TemporalDoRA: Temporal PEFT for Robust Surgical Video Question Answering

Surgical Video Question Answering (VideoQA) requires accurate temporal grounding while remaining robust to natural variation in how clinicians phrase questions, where linguistic bias can arise. Standard Parameter Efficient Fine Tuning (PEFT) methods adapt pretrained projections without explicitly modeling frame-to-frame interactions within the adaptation pathway, limiting their ability to exploit sparse temporal evidence. We introduce TemporalDoRA, a video-specific PEFT formulation that extends Weight-Decomposed Low-Rank Adaptation by (i) inserting lightweight temporal Multi-Head Attention (MHA) inside the low-rank bottleneck of the vision encoder and (ii) selectively applying weight decomposition only to the trainable low-rank branch rather than the full adapted weight. This design enables temporally-aware updates while preserving a frozen backbone and stable scaling. By mixing information across frames within the adaptation subspace, TemporalDoRA steers updates toward temporally consistent visual cues and improves robustness with minimal parameter overhead. To benchmark this setting, we present REAL-Colon-VQA, a colonoscopy VideoQA dataset with 6,424 clip--question pairs, including paired rephrased Out-of-Template questions to evaluate sensitivity to linguistic variation. TemporalDoRA improves Out-of-Template performance, and ablation studies confirm that temporal mixing inside the low-rank branch is the primary driver of these gains. We also validate on EndoVis18-VQA adapted to short clips and observe consistent improvements on the Out-of-Template split. Code and dataset available at~\href{https://anonymous.4open.science/r/TemporalDoRA-BFC8/}{Anonymous GitHub}.

SORT: A Systematically Optimized Ranking Transformer for Industrial-scale Recommenders

While Transformers have achieved remarkable success in LLMs through superior scalability, their application in industrial-scale ranking models remains nascent, hindered by the challenges of high feature sparsity and low label density. In this paper, we propose SORT (Systematically Optimized Ranking Transformer), a scalable model designed to bridge the gap between Transformers and industrial-scale ranking models. We address the high feature sparsity and low label density challenges through a series of optimizations, including request-centric sample organization, local attention, query pruning and generative pre-training. Furthermore, we introduce a suite of refinements to the tokenization, multi-head attention (MHA), and feed-forward network (FFN) modules, which collectively stabilize the training process and enlarge the model capacity. To maximize hardware efficiency, we optimize our training system to elevate the model FLOPs utilization (MFU) to 22%. Extensive experiments demonstrate that SORT outperforms strong baselines and exhibits excellent scalability across data size, model size and sequence length, while remaining flexible at integrating diverse features. Finally, online A/B testing in large-scale e-commerce scenarios confirms that SORT achieves significant gains in key business metrics, including orders (+6.35%), buyers (+5.97%) and GMV (+5.47%), while simultaneously halving latency (-44.67%) and doubling throughput (+121.33%).

Whisper-MLA: Reducing GPU Memory Consumption of ASR Models based on MHA2MLA Conversion

The Transformer-based Whisper model has achieved state-of-the-art performance in Automatic Speech Recognition (ASR). However, its Multi-Head Attention (MHA) mechanism results in significant GPU memory consumption due to the linearly growing Key-Value (KV) cache usage, which is problematic for many applications especially with long-form audio. To address this, we introduce Whisper-MLA, a novel architecture that incorporates Multi-Head Latent Attention (MLA) into the Whisper model. Specifically, we adapt MLA for Whisper's absolute positional embeddings and systematically investigate its application across encoder self-attention, decoder self-attention, and cross-attention modules. Empirical results indicate that applying MLA exclusively to decoder self-attention yields the desired balance between performance and memory efficiency. Our proposed approach allows conversion of a pretrained Whisper model to Whisper-MLA with minimal fine-tuning. Extensive experiments on the LibriSpeech benchmark validate the effectiveness of this conversion, demonstrating that Whisper-MLA reduces the KV cache size by up to 87.5% while maintaining competitive accuracy.

