55–6% for standard decoding at comparable temperatures. Supplementary experiments on 60 SEP texts (96 completions each, 4,000 tokens) and on Qwen3-4B-Instruct and Llama3.1-8B-Instruct confirm that RMR maintains higher correlation dimension over long horizons across models, with instruction-tuned models showing somewhat greater intrinsic resistance to collapse.

Limitations and open questions

The paper is candid about scope. RMR requires white-box access to the KV cache, precluding black-box API settings and requiring inference-stack integration. It adds overhead for running statistics and eigenvector estimation, relevant under strict latency constraints. The fixed threshold $m_t \leftarrow (1-\eta)m_t$4 and strength $m_t \leftarrow (1-\eta)m_t$5 work well empirically but may be suboptimal for other architectures, layer/head selections, or decoding regimes—the threshold's justification rests partly on observed spectral separation in Qwen3-4B. Non-autoregressive models (e.g., diffusion LLMs) are not considered. Evaluation is confined to mode collapse as characterized by correlation dimension; interactions with safety behaviors, downstream task performance, and other failure modes remain unexamined. An additional unstated dependency: the quality evaluation excludes collapsing runs entirely, so RMR's effect on the content of borderline trajectories is not directly assessed.

Conclusion

This work connects symbolic degeneration in LLM decoding to measurable geometric collapse of the internal trajectory, using finite-time correlation dimension as a trajectory-level diagnostic that precedes and predicts surface-level failure better than entropy or Distinct-$m_t \leftarrow (1-\eta)m_t$6. Its intervention, RMR, operationalizes the classical slow-variable regulation principle within the value cache through a bounded-spectrum generalized eigenvalue formulation, yielding large gains in non-collapse rates at low randomness without detectable degradation of text quality. The main open questions are whether the spectral-threshold calibration generalizes across architectures and scales, whether the approach extends to black-box or non-autoregressive generation, and how geometric regulation interacts with task performance beyond free-form completion.