Model Collapse in Alignment

• Model collapse in alignment is the degeneration of learned representations during or after alignment procedures, leading to reduced diversity and impaired task performance.
• It is characterized by geometric misalignment, such as the breakdown of Neural Collapse, where class means and classifier weights lose their optimal configuration, undermining generalization.
• Mitigation strategies like geometric regularization and prototype freezing have been shown to improve fairness, robustness, and accuracy across various domains.

Model collapse in alignment represents a unified geometric and functional degeneration of learned representations during or after explicit alignment procedures in deep learning models. It manifests as a pathological reduction in representational diversity, robustness, or task competence—often triggered by over-regularization, misapplication of safety constraints, or reward over-optimization—across vision, language, and generative architectures. The phenomenon is best characterized by instances where the intended alignment—safety, fairness, or task policy—is achieved in a degenerate, overly narrow, or brittle sense, undermining the model’s generalization or responsiveness.

1. Geometric Foundations: Neural Collapse and Feature-Classifier Alignment

The canonical geometric paradigm underpinning model collapse is Neural Collapse (NC), first observed in the terminal phase of deep network training. NC encompasses four statistical-geometry properties:

• NC1 (Variability collapse): Within-class covariance of last-layer representations vanishes, i.e., $\Sigma_W \rightarrow 0$.
• NC2 (Simplex ETF): Class means, centered by the global mean, form a simplex equiangular tight frame (ETF): the vectors are equally separated at maximal angles.
• NC3 (Self-duality/alignment): Classifier weights align precisely with centered class means; $w_k \propto \mu_k - \mu_G$; the classifier and feature spaces become congruent.
• NC4 (Nearest-center classification): Decisions reduce to nearest-mean in Euclidean space: $$\arg\max_k w_k^\top \mu \equiv \arg\min_k \|\mu - \mu_k\|$$.

In tasks such as few-shot class-incremental learning (FSCIL), this geometric alignment is exploited for optimal feature-classifier mutual configuration, maximizing the Fisher Discriminant Ratio under simplex ETF geometry (Yang et al., 2023).

Model collapse arises when this alignment degrades under non-ideal conditions, such as domain shift or extreme data imbalance, manifesting as misalignment or over-concentration of representations—breaking the symmetry and optimal decision geometry.

2. Collapse Manifestations Across Domains and Modalities

Collapse is not restricted to image classifiers; variants arise in LLMs, diffusion policies, multimodal retrieval, and safety-aligned large models:

• Class-Incremental and Long-Tailed Learning: Freezing prototypes (ETF) in FSCIL prevents cross-session misalignment and catastrophic forgetting. In long-tailed data, minority collapse occurs: tail classes’ classifier weights fail to align with their means; error exponents suffer a quadratic penalty proportional to the misalignment angle (Wang et al., 25 Nov 2025, Yang et al., 2023).
• LLMs and Fairness: Refusal-based alignment data ("I’m sorry…” answers) can poison instruction-tuned models, collapsing them into refusal-only agents and degrading reasoning benchmarks (MMLU, BBH, HumanEval) by 4–33% (Bekbayev et al., 2023). For intrinsic fairness, collapsed alignment—where fairness-sensitive word embeddings and their means are tightly aligned (low standard deviation in cosine direction)—correlates with improved stereotype and occupation-bias metrics (Xu et al., 2024).
• Multimodal and Generative Models: RLHF-aligned diffusion models exhibit Preference Mode Collapse (PMC), where reward maximization along biased embedding directions leads to loss of output diversity (identity, style, layout). Directional Decoupling Alignment ($D^2$-Align) mitigates PMC by learning and counteracting these reward-alignment biases (Chen et al., 30 Dec 2025).
• Test-Time Adaptation: Under domain shifts, sample-wise feature-classifier alignment collapses (NC3$^+$), causing pseudo-label drift and unreliable adaptation. Hybrid geometric-probabilistic losses can restore optimal NC alignment even without labels (Chen et al., 11 Dec 2025).
• Semantic Collapse in Video Retrieval: Uniform positive/negative contrastive learning collapses all event-representations within a video or query, reducing within-video and inter-query diversity. Text Correlation Preservation Learning (TCPL) and Cross-Branch Video Alignment (CBVA) preserve the semantic geometry of foundation models, disentangling events and restoring retrieval capacity (Moon et al., 31 Oct 2025).

3. Mechanisms and Metrics of Collapse

Collapse is quantifiable using geometric, statistical, and performance-based metrics:

Collapse Type	Key Metric/Indicator	Reference
Neural Collapse	ETF conformity; alignment: $\cos(\hat{\mu}_c, \hat{w}_c)$	(Yang et al., 2023, Wang et al., 25 Nov 2025)
Diversity reduction	Topic/plurality counts; entropy of outputs/topics	(Hamilton, 2024, Chen et al., 30 Dec 2025)
Alignment poisoning	Drop in reasoning benchmarks; mutual information loss	(Bekbayev et al., 2023)
Fairness alignment	Std. of alignment on protected vocabulary; bias metrics (ICAT, GAP)	(Xu et al., 2024)
Safety collapse	Harmfulness score (HS), $\Delta$HS after fine-tuning	(Hsiung et al., 5 Jun 2025)
Spectral collapse	Collapse of sign diversity in spectral alignment (SA)	(Qiu et al., 5 Oct 2025)

Collapse often results from inexorable or overly-constrained “pulls” in parameter or representation space—whether via explicit regularization, backpropagated losses, or discrete alignment operations (RLHF, SFT, proxy rewards)—that flatten the landscape of functionally relevant distinctions.

