CHERRY: Compressed Hierarchical Experts with Recurrent Representational Yield

• URL: http://arxiv.org/abs/2606.31796v1
• Date: Tue, 30 Jun 2026 15:14:38 GMT
• Title: CHERRY: Compressed Hierarchical Experts with Recurrent Representational Yield
• Abstract summary: We study three techniques for training compute-efficient language models.<n>SGT concentrates supervision on the 15% of output tokens that carry semantic payload.<n> depth compression with recurrent recovery.<n>We validate these techniques on CHERRY-1.8B, a Korean foundation model.
• Score: 1.2203679037551938
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: We study three complementary techniques for training compute-efficient language models. (1) Selective supervision and per-token efficiency. Selective Ground Truth Token Training (SGT) concentrates supervision on the ~15% of output tokens that carry semantic payload. Through positive gradient coupling in position-shared transformer weights -- a token-level instance of auxiliary-task transfer -- the remaining 85% of unsupervised tokens still improve substantially, giving a 4.5x per-supervised-token efficiency (at the step-100 eval optimum, ~67% of the full-sequence loss reduction is recovered from 15% of the supervision). We prove that this improvement on unsupervised tokens is guaranteed whenever the gradient coupling coefficient gamma-bar = 0.72 is positive (Theorem 1), and show the effect is a property of natural-language structure: it collapses on shuffled text. (2) Depth compression with recurrent recovery. A 48-layer, 1B-parameter transformer is compressed to 6 layers (227M) by averaging adjacent layers and restored through learned recurrent unrolling. With 34 effective recurrent layers it reaches a held-out loss of 2.934, within measurement noise of a 566M dense model at 2.926 -- a 2.5x reduction in parameters. (3) Fusion of compressed experts. Assembling several compressed models as a Mixture of Efficient Experts (MoEE) with multi-token prediction improves over each single expert at comparable active parameters: a 2-expert MoEE reaches loss 2.789 versus 2.926 for the best single compressed model. We validate these techniques on CHERRY-1.8B, a Korean foundation model whose every trainable parameter derives from our own training runs. We are explicit throughout about the scope of the evidence (one model family, Korean data, loss-based metrics) and about which claims are established versus prospective.

