No Generation without Representation: Efficient Causal Protein Language Models Enable Zero-Shot Fitness Estimation

Protein language models (PLMs) face a fundamental divide: masked language models (MLMs) excel at fitness prediction while causal models enable generation, forcing practitioners to maintain separate architectures. We introduce \textbf{Proust}, a 309M-parameter causal PLM that bridges this gap through architectural innovations adapted from recent LLM research, including grouped-query attention with shared K/V projections, cross-layer value residuals, and depthwise causal convolutions. Trained on 33B tokens in 40 B200 GPU-hours, Proust achieves Spearman $ρ= 0.390$ on ProteinGym substitutions, competitive with MLMs requiring 50--200$\times$ the compute. On indels, Proust sets a new state-of-the-art, outperforming models up to 20$\times$ larger. On EVEREST viral fitness benchmarks, it approaches structure-aware methods using sequence alone. These powerful representations position Proust in a sweet spot as it also retains native generative capabilities that MLMs lack by design. Interpretability analysis reveals that per-position entropy variance predicts, to an extent, when retrieval augmentation helps and hurts. Such insights can grow in both quantity and quality at scale and inform capabilities such as test-time scaling. Code and weights are available at https://github.com/Furkan9015/proust-inference

Low-Rank Key Value Attention

Transformer pretraining is increasingly constrained by memory and compute requirements, with the key-value (KV) cache emerging as a dominant bottleneck during training and autoregressive decoding. We propose \textit{low-rank KV adaptation} (LRKV), a simple modification of multi-head attention that reduces KV cache memory by exploiting redundancy across attention heads while preserving full token-level resolution. Each layer uses a shared full-rank KV projection augmented with low-rank, head-specific residuals, yielding a continuous trade-off between complete sharing and fully independent attention. LRKV is a drop-in replacement for standard multi-head attention and directly subsumes query-sharing approaches such as multi-query and grouped-query attention, while remaining distinct from latent-compression methods such as multi-latent attention (MLA). Across large-scale pretraining experiments, LRKV consistently achieves faster loss reduction, lower validation perplexity, and stronger downstream task performance than standard attention, MQA/GQA, and MLA. At the 2.5B scale, LRKV outperforms standard attention while using roughly half the KV cache, and reaches equivalent model quality with up to \textbf{20-25\% less training compute} when measured in cumulative FLOPs. To explain these gains, we analyze attention head structure in operator space and show that LRKV preserves nearly all functional head diversity relative to standard attention, whereas more aggressive KV-sharing mechanisms rely on compensatory query specialization. Together, these results establish LRKV as a practical and effective attention mechanism for scaling Transformer pretraining under memory- and compute-constrained regimes.

Mugi: Value Level Parallelism For Efficient LLMs

Value level parallelism (VLP) has been proposed to improve the efficiency of large-batch, low-precision general matrix multiply (GEMM) between symmetric activations and weights. In transformer based large language models (LLMs), there exist more sophisticated operations beyond activation-weight GEMM. In this paper, we explore how VLP benefits LLMs. First, we generalize VLP for nonlinear approximations, outperforming existing nonlinear approximations in end-to-end LLM accuracy, performance, and efficiency. Our VLP approximation follows a value-centric approach, where important values are assigned with greater accuracy. Second, we optimize VLP for small-batch GEMMs with asymmetric inputs efficiently, which leverages timely LLM optimizations, including weight-only quantization, key-value (KV) cache quantization, and group query attention. Finally, we design a new VLP architecture, Mugi, to encapsulate the innovations above and support full LLM workloads, while providing better performance, efficiency and sustainability. Our experimental results show that Mugi can offer significant improvements on throughput and energy efficiency, up to $45\times$ and $668\times$ for nonlinear softmax operations, and $2.07\times$ and $3.11\times$ for LLMs, and also decrease operational carbon for LLM operation by $1.45\times$ and embodied carbon by $1.48\times$.

