Logits Replay + MoClip: Stabilized, Low-Cost Post-Training with Minimal Forgetting

• URL: http://arxiv.org/abs/2510.09152v1
• Date: Fri, 10 Oct 2025 08:55:32 GMT
• Title: Logits Replay + MoClip: Stabilized, Low-Cost Post-Training with Minimal Forgetting
• Authors: Suming Qiu, Jing Li, Zhicheng Zhou, Junjie Huang, Linyuan Qiu, Zhijie Sun,
• Abstract summary: We introduce Logits Replay + MoClip, a framework that compresses supervision in the logit space and stabilizes optimization at the update level.<n> Empirically, our method improves domain performance on Communication Technology tasks while mitigating forgetting on general benchmarks.
• Score: 6.653834890554154
• License: http://creativecommons.org/licenses/by-nc-nd/4.0/
• Abstract: Large language models (LLMs) often face a trade-off in post-training: improvements on specialized domains frequently come at the expense of general capabilities. Existing solutions attempt to mitigate this tension via regularization, selective parameter updates, or data-centric replay, but each imposes significant costs in computation, data access, or adaptability. Recent work has shown that training signals can be compressed to subsets of logits without severe accuracy loss, suggesting a path toward efficient adaptation. However, naive truncation destabilizes optimization and exacerbates forgetting. We introduce Logits Replay + MoClip, a two-stage framework that compresses supervision in the logit space and stabilizes optimization at the update level. In Stage 0, we record dynamic Top-K token subsets that cover a probability threshold, always including the gold label. In Stage 1, we replay these compact subsets to compute exact renormalized losses, avoiding full softmax computation and implicitly regularizing. To ensure stability, we design MoClip, an optimizer that caps gradient-momentum rotation and applies an arctan2-based rescaling of updates. Empirically, our method improves domain performance on Communication Technology (CT) and NL2SQL tasks while mitigating forgetting on general benchmarks (MMLU, BBH, GPQA, MATH), and reduces training cost by over 40%. Together, these contributions offer a scalable, architecture-agnostic path for domain adaptation of LLMs without sacrificing generalization.

Related papers

• Training Large Reasoning Models Efficiently via Progressive Thought Encoding [63.254758972725654]
  Large reasoning models (LRMs) excel on complex problems but face a critical barrier to efficiency.<n>We introduce Progressive Thought, a parameter-efficient fine-tuning method that enables LRMs to reason effectively under fixed-size caches.
  arXiv  Detail & Related papers  (2026-02-18T20:03:38Z)
• RaBiT: Residual-Aware Binarization Training for Accurate and Efficient LLMs [5.782015253162346]
  Residual binarization enables matmul-free inference by stacking binary layers.<n>We propose RaBiT, a novel quantization framework that resolves coadaptation by algorithmically enforcing a residual hierarchy.<n>RaBiT achieves state-of-the-art performance, rivals even hardware-intensive Vector Quantization (VQ) methods, and delivers a $4.49times$ inference speed-up over full-precision models.
  arXiv  Detail & Related papers  (2026-02-05T06:41:11Z)
• Zero Sum SVD: Balancing Loss Sensitivity for Low Rank LLM Compression [11.908793753919745]
  We propose textbfZero Sum SVD (textbfZS-SVD), a post-training method that performs singular component selection in whitened coordinates.<n>textbfZS-SVD prunes components across the whole model with a textbfzero sum rule that keeps the cumulative predicted loss change near zero.<n>Experiments show consistent gains across diverse benchmarks and compression ratios.
  arXiv  Detail & Related papers  (2026-02-02T21:51:01Z)
• Gated KalmaNet: A Fading Memory Layer Through Test-Time Ridge Regression [53.48692193399171]
  Gated KalmaNet (GKA) is a layer that reduces the gap by accounting for the full past when predicting the next token.<n>We solve an online ridge regression problem at test time, with constant memory and linear compute cost in the sequence length.<n>On long-context, GKA excels at real-world RAG and LongQA tasks up to 128k tokens, achieving more than $10$% relative improvement over other fading memory baselines.
  arXiv  Detail & Related papers  (2025-11-26T03:26:37Z)
• PT$^2$-LLM: Post-Training Ternarization for Large Language Models [52.4629647715623]
  Large Language Models (LLMs) have shown impressive capabilities across diverse tasks, but their large memory and compute demands hinder deployment.<n>We propose PT$2$-LLM, a post-training ternarization framework tailored for LLMs.<n>At its core is an Asymmetric Ternary Quantizer equipped with a two-stage refinement pipeline.
  arXiv  Detail & Related papers  (2025-09-27T03:01:48Z)
• Constrained Entropic Unlearning: A Primal-Dual Framework for Large Language Models [7.566515311806724]
  Large Language Models (LLMs) deployed in real-world settings increasingly face the need to unlearn sensitive, outdated, or proprietary information.<n>Existing unlearning methods formulate forgetting and retention as a regularized trade-off, combining both objectives into a single scalarized loss.<n>We propose a new formulation of LLM unlearning as a constrained optimization problem: forgetting is enforced via a novel logit-margin flattening loss.
  arXiv  Detail & Related papers  (2025-06-05T17:55:23Z)
• Forget Forgetting: Continual Learning in a World of Abundant Memory [55.64184779530581]
  Continual learning has traditionally focused on minimizing exemplar memory.<n>This paper challenges this paradigm by investigating a more realistic regime.<n>We find that the core challenge shifts from stability to plasticity, as models become biased toward prior tasks and struggle to learn new ones.
  arXiv  Detail & Related papers  (2025-02-11T05:40:52Z)
• HALO: Hadamard-Assisted Lower-Precision Optimization for LLMs [45.37278584462772]
  We present HALO, a novel quantization-aware training approach for Transformers.<n>Our approach ensures that all large matrix multiplications during the forward and backward passes are executed in lower precision.<n>Applying to LLAMA-family models, HALO achieves near-full-precision-equivalent results during fine-tuning on various tasks.
  arXiv  Detail & Related papers  (2025-01-05T18:41:54Z)
• Initialization using Update Approximation is a Silver Bullet for Extremely Efficient Low-Rank Fine-Tuning [20.74016978787039]
  We propose a method that approximates full fine-tuning within low-rank subspaces.<n>We leverage the architecture of LoRA-XS to achieve optimal scaling for high-rank gradient updates.<n>Our findings establish that it is possible to simulate full fine-tuning in low-rank subspaces.
  arXiv  Detail & Related papers  (2024-11-29T09:10:30Z)
• Winner-Take-All Column Row Sampling for Memory Efficient Adaptation of Language Model [89.8764435351222]
  We propose a new family of unbiased estimators called WTA-CRS, for matrix production with reduced variance. Our work provides both theoretical and experimental evidence that, in the context of tuning transformers, our proposed estimators exhibit lower variance compared to existing ones.
  arXiv  Detail & Related papers  (2023-05-24T15:52:08Z)

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.