4. Theoretical Consequences and Error Exponent Analysis

A fundamental result is that misalignment between feature means and classifier weights, even when both lie on simplex ETFs, reduces the optimal error exponent by a factor proportional to $\cos^2\alpha$, where $\alpha$ is the alignment angle (Wang et al., 25 Nov 2025). This is true even if class means and weights are maximally separated within their own spaces. Thus, recovery of geometric alignment is strictly necessary for optimal generalization.

Similarly, collapse of sign-diversity in spectral alignment metrics reliably predicts impending loss explosion and divergence in LLM training, with SA collapse preceding loss blow-ups by thousands of steps—outperforming traditional weight norm or gradient metrics (Qiu et al., 5 Oct 2025).

5. Algorithmic Guardrails and Mitigation Strategies

To prevent or reverse collapse in alignment:

• Geometric regularization: Plug-and-play alignment penalties (e.g., cosine similarity) between features and classifier weights (SpA-Reg), periodic SLERP rotation, or gradient projection can maintain alignment during training, supporting robust neural collapse in imbalanced regimes (Wang et al., 25 Nov 2025).
• Prototype freezing: ETF-based frozen classifiers ensure all class features can converge to predefined, maximally separated vertices across incremental learning sessions, eliminating misalignment and catastrophic forgetting (Yang et al., 2023).
• Dual-objective and hybrid losses: In domain adaptation, hybrid geometric/confidence targets for alignment losses address breakdowns of pseudo-label reliability under domain shift, preserving class separation even without supervision (Chen et al., 11 Dec 2025).
• Reward de-biasing: In diffusion RLHF, decoupling the reward direction prevents collapse to a high-reward but low-diversity manifold, balancing quality and diversity (Chen et al., 30 Dec 2025).
• Dataset curation: Removing refusal-alignment examples from instructional datasets and controlling representation similarity between alignment and downstream fine-tuning corpora can prevent reasoning or safety collapse in LLMs (Bekbayev et al., 2023, Hsiung et al., 5 Jun 2025).
• Diversity objectives: Explicit encouragement of topic entropy or writing-style conditional entropy, as well as cognitively grounded perturbation operators (macro/micro-layer diversifiers), ensure preservation of output variation and resistance to expressive collapse (Hamilton, 2024, Jiang, 1 Dec 2025).

6. Applications and Empirical Impact

Experiments consistently confirm that collapse-aware geometric and algorithmic strategies yield superior final accuracies, fairness, and robustness metrics:

• FSCIL: ETF+DR loss outperforms state-of-the-art by 2–4 points in last-session accuracy and shows dramatic reduction in forgetting (Yang et al., 2023).
• Long-tailed classification: Explicit alignment relieves tail-class collapse and yields up to 2.6% accuracy gains over strong NC-based baselines (Wang et al., 25 Nov 2025).
• Fairness: Neural Collapse regularization reduces Gender Association and TPR gaps on fairness benchmarks, raising ICAT scores while minimally affecting GLUE task accuracy (Xu et al., 2024).
• Safety: Downstream fine-tuning on data dissimilar to the original alignment corpus yields up to 10.33 pp lower harmfulness scores and more durable safety barriers against jailbreak attacks (Hsiung et al., 5 Jun 2025).
• Multimodal reasoning: Safety-CoT fine-tuning reduces attack success rates to near-zero across all evaluated jailbreak tests without harming reasoning ability (Lou et al., 10 May 2025).
• Diversity in generative models: Directional decoupling achieves simultaneously state-of-the-art human preference and generative diversity on DivGenBench (Chen et al., 30 Dec 2025).

7. Broader Implications and Future Challenges

Model collapse in alignment exposes universal vulnerabilities in empirical and algorithmic alignment strategies. The phenomenon bridges disparate domains—incremental vision, language reasoning, safety, fairness, and generation—indicating that geometric congruence and meaningful diversity must be preserved alongside explicit alignment objectives. Theoretical analysis (error exponents, information bottlenecks) and practical metrics (cosine similarity, entropy, diversity indices) jointly inform interventions.

Future research should pursue adaptive alignment schedules, multi-prototype and non-symmetric collapse analysis, and modalities beyond classification—e.g., regression and structured prediction with tight-frame regularization. Robust alignment must internalize not only static constraints but also guard against dynamic or interactive collapse, especially in continual, open-world, or adversarial contexts.

Model collapse in alignment thus encapsulates both a warning and a design principle: optimal, interpretable alignment is always geometric, never merely functional, and it is the collapse—subtly measured, precisely defined, stringently mitigated—that ultimately delimits the expressive, fair, and safe capacity of all modern machine learning systems (Yang et al., 2023, Wang et al., 25 Nov 2025, Xu et al., 2024, Bekbayev et al., 2023, Chen et al., 30 Dec 2025, Chen et al., 11 Dec 2025, Hsiung et al., 5 Jun 2025, Lou et al., 10 May 2025, Hamilton, 2024, Qiu et al., 5 Oct 2025, Moon et al., 31 Oct 2025).