Long-context modeling is increasingly important for language understanding, reasoning, coding agents, and recommendation systems, but standard softmax attention becomes expensive as sequences grow. KSA introduces a middle ground between per-token KV caching and aggressively compressed alternatives: it inserts learnable summary tokens at regular chunk boundaries to retain distant information at a controllable semantic compression ratio. This design reduces the cache required for long sequences while preserving referential and interpretable access to earlier context, and it can be combined with head- and dimension-level compression methods such as GQA and MLA.

Modern Parameter-Efficient Fine-Tuning (PEFT) methods such as low-rank adaptation (LoRA) reduce the cost of customizing large language models (LLMs), yet still require a separate optimization run for every downstream dataset. We introduce **Drag-and-Drop LLMs (DnD)**, a prompt-conditioned parameter generator that eliminates per-task training by mapping a handful of unlabeled task prompts directly to LoRA weight updates. A lightweight text encoder distills each prompt batch into condition embeddings, which are then transformed by a cascaded hyper-convolutional decoder into the full set of LoRA matrices. Once trained in a diverse collection of prompt-checkpoint pairs, DnD produces task-specific parameters in seconds, yielding i) up to **12,000$\times$** lower overhead than full fine-tuning, ii) average gains up to **30\%** in performance over the strongest training LoRAs on unseen common-sense reasoning, math, coding, and multimodal benchmarks, and iii) robust cross-domain generalization despite never seeing the target data or labels. Our results demonstrate that prompt-conditioned parameter generation is a viable alternative to gradient-based adaptation for rapidly specializing LLMs. Our project is available at https://jerryliang24.github.io/DnD.

Diffusion Transformers (DiTs) deliver state-of-the-art image quality, yet their training remains notoriously slow. A recent remedy—representation alignment (REPA) that matches DiT hidden features to those of a non-generative teacher (e.g. DINO)—dramatically accelerates the early epochs but plateaus or even de grades performance later. We trace this failure to a capacity mismatch: once the generative student begins modelling the joint data distribution, the teacher's lower-dimensional embeddings and attention patterns become a straitjacket rather than a guide. We then introduce HASTE (Holistic Alignment with Stage-wise Termination for Efficient training), a two-phase schedule that keeps the help and drops the hindrance. Phase I applies a holistic alignment loss that simultaneously distills attention maps (relational priors) and feature projections (semantic anchors) from the teacher into mid-level layers of the DiT, yielding rapid convergence. Phase II then performs one-shot termination that deactivates the alignment loss, once a simple trigger such as a fixed iteration is hit, freeing the DiT to focus on denoising and exploit its generative capacity. HASTE speeds up training of diverse DiTs without architecture changes. On ImageNet 256×256, it reaches the vanilla SiT-XL/2 baseline FID in 50epochs and matches REPA's best FID in 500epochs, amounting to a 28× reduction in optimization steps. HASTE also improves text to-image DiTs on MS-COCO, demonstrating to be a simple yet principled recipe for efficient diffusion training across various tasks. Our code is available here.

LLM-as-a-Judge to iteratively improve the alignment performance without the need of human annotations for preference data. These methods commonly utilize the same LLM to act as both the policy model (which generates responses) and the reward model (which scores and ranks those responses). The ranked responses are then used as preference pairs to train the LLM via direct alignment technolo gies (e.g. DPO). However, it is noteworthy that throughout this process, there is no guarantee of accuracy in the rewarding and ranking, which is critical for en suring accurate rewards and high-quality preference data. Empirical results from relatively small LLMs (e.g., 7B parameters) also indicate that improvements from self-rewarding may diminish after several iterations in certain situations, which we hypothesize is due to accumulated bias in the reward system. This bias can lead to unreliable preference data for training the LLM. To address this issue, we first formulate and analyze the generalized iterative preference fine-tuning frame work for self-rewarding language model. We then introduce the regularization to this generalized framework to mitigate the overconfident preference labeling in the self-rewarding process. Based on this theoretical insight, we propose a Consistency Regularized sElf-rewarding lAnguage Model (CREAM) that lever ages the consistency of rewards across different iterations to regularize the self rewarding training, helping the model to learn from more reliable preference data. With this explicit regularization, our empirical results demonstrate the superior ity of CREAM in improving both reward consistency and alignment performance. The code is publicly available at https://github.com/Raibows/CREAM.