Interleaved Head Attention

Multi-Head Attention (MHA) is the core computational primitive underlying modern Large Language Models (LLMs). However, MHA suffers from a fundamental linear scaling limitation: $H$ attention heads produce exactly $H$ independent attention matrices, with no communication between heads during attention computation. This becomes problematic for multi-step reasoning, where correct answers depend on aggregating evidence from multiple parts of the context and composing latent token-to-token relations over a chain of intermediate inferences. To address this, we propose Interleaved Head Attention (IHA), which enables cross-head mixing by constructing $P$ pseudo-heads per head (typically $P=H$), where each pseudo query/key/value is a learned linear combination of all $H$ original queries, keys and values respectively. Interactions between pseudo-query and pseudo-key heads induce up to $P^2$ attention patterns per head with modest parameter overhead $\mathcal{O}(H^2P)$. We provide theory showing improved efficiency in terms of number of parameters on the synthetic Polynomial task (IHA uses $Θ(\sqrt{k}n^2)$ parameters vs. $Θ(kn^2)$ for MHA) and on the synthetic order-sensitive CPM-3 task (IHA uses $\lceil\sqrt{N_{\max}}\rceil$ heads vs. $N_{\max}$ for MHA). On real-world benchmarks, IHA improves Multi-Key retrieval on RULER by 10-20% (4k-16k) and, after fine-tuning for reasoning on OpenThoughts, improves GSM8K by 5.8% and MATH-500 by 2.8% (Majority Vote) over full attention.

A BERTology View of LLM Orchestrations: Token- and Layer-Selective Probes for Efficient Single-Pass Classification

Production LLM systems often rely on separate models for safety and other classification-heavy steps, increasing latency, VRAM footprint, and operational complexity. We instead reuse computation already paid for by the serving LLM: we train lightweight probes on its hidden states and predict labels in the same forward pass used for generation. We frame classification as representation selection over the full token-layer hidden-state tensor, rather than committing to a fixed token or fixed layer (e.g., first-token logits or final-layer pooling). To implement this, we introduce a two-stage aggregator that (i) summarizes tokens within each layer and (ii) aggregates across layer summaries to form a single representation for classification. We instantiate this template with direct pooling, a 100K-parameter scoring-attention gate, and a downcast multi-head self-attention (MHA) probe with up to 35M trainable parameters. Across safety and sentiment benchmarks our probes improve over logit-only reuse (e.g., MULI) and are competitive with substantially larger task-specific baselines, while preserving near-serving latency and avoiding the VRAM and latency costs of a separate guard-model pipeline.

Guardians of the Hair: Rescuing Soft Boundaries in Depth, Stereo, and Novel Views

Soft boundaries, like thin hairs, are commonly observed in natural and computer-generated imagery, but they remain challenging for 3D vision due to the ambiguous mixing of foreground and background cues. This paper introduces Guardians of the Hair (HairGuard), a framework designed to recover fine-grained soft boundary details in 3D vision tasks. Specifically, we first propose a novel data curation pipeline that leverages image matting datasets for training and design a depth fixer network to automatically identify soft boundary regions. With a gated residual module, the depth fixer refines depth precisely around soft boundaries while maintaining global depth quality, allowing plug-and-play integration with state-of-the-art depth models. For view synthesis, we perform depth-based forward warping to retain high-fidelity textures, followed by a generative scene painter that fills disoccluded regions and eliminates redundant background artifacts within soft boundaries. Finally, a color fuser adaptively combines warped and inpainted results to produce novel views with consistent geometry and fine-grained details. Extensive experiments demonstrate that HairGuard achieves state-of-the-art performance across monocular depth estimation, stereo image/video conversion, and novel view synthesis, with significant improvements in soft boundary regions.

