Representation Collapse in Sequential Post-Training of Large Language Models

• URL: http://arxiv.org/abs/2605.30524v1
• Date: Thu, 28 May 2026 19:59:17 GMT
• Title: Representation Collapse in Sequential Post-Training of Large Language Models
• Abstract summary: We study whether sequential post-training gradually compresses internal representations into low-rank, anisotropic, and homogeneous feature spaces.<n>We analyze supervised fine-tuning, preference optimization, safety/refusal tuning, math and code specialization, and long chain-of-thought tuning under controlled stage orderings.
• Score: 17.437127662981847
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language models are now adapted through chains of post-training stages rather than through a single instruction-tuning pass. This paper studies whether such sequential post-training gradually compresses internal representations into low-rank, anisotropic, and homogeneous feature spaces. We define a measurement suite for hidden states, logits, token trajectories, and LoRA updates, and we use it to analyze supervised fine-tuning, preference optimization, safety/refusal tuning, math and code specialization, and long chain-of-thought tuning under controlled stage orderings. The central hypothesis is that excessive representation concentration is not merely a geometric curiosity: it predicts reduced plasticity during later adaptation, weaker out-of-domain generalization, and poorer calibration. We further evaluate lightweight interventions, including mixed-domain replay, feature refresh, representation diversity regularization, and LoRA update decorrelation, as ways to preserve future learnability without giving up the behavioral gains of post-training.

Related papers

• Measuring Structured Predictability in Neural Training Dynamics: A Cross-Regime Study [2.8019744694531554]
  Modern deep networks are trained through long update trajectories, yet their temporal organization remains less systematically characterized than architectures, losses, or redistributions.<n>We study short-horizon predictability as a measure of temporal redundancy, where, when, and under which training conditions recent updates contain information about near-future parameter motion.
  arXiv  Detail & Related papers  (2026-08-16T02:17:51Z)
• Dive Into the Implicit Biases of Low-rank Vision-language Alignment [22.80937850155209]
  We find that low-rank alignment not only reduces computational costs but also outperforms full- parameter alignment on most benchmarks.<n> Empirically, we find that low-rank alignment shifts model behavior from hallucinatory to conservative and preserves per-language linear separability of visual features.<n>We establish two theorems showing that low-rank alignment induces preferences for parameter subspaces with flat gradients and feature subspaces robust to perturbations.
  arXiv  Detail & Related papers  (2026-07-09T07:50:49Z)
• MIITA: Memory-Induced Inference-Time Adaptation for Continual Learning with Small Language Models [64.2820121997882]
  MIITA is a Memory-Induced Inference-Time Adaptation framework for supervised CL under constrained storage.<n>It stores supervised experiences as compact correction-direction prototypes with semantic anchors, and retrieves them at inference time using semantic and uncertainty-based cues.<n>Experiments show that MIITA consistently improves final performance and mitigates forgetting under fixed memory budgets.
  arXiv  Detail & Related papers  (2026-05-20T03:03:04Z)
• Continual Segmentation under Joint Nonstationarity [8.522964568257455]
  Evolving data streams induce joint nonstationarity in continual semantic segmentation.<n>This setting reflects practical structured prediction systems, yet remains largely unexplored in prior continual learning work.<n>We formalize continual segmentation under coupled class, domain, and label shifts and investigate learning in heterogeneous dense prediction environments.
  arXiv  Detail & Related papers  (2026-05-19T22:20:00Z)
• Next-Acceleration-Scale Prediction for Autoregressive MRI Reconstruction [52.32112533846212]
  MRI reconstruction is an inherently ill-posed inverse problem, since incomplete measurements admit many plausible solutions.<n>We address this limitation by moving reconstruction to discrete multi-scale latent space and posing it as autoregressive next-acceleration-scale prediction.<n>We show that our approach delivers improved reconstruction performance across diverse sampling patterns under extreme undersampling.
  arXiv  Detail & Related papers  (2026-05-19T04:40:50Z)
• Power Reinforcement Post-Training of Text-to-Image Models with Super-Linear Advantage Shaping [66.25536973294726]
  Post-training methods for text-to-image (T2I) models are prone to reward hacking.<n>Super-Linear Advantage Shaping (SLAS) reshapes the local policy space.<n>SLAS consistently surpasses the DanceGRPO baseline across multiple backbones and benchmarks.
  arXiv  Detail & Related papers  (2026-05-11T17:59:25Z)
• Deriving Hyperparameter Scaling Laws via Modern Optimization Theory [55.63126290312615]
  We study modern first-order boundss for methods based on the Linear Minimization Oracle (LMO)<n>Treating bounds in recent literature as a proxy and minimizing them across different tuning regimes yields closed-form power-law schedules for learning rate, momentum, and batch size.<n>Our results draw particular attention to the interaction between momentum and batch-size scaling, suggesting that optimal performance may be achieved with several scaling strategies.
  arXiv  Detail & Related papers  (2026-03-16T22:21:27Z)
• Alternating Reinforcement Learning for Rubric-Based Reward Modeling in Non-Verifiable LLM Post-Training [29.56905427210088]
  gradient-ARM is a framework that jointly optimize a rubric generator and a judge using reinforcement learning from preference feedback.<n>We show that gradient-ARM achieves state-of-the-art performance among baselines on benchmarks and significantly improves downstream policy alignment in both offline and online reinforcement learning settings.
  arXiv  Detail & Related papers  (2026-02-02T00:50:53Z)
• Enhancing Robustness of Vision-Language Models through Orthogonality Learning and Self-Regularization [77.62516752323207]
  We introduce an orthogonal fine-tuning method for efficiently fine-tuning pretrained weights and enabling enhanced robustness and generalization. A self-regularization strategy is further exploited to maintain the stability in terms of zero-shot generalization of VLMs, dubbed OrthSR. For the first time, we revisit the CLIP and CoOp with our method to effectively improve the model on few-shot image classficiation scenario.
  arXiv  Detail & Related papers  (2024-07-11T10:35:53Z)
• Trajectory Forecasting through Low-Rank Adaptation of Discrete Latent Codes [36.12653178844828]
  Trajectory forecasting is crucial for video surveillance analytics, as it enables the anticipation of future movements for a set of agents. We introduce Vector Quantized Variational Autoencoders (VQ-VAEs), which utilize a discrete latent space to tackle the issue of posterior collapse. We show that such a two-fold framework, augmented with instance-level discretization, leads to accurate and diverse forecasts.
  arXiv  Detail & Related papers  (2024-05-31T10:13:17Z)
• DR-Tune: Improving Fine-tuning of Pretrained Visual Models by Distribution Regularization with Semantic Calibration [38.4461170690033]
  We propose a novel fine-tuning framework, namely distribution regularization with semantic calibration (DR-Tune) DR-Tune employs distribution regularization by enforcing the downstream task head to decrease its classification error on the pretrained feature distribution. To alleviate the interference by semantic drift, we develop the semantic calibration (SC) module.
  arXiv  Detail & Related papers  (2023-08-23T10:59:20Z)
• Debiased Fine-Tuning for Vision-language Models by Prompt Regularization [56.48290708901531]
  We present a new paradigm for fine-tuning large-scale vision pre-trained models on downstream task, dubbed Prompt Regularization (ProReg)<n>ProReg uses the prediction by prompting the pretrained model to regularize the fine-tuning.<n>We show the consistently strong performance of ProReg compared with conventional fine-tuning, zero-shot prompt, prompt tuning, and other state-of-the-art methods.
  arXiv  Detail & Related papers  (2023-01-29T11:53:55Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.