Related papers

• On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability [57.03393882933515]
  We describe the architecture and ablations of Qwen3.8-Flash-Next, a sparse mixture-of-experts model with 125B parameters, 6B activated per token, and 51B parameters of n-gram embedding tables held off the accelerator.<n>On fourteen pre-training benchmarks the model leads the 397B-A17B predecessor on eight and trails it on the rest by at most 2.6 points, at 1/3 the activated parameters, 1/3 the training tokens, and roughly 1/9 the training FLOPs.
  arXiv  Detail & Related papers  (2026-08-31T06:35:07Z)
• How Much Capacity Does EEG Denoising Need? Ultra-Compact Networks reveal Benchmark Saturation and Metric-Utility Gap [0.0]
  Deep learning EEG denoising architectures have scaled from tens of thousands to tens of millions of parameters, yet no prior study has isolated model capacity as the experimental variable.<n>We address both gaps by fixing architecture, loss, data split, and training recipe while sweeping only channel width from 1.05K to 40.26K parameters in a minimal depthwiseseparable convolutional U-Net.
  arXiv  Detail & Related papers  (2026-06-07T12:17:25Z)
• On the Role of Reasoning Patterns in the Generalization Discrepancy of Long Chain-of-Thought Supervised Fine-Tuning [63.41902113656453]
  Supervised Fine-Tuning (SFT) on long Chain-of-Thought (CoT) trajectories has become a pivotal phase in building large reasoning models.<n>We conduct a comparative study using two sources of verified CoT trajectories generated by two competing models.<n>We find that SFT on textttDeepSeek-R1-0528 data achieves remarkably lower training loss, yet exhibits significantly worse generalization performance.
  arXiv  Detail & Related papers  (2026-04-02T07:00:54Z)
• Low-Rank Adaptation Reduces Catastrophic Forgetting in Sequential Transformer Encoder Fine-Tuning: Controlled Empirical Evidence and Frozen-Backbone Representation Probes [3.305265383862785]
  We present an empirical study of Low-Rank Adaptation (LoRA) in sequential transformer encoder finetuning with companion representation probes.<n>In five full-validation BERT-base reruns on an RTE->MRPC->CoLA->SST-2 sequence, full fine-tuning yields 19.9% +/- average forgetting, whereas standard LoRA (r, query/value modules) yields 0.6% +/-1.4%.
  arXiv  Detail & Related papers  (2026-03-29T14:14:36Z)
• Separating Constraint Compliance from Semantic Accuracy: A Novel Benchmark for Evaluating Instruction-Following Under Compression [0.0]
  Large language models (LLMs) exhibit degraded performance under prompt compression.<n>We introduce the CompressionDecay Test (CDCT), a benchmark that measures constraint compliance (CC) and semantic accuracy (SA)<n>We observe a universal U-curve pattern in constraint compliance (97.2% prevalence), with violations peaking at medium compression.
  arXiv  Detail & Related papers  (2025-12-02T13:25:48Z)
• E$^3$-Pruner: Towards Efficient, Economical, and Effective Layer Pruning for Large Language Models [24.195465096877196]
  layer pruning is a hardware-friendly approach for model compression.<n> introduces two key innovations: a differentiable mask optimization method and an entropy-aware adaptive knowledge distillation strategy.<n> achieves 96% accuracy, a mere 0.8% drop from the original model.
  arXiv  Detail & Related papers  (2025-11-21T12:32:01Z)
• Unifying Mixture of Experts and Multi-Head Latent Attention for Efficient Language Models [1.7272658301768147]
  MoE-MLA-RoPE is a novel architecture combination that combines Mixture of Experts (MoE) with Multi-head Latent Attention (MLA) and Rotary Position Embeddings (RoPE) for efficient language modeling.<n>Our approach addresses the fundamental trade-off between model capacity and computational efficiency through three key innovations.
  arXiv  Detail & Related papers  (2025-08-02T08:33:30Z)
• Mixture Compressor for Mixture-of-Experts LLMs Gains More [71.0473038084673]
  We propose a training-free Mixture-Compressor for Mixture-of-Experts large language models (MoE-LLMs)<n>Our MC integrates static quantization and dynamic pruning to collaboratively achieve extreme compression for MoE-LLMs with less accuracy loss.<n>For instance, at 2.54 bits, MC compresses 76.6% of the model, with only a 3.8% average accuracy loss.
  arXiv  Detail & Related papers  (2024-10-08T18:09:38Z)
• Token Compensator: Altering Inference Cost of Vision Transformer without Re-Tuning [63.43972993473501]
  Token compression expedites the training and inference of Vision Transformers (ViTs) However, when applied to downstream tasks, compression degrees are mismatched between training and inference stages. We propose a model arithmetic framework to decouple the compression degrees between the two stages.
  arXiv  Detail & Related papers  (2024-08-13T10:36:43Z)
• Sub-token ViT Embedding via Stochastic Resonance Transformers [51.12001699637727]
  Vision Transformer (ViT) architectures represent images as collections of high-dimensional vectorized tokens, each corresponding to a rectangular non-overlapping patch. We propose a training-free method inspired by "stochastic resonance" The resulting "Stochastic Resonance Transformer" (SRT) retains the rich semantic information of the original representation, but grounds it on a finer-scale spatial domain, partly mitigating the coarse effect of spatial tokenization.
  arXiv  Detail & Related papers  (2023-10-06T01:53:27Z)
• (Certified!!) Adversarial Robustness for Free! [116.6052628829344]
  We certify 71% accuracy on ImageNet under adversarial perturbations constrained to be within a 2-norm of 0.5. We obtain these results using only pretrained diffusion models and image classifiers, without requiring any fine tuning or retraining of model parameters.
  arXiv  Detail & Related papers  (2022-06-21T17:27:27Z)
• ERNIE-SPARSE: Learning Hierarchical Efficient Transformer Through Regularized Self-Attention [48.697458429460184]
  Two factors, information bottleneck sensitivity and inconsistency between different attention topologies, could affect the performance of the Sparse Transformer. This paper proposes a well-designed model named ERNIE-Sparse. It consists of two distinctive parts: (i) Hierarchical Sparse Transformer (HST) to sequentially unify local and global information, and (ii) Self-Attention Regularization (SAR) to minimize the distance for transformers with different attention topologies.
  arXiv  Detail & Related papers  (2022-03-23T08:47:01Z)

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.