Mixture of Attention Schemes (MoAS): Learning to Route Between MHA, GQA, and MQA

The choice of attention mechanism in Transformer models involves a critical trade-off between modeling quality and inference efficiency. Multi-Head Attention (MHA) offers the best quality but suffers from large Key-Value (KV) cache memory requirements during inference. Multi-Query Attention (MQA) and Grouped-Query Attention (GQA) reduce memory usage but often at the cost of model performance. In this work, we propose Mixture of Attention Schemes (MoAS), a novel architecture that dynamically selects the optimal attention scheme (MHA, GQA, or MQA) for each token via a learned router. We demonstrate that dynamic routing performs better than static averaging of schemes and achieves performance competitive with the MHA baseline while offering potential for conditional compute efficiency. Experimental results on WikiText-2 show that dynamic routing (val loss 2.3074) outperforms a static mixture (2.3093), validating the effectiveness of the proposed method. Our code is available at https://github.com/Esmail-ibraheem/Mixture-of-Attention-Schemes-MoAS.

BLASST: Dynamic BLocked Attention Sparsity via Softmax Thresholding

The growing demand for long-context inference capabilities in Large Language Models (LLMs) has intensified the computational and memory bottlenecks inherent to the standard attention mechanism. To address this challenge, we introduce BLASST, a drop-in sparse attention method that dynamically prunes the attention matrix without any pre-computation or proxy scores. Our method uses a fixed threshold and existing information from online softmax to identify negligible attention scores, skipping softmax computation, Value block loading, and the subsequent matrix multiplication. This fits seamlessly into existing FlashAttention kernel designs with negligible latency overhead. The approach is applicable to both prefill and decode stages across all attention variants (MHA, GQA, MQA, and MLA), providing a unified solution for accelerating long-context inference. We develop an automated calibration procedure that reveals a simple inverse relationship between optimal threshold and context length, enabling robust deployment across diverse scenarios. Maintaining high accuracy, we demonstrate a 1.62x speedup for prefill at 74.7% sparsity and a 1.48x speedup for decode at 73.2% sparsity on modern GPUs. Furthermore, we explore sparsity-aware training as a natural extension, showing that models can be trained to be inherently more robust to sparse attention patterns, pushing the accuracy-sparsity frontier even further.

Focusing on Language: Revealing and Exploiting Language Attention Heads in Multilingual Large Language Models

Large language models (LLMs) increasingly support multilingual understanding and generation. Meanwhile, efforts to interpret their internal mechanisms have emerged, offering insights to enhance multilingual performance. While multi-head self-attention (MHA) has proven critical in many areas, its role in multilingual capabilities remains underexplored. In this work, we study the contribution of MHA in supporting multilingual processing in LLMs. We propose Language Attention Head Importance Scores (LAHIS), an effective and efficient method that identifies attention head importance for multilingual capabilities via a single forward and backward pass through the LLM. Applying LAHIS to Aya-23-8B, Llama-3.2-3B, and Mistral-7B-v0.1, we reveal the existence of both language-specific and language-general heads. Language-specific heads enable cross-lingual attention transfer to guide the model toward target language contexts and mitigate off-target language generation issue, contributing to addressing challenges in multilingual LLMs. We also introduce a lightweight adaptation that learns a soft head mask to modulate attention outputs over language heads, requiring only 20 tunable parameters to improve XQuAD accuracy. Overall, our work enhances both the interpretability and multilingual capabilities of LLMs from the perspective of MHA.

The Strong Lottery Ticket Hypothesis for Multi-Head Attention Mechanisms

The strong lottery ticket hypothesis (SLTH) conjectures that high-performing subnetworks, called strong lottery tickets (SLTs), are hidden in randomly initialized neural networks. Although recent theoretical studies have established the SLTH across various neural architectures, the SLTH for transformer architectures still lacks theoretical understanding. In particular, the current theory of the SLTH does not yet account for the multi-head attention (MHA) mechanism, a core component of transformers. To address this gap, we introduce a theoretical analysis of the existence of SLTs within MHAs. We prove that, if a randomly initialized MHA of $H$ heads and input dimension $d$ has the hidden dimension $O(d\log(Hd^{3/2}))$ for the key and value, it contains an SLT that approximates an arbitrary MHA with the same input dimension with high probability. Furthermore, by leveraging this theory for MHAs, we extend the SLTH to transformers without normalization layers. We empirically validate our theoretical findings, demonstrating that the approximation error between the SLT within a source model (MHA and transformer) and an approximate target counterpart decreases exponentially by increasing the hidden dimension of the source model.