Compressed context is usually carried as human-readable text or as rendered images that must be decoded, even when its consumer is a language model. We introduce LatentPress, which writes conversational histories and long documents into a third representation: continuous memory tokens that a frozen decoder reads directly through its input-embedding interface, with no text reconstruction at inference. A small reader-matched writer compresses $4$-$16\times$ while training only an adapter (4.2M-26.2M parameters, $\sim\!0.1\%$ of the decoder). On LongMemEval, LatentPress reaches $0.504$ accuracy at $7.70\times$ compression versus $0.490$ for uncompressed evidence, outperforming text summaries (0.184) and OCR-based compression (0.426 to 0.312). On LongBench-QA, in-domain writers match or exceed raw-context reading at $4$-$8\times$ compression, while $16\times$ trails raw. Writing takes 43ms per conversation, roughly an order of magnitude faster than text summarization or OCR reconstruction, and reading is $5$-$9\times$ faster than raw context or cached OCR. We validate the interface under two transfer settings, zero-shot from UltraChat to LongMemEval memory QA and from LongMemEval-derived QA to unseen LongBench document domains, establishing direct soft tokens as a practical machine-facing context interface beyond text and vision. The implementation of the experiments could be found at: https://github.com/xuyd16ai/context_softtoken_compress .
Mohammad Aref Jafari-Raddani, Morteza Mohajjel Kafshdoozcs.LG cs.AI
While Parameter-Efficient Fine-Tuning (PEFT) has substantially reduced the hardware cost of adapting Large Language Models (LLMs) by decreasing the number of trainable parameters, recent studies have sought to further improve PEFT through parameter sharing. However, these approaches either employ uniform parameter sharing across layers, which can delay convergence, or rely on dynamic masking strategies, which add computational overhead. The potential of sharing patterns inspired by the inherent hierarchical structure of Transformer architectures remains unexplored in PEFT. To address this gap, we introduce SAPE (Sandwich Adapters for Parameter Efficiency), a PEFT framework based on a sandwich-style hard weight-sharing topology. SAPE routes intermediate Transformer layers through balanced shared group adapters while strictly isolating the input embedding and final projection boundary transformations to prevent gradient interference. This design significantly reduces memory consumption while eliminating the computational overhead associated with dynamic parameter-sharing methods. Extensive evaluations across encoder-only and causal decoder architectures demonstrate that SAPE achieves state-of-the-art performance in low-parameter regimes. On natural language understanding, SAPE outperforms proPETL on RoBERTa-large while utilizing only 10% of the baseline's parameter budget. On natural language generation and world knowledge reasoning with LLaMA-3.2 (3B) under a strict ~0.6M parameter constraint, SAPE outperforms AdaLoRA, yielding absolute improvements of +4.85% on GSM8K and +3.11% on CommonsenseQA. Furthermore, through comprehensive topological ablations, we formalize an inherent capacity trade-off: while hard parameter sharing strongly regularizes semantic generalization, it slightly smooths the sharp layer-wise transformations required for rigid multi-step arithmetic reasoning.
Multi-task supervised fine-tuning (SFT) often casts a heterogeneous data mixture as a single optimization problem, even though different tasks may reach their best generalization at different times. msft exposes this mismatch through task-wise roll-out, exclusion, and rollback, but its original formulation materializes the scheduler state as full-model checkpoints, making stage transitions costly to store, restore, and deploy. This paper introduces AuroSFT, a parameter-efficient framework that recasts the carried state of overfitting-aware multi-task SFT as a compact, mergeable adapter state. AuroSFT freezes the pretrained backbone, trains only injected adapters, rolls back adapter checkpoints at task-wise peaks, and continues on the remaining active mixture. At the layer level, each adapter applies an AuroRA-inspired adaptive nonlinear layer to a low-rank weight factor rather than to the sample representation. The resulting update remains linear in the input, rank-bounded, and exactly mergeable into the frozen projection. Under the retained-backbone comparison protocol, AuroSFT achieves 61.36% average accuracy, compared with 59.85% for the corresponding msft reference row, and obtains higher accuracy on all five backbones. Our code is available at the anonymous repository: https://anonymous.4open.science/r/AuroSFT-80D1.
High-dimensional embeddings from large language models impose significant storage and computational costs on vector search systems. Recent embedding compression methods, including Matryoshka-Adaptor (EMNLP 2024), Search-Adaptor (ACL 2024), and SMEC (EMNLP 2025), enable dimensionality reduction through lightweight residual adapters, but their training objectives cause severe overfitting when labeled data is scarce, degrading retrieval performance below the frozen baseline. We propose \textsc{DIVE} (\textbf{D}imensionality reduction with \textbf{I}mplicit \textbf{V}iew \textbf{E}nsembles), a compression adapter that addresses this failure through two mechanisms. First, a self-limiting hinge-based triplet loss produces zero gradient once a triplet satisfies the margin constraint, bounding the total perturbation applied to the pretrained embedding space. Second, a head-wise NT-Xent contrastive loss treats multiple learned projections of each embedding as implicit views, providing dense self-supervised gradients that compensate for the sparsity of the triplet signal on small datasets. Across six BEIR datasets, \textsc{DIVE} outperforms all three baseline adapters on every dataset and at every evaluated compression ratio, with a 14M-parameter open-source implementation.