Autoregressive models have transformed protein engineering by enabling the generation of novel protein sequences beyond those found in nature. However, their sequential inference introduces significant latency, limiting their utility in high-throughput protein screening. Speculative decoding accelerates generation by employing a lightweight draft model to sample tokens, which a larger target model then verifies and refines. Yet, in protein sequence generation, draft models are typically agnostic to the structural and functional constraints of the target protein, leading to biologically implausible outputs and a shift in the likelihood distribution of generated sequences. We introduce SpecMER (Speculative Decoding via k-mer Guidance), a novel framework that incorporates biological, structural, and functional priors using k-mer motifs extracted from multiple sequence alignments. By scoring candidate sequences in parallel and selecting those most consistent with known biological patterns, SpecMER significantly improves sequence plausibility while retaining the efficiency of speculative decoding. SpecMER achieves 24-32% speedup over standard autoregressive decoding, along with higher acceptance rates and improved sequence likelihoods.

Peer review serves as a backbone of academic research, but in most AI conferences, the review quality is degrading as the number of submissions explodes. To reliably detect low-quality reviews, we define misinformed review points as either "weaknesses" in a review that contain incorrect premises, or "questions" in a review that can be already answered by the paper. We verify that 15.2% of weaknesses and 26.4% of questions are misinformed and introduce ReviewScore indicating if a review point is misinformed. To evaluate the factuality of each premise of weaknesses, we propose an automated engine that reconstructs every explicit and implicit premise from a weakness. We build a human expert-annotated ReviewScore dataset to check the ability of LLMs to automate ReviewScore evaluation. Then, we measure human-model agreements on ReviewScore using eight current state-of-the-art LLMs and verify moderate agreements. We also prove that evaluating premise-level factuality shows significantly higher agreements than evaluating weakness-level factuality. A thorough disagreement analysis further supports a potential of fully automated ReviewScore evaluation.

Abstract--Current anti-spoofing systems remain vulnerable to expressive and emotional synthetic speech, since they rarely leverage prosody as a discriminative cue. In this paper, we propose HuLA, a two-stage prosody-aware multi-task learning framework for spoof detection. In Stage 2, the model is jointly optimized for spoof detection and prosody tasks on both real and synthetic data, leveraging prosodic awareness to detect mismatches between natural and expressive synthetic speech. Experiments show that HuLA consistently outperforms strong baselines on challenging out-of-domain dataset, including expressive, emotional, and cross-lingual attacks. These results demonstrate that explicit prosodic supervision, combined with SSL embeddings, substantially improves robustness against advanced synthetic speech attacks. Anti-spoofing aims to detect audio generated through replay attacks, speech synthesis, and voice conversion (VC) [1]. Recent progress in text-to-speech (TTS) [2]-[6] and VC systems [7]-[11] has amplified concerns about expressive synthetic speech, which can be misused to compromise bio-metric authentication or impersonate speakers for spreading misinformation [12], [13]. One of the goals of speech generation is to produce speech that is natural and indistinguishable from human speech. Expressiveness and emotion are the defining characteristics of human speech. While this limitation is a weakness for synthesis, it represents a valuable opportunity for anti-spoofing: imperfect expressive-A. Mahapatra is with the Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD 21218 USA (e-mail: amahapa2@jhu.edu). Ismail Rasim Ulgen is with the Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, MD 21218 USA (e-mail: iulgen1@jhu.edu).

While deep learning models excel at predictive tasks, they often overfit due to their complex structure and large number of parameters, causing them to memorize training data, including noise, rather than learn patterns that generalize to new data. To tackle this challenge, this paper proposes a new regularization method, i.e., Enforcing Domain-Informed Monotonicity in Deep Neural Networks (DIM), which maintains domain-informed monotonic relationships in complex deep learning models to further improve predictions. Specifically, our method enforces monotonicity by penalizing violations relative to a linear baseline, effectively encouraging the model to follow expected trends while preserving its predictive power. We formalize this approach through a comprehensive mathematical framework that establishes a linear reference, measures deviations from monotonic behavior, and integrates these measurements into the training objective. We test and validate the proposed methodology using a real-world ridesourcing dataset from Chicago and a synthetically created dataset. Experiments across various neural network architectures show that even modest monotonicity constraints consistently enhance model performance. DIM enhances the predictive performance of deep neural networks by applying domain-informed monotonicity constraints to regularize model behavior and mitigate overfitting.

Mask-based pretraining has become a cornerstone of modern large-scale models across language, vision, and recently biology. Despite its empirical success, its role and limits in learning data representations have been unclear. In this work, we show that the behavior of mask-based pretraining can be directly characterized by test risk in high-dimensional minimum-norm ("ridge-less") linear regression, without relying on further model specifications. Further analysis of linear models uncovers several novel aspects of mask-based pretraining. The theoretical framework and its implications have been validated across diverse neural architectures (including MLPs, CNNs, and Transformers) applied to both vision and language tasks. Guided by our theory, we propose an embarrassingly simple yet overlooked pretraining scheme named Randomly Random Mask AutoEncoding (R$^2$MAE), which enforces capturing multi-scale features from data and is able to outperform optimal fixed mask ratio settings in our linear model framework. We implement R$^2$MAE in vision, language, DNA sequence, and single-cell models, where it consistently outperforms standard and more complicated masking schemes, leading to improvements for state-of-the-art models. Our code is available at: https://github.com/MingzeDong/r2mae

