Block diffusion speculative decoding accelerates LLM inference by predicting all tokens within a block simultaneously for the target model to verify in parallel. Predicting an entire block at once requires a sufficiently capable draft model and effective utilization of the target model’s internal knowledge. However, the state-of-the-art method DFlash constrains all draft layers to share a single fused representation derived from only a few target layers, limiting per-layer expressiveness and hindering further scaling of draft capacity. In this paper, we present DFLare, which flares out the narrow conditioning bottleneck of DFlash through a lightweight layer-wise fusion mechanism: each draft layer attends to its own learnable combination of a broad set of target layers at negligible overhead, simultaneously injecting richer target knowledge and providing every draft layer with a distinct input. This enhanced per-layer expressiveness enables scaling the draft model to deeper architectures with consistent gains. We further scale training data from 800K to 2.4M samples to fully exploit the enlarged capacity. On six benchmarks spanning mathematical reasoning, code generation, and conversation, DFLare attains average wall-clock speedups of 5.52x on Qwen3-4B, 5.46x on Qwen3-8B, and 3.91x on GPT-OSS-20B, improving over DFlash by roughly 11%, 8%, and 5% respectively.

Dealing with language heterogeneity has always been one of the challenges in neural machine translation (NMT).The idea of using mixture-of-experts (MoE) naturally excels in addressing this issue by employing different experts to take responsibility for different problems.However, the parameter-inefficiency problem in MoE results in less performance improvement when boosting the number of parameters.Moreover, most of the MoE models are suffering from the training instability problem.This paper proposes MoA (Mixture-of-Adapters), a lightweight MoE-based NMT model that is trained via an elaborately designed stage-wise training strategy.With the standard Transformer as the backbone model, we introduce lightweight adapters as experts for easy expansion.To improve the parameter efficiency, we explicitly model and distill the language heterogeneity into the gating network with clustering.After freezing the gating network, we adopt the Gumbel-Max sampling as the routing scheme when training experts to balance the knowledge of generalization and specialization while preventing expert over-fitting.Empirical results show that MoA achieves stable improvements in different translation tasks by introducing much fewer extra parameters compared to other MoE baselines.Additionally, the performance evaluations on a multi-domain translation task illustrate the effectiveness of our training strategy.

The Visual Word Sense Disambiguation (VWSD) shared task aims at selecting the image among candidates that best interprets the semantics of a target word with a short-length phrase for English, Italian, and Farsi. The limited phrase context, which only contains 2-3 words, challenges the model’s understanding ability, and the visual label requires image-text matching performance across different modalities. In this paper, we propose a prompt based and multimodal retrieval enhanced VWSD system, which uses the rich potential knowledge of large-scale pretrained models by prompting and additional text-image information from knowledge bases and open datasets. Under the English situation and given an input phrase, (1) the context retrieval module predicts the correct definition from sense inventory by matching phrase and context through a biencoder architecture. (2) The image retrieval module retrieves the relevant images from an image dataset.(3) The matching module decides that either text or image is used to pair with image labels by a rule-based strategy, then ranks the candidate images according to the similarity score. Our system ranks first in the English track and second in the average of all languages (English, Italian, and Farsi).

This paper describes our system for SemEval-2023 Task 2 Multilingual Complex Named EntityRecognition (MultiCoNER II). Our teamSamsung Research China - Beijing proposesan AL-R (Adjustable Loss RoBERTa) model toboost the performance of recognizing short andcomplex entities with the challenges of longtaildata distribution, out of knowledge base andnoise scenarios. We first employ an adjustabledice loss optimization objective to overcomethe issue of long-tail data distribution, which isalso proved to be noise-robusted, especially incombatting the issue of fine-grained label confusing. Besides, we develop our own knowledgeenhancement tool to provide related contextsfor the short context setting and addressthe issue of out of knowledge base. Experimentshave verified the validation of our approaches.