Pairing-free Group-level Knowledge Distillation for Robust Gastrointestinal Lesion Classification in White-Light Endoscopy

White-Light Imaging (WLI) is the standard for endoscopic cancer screening, but Narrow-Band Imaging (NBI) offers superior diagnostic details. A key challenge is transferring knowledge from NBI to enhance WLI-only models, yet existing methods are critically hampered by their reliance on paired NBI-WLI images of the same lesion, a costly and often impractical requirement that leaves vast amounts of clinical data untapped. In this paper, we break this paradigm by introducing PaGKD, a novel Pairing-free Group-level Knowledge Distillation framework that that enables effective cross-modal learning using unpaired WLI and NBI data. Instead of forcing alignment between individual, often semantically mismatched image instances, PaGKD operates at the group level to distill more complete and compatible knowledge across modalities. Central to PaGKD are two complementary modules: (1) Group-level Prototype Distillation (GKD-Pro) distills compact group representations by extracting modality-invariant semantic prototypes via shared lesion-aware queries; (2) Group-level Dense Distillation (GKD-Den) performs dense cross-modal alignment by guiding group-aware attention with activation-derived relation maps. Together, these modules enforce global semantic consistency and local structural coherence without requiring image-level correspondence. Extensive experiments on four clinical datasets demonstrate that PaGKD consistently and significantly outperforms state-of-the-art methods, achieving relative AUC improvements of 3.3%, 1.1%, 2.8%, and 3.2%, respectively, establishing a new direction for cross-modal learning from unpaired data.

Nonparametric Kernel Clustering with Bandit Feedback

Clustering with bandit feedback refers to the problem of partitioning a set of items, where the clustering algorithm can sequentially query the items to receive noisy observations. The problem is formally posed as the task of partitioning the arms of an N-armed stochastic bandit according to their underlying distributions, grouping two arms together if and only if they share the same distribution, using samples collected sequentially and adaptively. This setting has gained attention in recent years due to its applicability in recommendation systems and crowdsourcing. Existing works on clustering with bandit feedback rely on a strong assumption that the underlying distributions are sub-Gaussian. As a consequence, the existing methods mainly cover settings with linearly-separable clusters, which has little practical relevance. We introduce a framework of ``nonparametric clustering with bandit feedback'', where the underlying arm distributions are not constrained to any parametric, and hence, it is applicable for active clustering of real-world datasets. We adopt a kernel-based approach, which allows us to reformulate the nonparametric problem as the task of clustering the arms according to their kernel mean embeddings in a reproducing kernel Hilbert space (RKHS). Building on this formulation, we introduce the KABC algorithm with theoretical correctness guarantees and analyze its sampling budget. We introduce a notion of signal-to-noise ratio for this problem that depends on the maximum mean discrepancy (MMD) between the arm distributions and on their variance in the RKHS. Our algorithm is adaptive to this unknown quantity: it does not require it as an input yet achieves instance-dependent guarantees.

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.

Leveraging Parameter Space Symmetries for Reasoning Skill Transfer in LLMs

Task arithmetic is a powerful technique for transferring skills between Large Language Models (LLMs), but it often suffers from negative interference when models have diverged during training. We address this limitation by first aligning the models' parameter spaces, leveraging the inherent permutation, rotation, and scaling symmetries of Transformer architectures. We adapt parameter space alignment for modern Grouped-Query Attention (GQA) and SwiGLU layers, exploring both weight-based and activation-based approaches. Using this alignment-first strategy, we successfully transfer advanced reasoning skills to a non-reasoning model. Experiments on challenging reasoning benchmarks show that our method consistently outperforms standard task arithmetic. This work provides an effective approach for merging and transferring specialized skills across evolving LLM families, reducing redundant fine-tuning and enhancing model adaptability.