Parameter-efficient fine-tuning (PEFT) methods must be resource-efficient yet handle heterogeneous reasoning transformations, and classical low-rank adaptation (LoRA) is constrained by the nominal rank $r$. Hadamard-style extensions like HiRA raise the nominal rank but couple every update to the global energy pattern of the frozen weight matrix, while ABBA trades this inductive bias for fully learned dense intermediates. To address the limitation of global modulation, we propose Block Hadamard high-Rank Adaptation (BHRA), which partitions each weight matrix and applies HiRA-style multiplicative modulation independently within every block, preserving the PEFT parameter footprint while unlocking localized rank amplification. Our empirical analyses reveal that this blockwise design maintains rich spectra across rank budgets, mitigating the collapse induced by global modulation. Across eight commonsense reasoning tasks and two arithmetic benchmarks with Llama-3.2 1B/3B, Mistral-7B, and Gemma-2 9B, BHRA consistently surpasses strong PEFT baselines under matched parameter budgets.

Language Models (LMs) have significantly advanced natural language processing and enabled remarkable progress across diverse domains, yet their black-box nature raises critical concerns about the interpretability of their internal mechanisms and decision-making processes. This lack of transparency is particularly problematic for adoption in high-stakes domains, where stakeholders need to understand the rationale behind model outputs to ensure accountability. On the other hand, while explainable artificial intelligence (XAI) methods have been well studied for non-LMs, they face many limitations when applied to LMs due to their complex architectures, considerable training corpora, and broad generalization abilities. Although various surveys have examined XAI in the context of LMs, they often fail to capture the distinct challenges arising from the architectural diversity and evolving capabilities of these models. To bridge this gap, this survey presents a comprehensive review of XAI techniques with a particular emphasis on LMs, organizing them according to their underlying transformer architectures: encoder-only, decoder-only, and encoder-decoder, and analyzing how methods are adapted to each while assessing their respective strengths and limitations. Furthermore, we evaluate these techniques through the dual lenses of plausibility and faithfulness, offering a structured perspective on their effectiveness. Finally, we identify open research challenges and outline promising future directions, aiming to guide ongoing efforts toward the development of robust, transparent, and interpretable XAI methods for LMs.

Program verification relies on loop invariants, yet automatically discovering strong invariants remains a long-standing challenge. We introduce a principled framework for evaluating LLMs on invariant synthesis. Our approach uses a verifier-based decision procedure with a formal soundness guarantee and assesses not only correctness but also the speedup that invariants provide in verification. We evaluate 7 state-of-the-art LLMs, and existing LLM-based verifiers against the traditional solver UAutomizer. While LLM-based verifiers represent a promising direction, they do not yet offer a significant advantage over UAutomizer. Model capability also proves critical, as shown by sharp differences in speedups across models, and our benchmark remains an open challenge for current LLMs. Finally, we show that supervised fine-tuning and Best-of-N sampling can improve performance: fine-tuning on 3589 instances raises the percentage of speedup cases for Qwen3-Coder-480B from 8% to 29.2%, and Best-of-N sampling with N=16 improves Claude-sonnet-4 from 8.8% to 22.1%.

The expanding long-context capabilities of large language models are constrained by a significant memory bottleneck: the key-value (KV) cache required for autoregressive generation. This bottleneck is substantial; for instance, a Llama-3.1-8B model processing a 32K-token prompt at a batch size of 4 requires approximately 16GB for its KV cache, a size exceeding the model's weights. While KV-cache compression via low-rank projection is a promising direction, existing methods rely on a static, offline-learned subspace that performs poorly under data distribution shifts. To overcome these limitations, we introduce OjaKV, a novel framework that integrates a strategic hybrid storage policy with online subspace adaptation. First, OjaKV recognizes that not all tokens are equally important for compression; it preserves the crucial first and most recent tokens in full-rank, maintaining high-fidelity anchors for attention. Second, for the vast majority of intermediate tokens, it applies low-rank compression by incrementally adapting the projection basis using Oja's algorithm for online principal component analysis. This adaptation involves a comprehensive update during prompt prefilling and lightweight periodic updates during decoding, ensuring the subspace remains aligned with the evolving context. Crucially, our framework is fully compatible with modern attention modules like FlashAttention. Experiments demonstrate that OjaKV maintains or even improves zero-shot accuracy at high compression ratios. In particular, OjaKV achieves its strongest gains on very long-context benchmarks that require complex reasoning, highlighting the importance of online subspace adaptation in dynamically tracking context shifts. These results establish our hybrid framework as a practical, plug-and-play solution for memory-efficient long-context inference without requiring model fine-tuning.

Training large models ranging from millions to billions of parameters is highly resource-intensive, requiring significant time, compute, and memory. It is observed that most of the learning (higher change in weights) takes place in the earlier stage of the training loop. These changes stabilize as training continues enabling them to be captured by matrices of a low intrinsic rank. Therefore, we propose an approach to identify such states of partial convergence and dynamically switch from full parameter training to Low Rank Adaptation (LoRA) on the ViT-Large model. W e introduce a flexible approach that leverages user-defined hyper-parameters to determine the switching point and assign a rank specific to each module layer based on its level of convergence. Experimental results show that this approach preserves model accuracy while reducing the number of train-able parameters to 10% of its original size, resulting in a 3 improvement in throughput, and a 1.5 reduction in average training time per epoch while also reducing GPU memory consumption by 20%.