Knocking-Heads Attention

Multi-head attention (MHA) has become the cornerstone of modern large language models, enhancing representational capacity through parallel attention heads. However, increasing the number of heads inherently weakens individual head capacity, and existing attention mechanisms - whether standard MHA or its variants like grouped-query attention (GQA) and grouped-tied attention (GTA) - simply concatenate outputs from isolated heads without strong interaction. To address this limitation, we propose knocking-heads attention (KHA), which enables attention heads to "knock" on each other - facilitating cross-head feature-level interactions before the scaled dot-product attention. This is achieved by applying a shared, diagonally-initialized projection matrix across all heads. The diagonal initialization preserves head-specific specialization at the start of training while allowing the model to progressively learn integrated cross-head representations. KHA adds only minimal parameters and FLOPs and can be seamlessly integrated into MHA, GQA, GTA, and other attention variants. We validate KHA by training a 6.1B parameter MoE model (1.01B activated) on 1T high-quality tokens. Compared to baseline attention mechanisms, KHA brings superior and more stable training dynamics, achieving better performance across downstream tasks.

Compressed Convolutional Attention: Efficient Attention in a Compressed Latent Space

Multi-headed Attention's (MHA) quadratic compute and linearly growing KV-cache make long-context transformers expensive to train and serve. Prior works such as Grouped Query Attention (GQA) and Multi-Latent Attention (MLA) shrink the cache, speeding decode, but leave compute, which determines prefill and training speed, largely unchanged. We introduce Compressed Convolutional Attention (CCA), a novel attention method which down-projects queries, keys, and values and performs the entire attention operation inside the shared latent space. This simple design dramatically cuts parameters, KV-cache, and FLOPs all at once by the desired compression factor. Because CCA is orthogonal to head-sharing, we combine the two to form Compressed Convolutional Grouped Query Attention (CCGQA), which further tightens the compute-bandwidth Pareto frontier so that users can tune compression toward either FLOP or memory limits without sacrificing quality. Experiments show that CCGQA consistently outperforms both GQA and MLA at equal KV-cache compression on dense and MoE models. Additionally, we show that CCGQA outperforms all other attention methods on MoE models with half the KV-cache of GQA and MLA, achieving an 8x KV-cache compression with no drop in performance compared to standard MHA. CCA and CCGQA also dramatically reduce the FLOP cost of attention which leads to substantially faster training and prefill than existing methods. On H100 GPUs, our fused CCA/CCGQA kernel reduces prefill latency by about 1.7x at a sequence length of 16k relative to MHA, and accelerates backward by about 1.3x.

MCP-RiskCue: Can LLM Infer Risk Information From MCP Server System Logs?

Large language models (LLMs) demonstrate strong capabilities in solving complex tasks when integrated with external tools. The Model Context Protocol (MCP) has become a standard interface for enabling such tool-based interactions. However, these interactions introduce substantial security concerns, particularly when the MCP server is compromised or untrustworthy. While prior benchmarks primarily focus on prompt injection attacks or analyze the vulnerabilities of LLM MCP interaction trajectories, limited attention has been given to the underlying system logs associated with malicious MCP servers. To address this gap, we present the first synthetic benchmark for evaluating LLMs ability to identify security risks from system logs. We define nine categories of MCP server risks and generate 1,800 synthetic system logs using ten state-of-the-art LLMs. These logs are embedded in the return values of 243 curated MCP servers, yielding a dataset of 2,421 chat histories for training and 471 queries for evaluation. Our pilot experiments reveal that smaller models often fail to detect risky system logs, leading to high false negatives. While models trained with supervised fine-tuning (SFT) tend to over-flag benign logs, resulting in elevated false positives, Reinforcement Learning from Verifiable Reward (RLVR) offers a better precision-recall balance. In particular, after training with Group Relative Policy Optimization (GRPO), Llama3.1-8B-Instruct achieves 83% accuracy, surpassing the best-performing large remote model by 9 percentage points. Fine-grained, per-category analysis further underscores the effectiveness of reinforcement learning in enhancing LLM safety within the MCP framework. Code and data are available at: https://github.com/PorUna-byte/MCP-RiskCue